The wealth channel of monetary policy in the global art market

TABLE OF CONTENTS

INTRODUCTION
1.1 The Core Argument
1.2 The Empirical Strategy
1.3 Structure of the Thesis

LITERATURE REVIEW
2.1 Measuring Art Returns: Index Construction, Bias, and Interpretation
2.1.1 The Repeat-Sales Regression
2.1.2 Selection Bias: The Korteweg-Kraussl-Verwijmeren Correction
2.1.3 The Hedonic Regression Alternative
2.1.4 Practical Data Sources and Their Limitations
2.2 Art’s Risk-Return Profile: The Asset Allocation Debate
2.2.1 Returns: The Convergence Toward Moderate Performance
2.2.2 Risk: Volatility, Illiquidity, and Transaction Costs
2.2.3 Correlation and Diversification: Conditional on the Regime
2.3 Macroeconomic and Financial Determinants of Art Prices
2.3.1 Wealth, Income Inequality, and the Demand for Art
2.3.2 Equity Market Co-Movement and the Pro-Cyclicality Debate
2.3.3 Art as an Inflation Hedge
2.3.4 Art Market Microstructure: Why Prices Adjust with Lags
2.4 Monetary Policy Transmission and the Wealth-Effect Channel
2.4.1 Monetary Policy and Asset Prices: The General Mechanism
2.4.2 The Wealth Effect of Monetary Policy: From Assets to Spending
2.4.3 The Missing Link: From Wealth Effects to Art Prices
2.4.4 Regime Dependence in Monetary Transmission
2.5 Synthesis and Identification of the Research Gap

THEORETICAL AND ECONOMIC FRAMEWORK
3.1 Art in Portfolio Theory: The Markowitz Foundation
3.1.1 Art’s Risk-Return Profile
3.1.2 Worked Example: The Diversification Benefit of Art
3.2 The CAPM and Multi-Factor Extensions for Art
3.3 The Wealth-Channel Hypothesis: The Novel Contribution
3.3.1 Stage 1: Monetary Policy ^ Financial Wealth
3.3.2 Stage 2: Financial Wealth ^ Art Demand
3.3.3 Stage 3: The Lag Structure
3.3.4 The Mediation Test: Empirical Identification Strategy
3.4 Regime Dependence: When the Wealth Channel Switches Off
3.4.1 Microfoundation: Loss Aversion and the Asymmetric Wealth Elasticity
3.4.2 Formalized Segment Theory: Buyer-Type Decomposition
3.5 Selection Bias and the Illiquidity Problem
3.6 Synthesis: From Theory to Testable Predictions

DATA AND METHODOLOGY
4.1 Art Return Data
4.2 Macro-Financial Data
4.3 Variable Transformations and Descriptive Statistics
4.4 The Five-Layer Estimation Architecture
4.4.2 Layer 1: Baseline OLS with HAC Standard Errors
4.4.3 Layer 2: ARDL Model
4.4.4 Layer 3: Wealth-Channel Mediation Test
4.4.5 Layer 4: Regime-Dependent Analysis
4.4.6 Layer 5: Segment-Level Decomposition
4.5 Robustness and Reproducibility

EMPIRICAL RESULTS
5.1 Data Verification
5.2 Stationarity Testing
5.3 Layer 1: Baseline OLS with HAC Standard Errors
5.4 Layer 2: ARDL Model: Testing H
5.5 Layer 3: Wealth-Channel Mediation Test: H2
5.6 Layer 4: Regime-Dependent Analysis: H3
5.7 Layer 5: Segment Decomposition: H4
5.8 Asset-Pricing Benchmark: CAPM
5.9 Robustness Checks
5.10 Selection-Bias Adjustment (Korteweg-Kraussl-Verwijmeren)
5.11 Stationary Block Bootstrap and Leave-One-Out Jackknife
5.12 Risk-Appetite Competing Hypothesis Test
5.13 Synthesis: Hypothesis Verdicts

DISCUSSION
6.1 Reconciling the Four Hypotheses with the Evidence
6.1.1 H1: Supported Under Conventional Inference
6.1.2 H2: Formal Failure, Honest Reporting
6.1.3 H3: Expansion Coefficient Robust; Formal Regime Test Fragile to Specification
6.1.4 H4: Non-Monotonic, Reinterpreted
6.2 Dialogue with the Existing Literature
6.3 Economic Magnitudes and Historical Validation
6.3.1 Magnitude of a One-Standard-Deviation Wealth Shock
6.3.2 Historical Decomposition: The 2008 Crash
6.3.3 Out-of-Sample Validation: The 2022-2024 Tightening Cycle
6.4 Transaction-Cost-Adjusted Investability
6.5 UK Tax and Regulatory Regimes as a Distinct Policy Channel
6.6 Implications for Portfolio Construction
6.7 Implications for the Art Industry and for Monetary Policy Analysis
6.8 Critical Self-Assessment

CONCLUSION
7.1 Primary Contributions
7.2 Limitations
7.3 Directions for Future Research
7.4 Final Statement: Is Art an Investable Asset?

ABSTRACT

This thesis develops and tests a wealth-channel framework for monetary-policy transmission to art-market prices. Drawing on an original dataset obtained directly from the Artprice Econometrics Department, covering quarterly returns on the Artprice Global Index from 1998Q1 through 2026Q1 and four art-historical segment indices at annual frequency from 2000 through 2024, the analysis employs a five-layer estimation architecture: baseline OLS with Newey-West HAC standard errors, an autoregressive distributed lag (ARDL) specification, a Baron-Kenny mediation test, threshold regressions for regime dependence, and a segment-level decomposition. The Federal Reserve top-1% wealth share (FRED: WFRBST01134) is adopted as the primary wealth proxy across all five layers, both because it is theoretically aligned with the wealth-channel hypothesis (which is about ultra-high-net-worth (UHNW) collectors specifically) and because it strictly excludes household art holdings, removing the construction-endogeneity concern that affects the Z.1 household net-worth series. The central thesis is that expansionary monetary policy is associated with growth in the financial wealth of ultra-high-net-worth collectors, whose deferred consignment and bidding decisions are transmitted to auction prices with a six-to-eighteen-month lag that is structurally anchored in the biannual auction calendar. Identification of the wealth channel does not rely on instrumental-variable methods, which are infeasible at annual frequency with T = 25; instead, it rests on a structural micro-foundation, a falsifiable lag prediction confirmed by the ARDL coefficient pattern and by practitioner testimony, and a natural-experiment test against the 2022-2024 Federal Reserve tightening cycle. The headline empirical findings are as follows. The Layer 1 baseline yields a wealth coefficient of +3.68 (HAC SE 1.39, p = 0.016), rising to +4.44 (HAC SE 1.15, p = 0.002) when high-influence years are excluded by Cook’s D diagnostics. The Layer 2 ARDL(1,1) cumulative long-run multipliers are simultaneously significant at the 5% level under conventional inference: the long-run multiplier on the policy rate is -5.36 (delta-method SE 2.45, p = 0.045, 95% CI [-10.15, -0.56]) and the long-run multiplier on the top-1% wealth share is +3.70 (SE 1.55, p = 0.019, 95% CI [+0.66, +6.73]). At quarterly frequency (T = 106), the long-run wealth multiplier rises to +14.44 (SE 4.85, p = 0.003, 95% CI [+4.94, +23.94]), the strongest single statistical result in the empirical chapter. The wealth coefficient is positive in every robustness specification examined (FFR-only, Fed balance-sheet- only, post-GFC, exclude-2020, exclude-2015), with magnitudes consistently in the +3.3 to +4.4 range and significance at the 5% level in five of six robustness rows. Within an expansion-only subsample, the wealth coefficient is +4.72 (HAC SE 1.22, p = 0.002), consistent with the wealth­channel mechanism being most cleanly identified during expansions; the formal regime­independence test does not reject under the cleaner interaction-term specification, so the regime­asymmetry claim is reported as a sub-finding rather than as a primary contribution. The reduced- form segment gradient under the top-1% share is approximately flat across the older three segments (19th C +3.15, Modern +3.40, Post-War +3.94, all p > 0.10) with the Contemporary segment near zero (+0.37, p = 0.88), a pattern consistent with the supply-side micro-foundation in which inventory-rigid segments produce similar reduced-form coefficients and the Contemporary segment’s high idiosyncratic-return variance dilutes the wealth signal. The formal Baron-Kenny mediation test fails to reject the null of no mediation (Sobel z = -0.24, p = 0.81), reflecting a weak first stage between annual policy variation and the slow-moving top-1% share; the second-stage wealth-art coefficient is, however, statistically significant (b = +3.68, p = 0.016), so the components of the channel are individually detectable in the data even when the formal three-step test cannot reject the absence of mediation. The findings extend the wealth-effect literature beyond standard consumption into luxury alternative assets, provide quantitative foundations for cyclical art-allocation strategy, and open an original window onto the distributional incidence of monetary accommodation.

Keywords: art market, monetary policy transmission, wealth effect, regime dependence, Artprice Global Index, luxury asset pricing, Baron-Kenny mediation, Markov-switching, ultra-high-net­worth collectors.

This thesis would not have been possible without the support of many. My deepest gratitude goes to my advisor, Professor Sara Farooqi, whose incisive feedback at every stage, from the initial problem formulation through the final manuscript, sharpened both the argument and its presentation. Her insistence on intellectual honesty in the face of inconvenient results shaped the analytical posture adopted throughout this work.

A sincere thank-you goes to Tom Legh, the Head of Valuations at Christie's, for agreeing to an interview in London and for providing his firsthand insights into the art market. I am also grateful to the Artprice Econometrics Department, particularly for its willingness to share segment­level index data that is not included in the publisher's standard public releases; the quality of the empirical contribution rests squarely on the quality of that data.

I thank my professors at IE University for the training that made this project possible, and my colleagues in the Bachelor in Economics program for the many conversations, though not always directly about art markets, that, in visible and invisible ways, contributed to the thinking that appears here.

In accordance with the disclosure expectations of contemporary economic research practice, I acknowledge the use of artificial intelligence assistance during the preparation of this thesis. AI tools were used to support the literature synthesis and research stages, to assist with the drafting and debugging of the Stata code underlying the empirical chapter, and to refine the expository prose at the editorial stage. All substantive analytical decisions, including the choice of identification strategy, the specification of the five-layer econometric architecture, the interpretation of results, and the framing of contributions and limitations, are my own. The numerical results, the structural micro-foundation, and the practical implications drawn in the discussion chapter were authored, reviewed, and verified against the underlying data and replication code; any remaining errors of judgment or interpretation are, accordingly, also my own.

LIST OF FIGURES AND TABLES

Figure 3.1 Wealth-Channel Transmission Mechanism

Figure 4.1 Wealth-Channel Identification DAG

Figure 5.1 Artprice Global Index (2000–2024) with Stress-Regime Shading

Figure 5.2 ARDL Cumulative Impulse Response Function

Figure 5.3 Regime-Conditional Wealth Coefficients

Figure 5.4 Segment-Level Wealth-Sensitivity Reduced-Form Coefficients

Figure 5.5 Wealth Coefficient Comparison: Z.1 vs Top-1% Share across Ten Specifications

Figure 6.1 Out-of-Sample Validation: 2022–2024 Tightening Cycle

Figure C.1 Wealth-Coefficient Comparison Across 19 Specifications

Figure C.2 Quarterly POOS Forecast Sequence (Top-1% Share)

Table 3.1 Cross-Asset Risk-Return Summary

Table 3.2 Two-Asset Portfolio Optimization Results

Table 3.3 Segment-Level Wealth-Channel Predictions

Table 3.4 Master Hypothesis Summary: Theoretical Predictions and Empirical Tests

Table 4.1 Artprice Global Index: Year-End Levels and Annual Returns

Table 4.2 Descriptive Statistics of Key Variables

Table 5.1 Verification of Computed Returns Against Published Artprice Figures

Table 5.2 Layer 1 OLS Results (Z.1 Net-Worth Reference Specification)

Table 5.3 Layer 1 OLS Results, Primary Specification (Top-1% Wealth Share)

Table 5.4 Layer 2 ARDL Results (Z.1 Net-Worth Reference Specification)

Table 5.5 Layer 2 ARDL(1,1) Long-Run Multipliers, Primary Specification

Table 5.6 Layer 3 Baron-Kenny Mediation Test

Table 5.7 Layer 4 Threshold Regression: Regime-Conditional Wealth Coefficients

Table 5.8 Layer 5 Segment-Level OLS Regressions

Table 5.9 Robustness of Wealth and Policy Coefficients

Table 5.10 Selection-Bias-Adjusted Wealth Coefficients (Korteweg–Kräussl–Verwijmeren)

Table 5.11 Post-2021 Risk-Appetite Subsample Regressions

Table 6.1 Master Hypothesis Test Summary

Table 6.2 Historical Natural Experiments: Model Predictions vs. Observed

Table A.1 Layer 1 Baseline OLS and Newey-West HAC Estimates

Table A.2 Residual Diagnostics for Layer 1 Baseline

Table A.3 Variance Inflation Factors for Layer 1 Baseline

Table A.4 Influence Statistics for Layer 1 Baseline

Table A.5 Layer 1 Re-estimation Excluding Cook’s D Outliers

Table B.1 Information Criteria Across 13 Specifications

Table B.2 Sign-Pattern Bootstrap and Aggregated Significance Tests

Table B.3 ARDL(1,1) Long-Run Multipliers: Delta-Method and Bootstrap CIs

Table C.1 Regime-Asymmetry Sensitivity to Stress Definition

Table C.2 Continuous-VIX Interaction Specification

Table C.3 Hansen-Style sup-Wald Threshold Scan

Table C.4 Davidson-MacKinnon J-Test (CAPM vs Wealth-Channel)

Table C.5 Policy-Rate Proxy Robustness (FFR / Wu-Xia / Composite)

Table C.6 Full Wealth-Proxy Comparison (19 Specifications)

CHAPTER 1

INTRODUCTION

The global art market cleared approximately $57.5 billion (USD) in dealer and auction sales in 2024, with an estimated $1.7 trillion in art and collectibles held in the investable portfolios of ultra-high-net-worth individuals worldwide (Deloitte, 2024; Art Basel & UBS, 2025). This is a consequential asset class, yet it occupies a peculiar position in the canon of financial economics. Equity returns are modeled, debated, and benchmarked against the CAPM; bond returns are decomposed into their yield-curve and credit components; even cryptocurrencies have acquired a small but rigorous research literature. Art, by contrast, sits uneasily at the intersection of portfolio theory and cultural economics, its price behavior documented primarily by practitioner commentary and isolated academic case studies rather than by an integrated theoretical framework.

This thesis proceeds from the observation that the conventional portfolio-optimization case for art is substantially weaker than popular discourse acknowledges. Once selection bias in observed returns is accounted for, the corrected Sharpe ratio for broad art portfolios is approximately 0.11 (Korteweg, Kraussl and Verwijmeren, 2016), compared to roughly 0.30 for U.S. equities over comparable periods. Round-trip transaction costs at major auction houses approach twenty-five to thirty percent of hammer value. The median time-to-sale for a work valued above one million dollars is six to twelve months. Individual lot prices range from thousands to hundreds of millions of dollars, precluding the diversification that meaningful portfolio allocations would require for all but the largest family offices. On any narrow reading of the Markowitz framework, art fails the allocation test.

And yet the allocation exists, persistently and at scale. Approximately four to five percent of ultra-high-net-worth portfolio wealth is allocated to art and collectibles (Knight Frank, 2024). Every major private bank with an ultra-wealthy client franchise maintains an art advisory practice. Every major auction house markets its Evening Sales explicitly as investment events, with pre­sale estimates, third-party guarantees, and curated dialogue between contemporary and historical works. The empirical pattern to be explained is not whether wealthy collectors buy art but how the

The Wealth Channel of Monetary Policy aggregate price behavior of the market responds to macroeconomic conditions that affect the wealth, sentiment, and liquidity of those collectors. The question this thesis poses is therefore mechanistic rather than normative: through what economic mechanism does monetary policy propagate to art market returns, and how does that mechanism vary across macroeconomic regimes and art-market segments?

1.1 The Core Argument

The argument advanced in this thesis can be stated compactly. Expansionary monetary policy, whether conventional rate cuts or quantitative easing, inflates the value of financial assets held disproportionately by ultra-high-net-worth households. The resulting wealth gains are redeployed into the art market with a lag that is structurally determined by the biannual auction calendar, by the discrete decision horizon of collectors considering consignment or acquisition, and by the sequential process of price discovery through comparable works trading at successive sales. The strength of this transmission varies systematically: it is strongest during macroeconomic expansions when wealth is growing, and risk appetite is high, and it operates with greater intensity in Contemporary and Post-War segments whose buyer base is most tightly concentrated among the newly wealthy. During periods of financial stress, the mechanism effectively reverses as wealth destruction, forced selling, and flight to liquidity combine to depress prices beyond what the underlying monetary variables alone would predict.

This is the wealth-channel hypothesis. Its intellectual antecedents are well-established piecewise. The macroeconomic literature has demonstrated that monetary policy transmits to asset prices through discount-rate, portfolio-balance, and wealth-effect channels (Bernanke and Kuttner, 2005; Gagnon et al., 2011). Research on the distributional incidence of unconventional policy has shown that the wealth gains accrue disproportionately to the top of the wealth distribution (Montecino and Epstein, 2015; Ampudia et al., 2018). The art-finance literature has separately documented long-run co-movement between art prices, equity returns, and measures of income inequality (Goetzmann, Renneboog and Spaenjers, 2011), and has modeled the microstructure of auction trading and consignment decisions (Lovo and Spaenjers, 2018). What this thesis contributes is the formal connection between these piecewise results: a specification of monetary policy transmitting through financial wealth to art demand to auction prices, with testable

The Wealth Channel of Monetary Policy predictions about lag structure, regime dependence, and segment heterogeneity that previous aggregate analyses have not delivered.

1.2 The Empirical Strategy

The empirical analysis employs an original dataset obtained directly from the Artprice Econometrics Department, covering the Artprice Global Index and four segment-level indices (nineteenth-century, Modern, Post-War, and Contemporary) from 1998Q1 through 2026Q1. The primary specification is estimated at an annual frequency with twenty-five observations, aligning with the annual reporting cycle of the macroeconomic variables that serve as explanatory candidates. The methodology proceeds in five layers: baseline OLS with Newey-West HAC standard errors establishes the unconditional associations; an autoregressive distributed lag model tests for the lagged transmission predicted by the auction-calendar microstructure; a Baron-Kenny mediation test formally evaluates whether household wealth is the intermediary between monetary policy and art returns; a threshold regression on the VIX identifies regime-dependent coefficient structures; and a segment-level decomposition tests whether the wealth channel intensifies moving from older to newer art-historical categories.

Throughout, the thesis adopts a deliberately conservative analytical posture. Small-sample t- distribution inference is used rather than asymptotic normal approximation; Newey-West HAC standard errors with the T/(T-k) finite-sample correction are applied uniformly; every coefficient is cross-verified between Stata 18 and R 4.3 to four decimal places. Where results contradict the hypothesized directions or fail to reach conventional significance, the thesis says so explicitly. The failure of the formal Baron-Kenny mediation test is treated with the same analytical seriousness as the success of the regime-dependence test. Intellectual honesty about negative findings is as critical to the thesis as the documentation of positive ones.

1.3 Structure of the Thesis

The thesis is organized in seven chapters. Chapter 2 reviews the academic literature relevant to the analysis, organized around four substantive debates: the measurement of art returns and the methodological challenge of selection bias; the risk-return profile of art relative to traditional assets; the macroeconomic and financial determinants of art prices; and the monetary-policy

The Wealth Channel of Monetary Policy transmission and wealth-effect literature from which the wealth-channel hypothesis derives. Chapter 3 develops the theoretical and economic framework, moving from the canonical Markowitz setup through the CAPM and Fama-French extensions to the novel three-stage specification of the wealth channel and its regime-dependent extension. Chapter 4 documents the data and econometric methodology: variable construction, data sources at the series-identifier level, stationarity testing protocol, and the five-layer estimation architecture. Chapter 5 presents the empirical results with diagnostic tests, robustness checks, and a transparent account of which hypotheses are supported and which are not. Chapter 6 interprets the findings in dialogue with the prior literature and develops the economic magnitudes through historical decomposition of the 2008 crisis and out-of-sample validation against the 2022-2024 Federal Reserve tightening cycle. Chapter 7 concludes with the primary contributions, acknowledged limitations, and directions for future research.

The structure is modular. The baseline analyses in Chapters 4 and 5 stand on their own; the more advanced regime-dependent and segment-level specifications are offered as extensions whose failure (should it occur) would not undermine the core argument. This modularity was an intentional design choice, reflecting the recognition that with twenty-five annual observations, no single econometric specification can carry the full weight of inference. The strength of the evidence rests on the convergence of multiple specifications toward a consistent economic narrative, not on any one coefficient t-statistic clearing a conventional threshold.

CHAPTER 2

LITERATURE REVIEW

This chapter examines the academic literature relevant to the analysis of art as an investable asset and the mechanisms through which macroeconomic and financial conditions transmit to art market returns. Rather than offering a chronological catalog of prior studies, the review is organized around four substantive debates that collectively define the research frontier and identify the specific gap this thesis addresses. Section 2.1 examines how art returns are measured and the methodological pitfalls that complicate their interpretation. Section 2.2 evaluates the evidence on art's risk-return profile relative to traditional asset classes. Section 2.3 reviews the macroeconomic and financial determinants of art prices, including the critical role of wealth and income inequality. Section 2.4 surveys the monetary-policy transmission literature and the wealth-effect channel, establishing the theoretical bridge to the empirical framework developed in Chapter 3. Each section concludes with a critical assessment of unresolved questions and limitations that motivate this thesis's contribution.

2.1 Measuring Art Returns: Index Construction, Bias, and Interpretation

Any investigation of art as an investment must first confront the fundamental measurement challenge: art is heterogeneous, illiquid, and trades endogenously. Unlike equities, where standardized securities are continuously quoted, each artwork is a unique asset that transacts infrequently, irregularly, and through an opaque market structure dominated by two major auction houses and a diffuse network of private dealers. The methods used to construct art price indices and the biases embedded in them have occupied a substantial portion of the cultural economics literature, and remain essential context for interpreting empirical results in this field.

2.1.1 The Repeat-Sales Regression

The repeat-sales regression (RSR), first developed by Bailey, Muth, and Nourse (1963) for real estate and adapted for art markets by Anderson (1974) and later Goetzmann (1993), estimates price appreciation by tracking the same artwork across multiple auction appearances. The method's

The Wealth Channel of Monetary Policy appeal lies in its control for quality differences by using each object as its own benchmark. Mei and Moses (2002) applied the RSR to a dataset of works sold at major New York auction houses, producing what became the widely cited Mei Moses Art Index, subsequently acquired by Sotheby's in 2016. Their influential finding that art returned approximately 4.0% per year in real terms between 1950 and 2000, with low correlation to equities, catalyzed a wave of interest in art as a portfolio diversifier.

However, the RSR suffers from a structural limitation that was only fully appreciated in subsequent research: it can only include artworks that have sold at least twice at public auction. This is a non-random sample. Goetzmann (1996) first flagged survivorship bias in the art market, noting that artists who fall permanently out of fashion may never re-enter the auction system, systematically inflating index returns. The severity of this problem was quantified by Korteweg, Kraussl, and Verwijmeren (2016) in a landmark paper in the Review of Financial Studies that, in this author's assessment, represents the single most important methodological advance in art return measurement over the past two decades.

2.1.2 Selection Bias: The Korteweg-Kraussl-Verwijmeren Correction

Korteweg, Kraussl, and Verwijmeren (2016) developed a generalized repeat-sales framework that explicitly models the probability of observing a sale as a function of the artwork's underlying (potentially unobserved) returns. Their central finding was that the sale decision is endogenous to performance, fundamentally challenging the reliability of all prior art index estimates. Using a dataset of 32,928 paintings that sold repeatedly between 1960 and 2013, they documented an asymmetric V-shaped relationship between sale probabilities and returns: paintings that have appreciated substantially are more likely to be consigned to auction (sellers realize gains), and paintings that have depreciated severely are also more likely to appear (forced sales due to estates, divorces, or liquidity needs). This endogeneity means that conventional RSR indices systematically over-represent extreme outcomes and under-represent the large population of artworks whose values have changed modestly and therefore remain in private hands.

The quantitative impact is striking. Correcting for selection bias reduced average annual index returns from 8.7% to 6.3% and Sharpe ratios from 0.27 to 0.11, a nearly 60% decline in risk- adjusted performance. The corrected Sharpe ratio of 0.11 renders broad passive art investment substantially less attractive than U.S. equities (Sharpe ratio of approximately 0.30 over the same

The Wealth Channel of Monetary Policy period). The authors concluded that investing in a diversified portfolio of paintings is not financially viable on a risk-adjusted basis, although targeting specific styles or top-selling artists may preserve positive abnormal returns.

This finding carries profound implications for the present thesis. Any empirical analysis using aggregate art return indices, including the Artprice Global Index and the Artprice100 indices employed in this study, operates with data that likely overstates true returns and understates true volatility. The Korteweg methodology cannot be implemented at the bachelor's thesis level, as it requires Markov Chain Monte Carlo estimation with latent price processes. Nevertheless, acknowledging this limitation is substantively important because the direction of the bias, overstatement of returns, means that any positive findings regarding art's investment properties should be interpreted as upper-bound estimates. If the wealth-channel hypothesis finds support despite operating with inflated return data, this constitutes conservative evidence in favor of the transmission mechanism.

2.1.3 The Hedonic Regression Alternative

The hedonic approach, formulated in the art context by Chanel, Gérard-Varet, and Ginsburgh (1996) and applied at scale by Renneboog and Spaenjers (2013), decomposes auction prices into observable characteristics, medium, size, artist reputation, auction house, provenance, and period, and extracts a time-varying price index from the residual temporal coefficients. This method's advantage over the RSR is that it does not require repeat sales: every transaction can contribute to index estimation, dramatically expanding the usable sample. Renneboog and Spaenjers (2013), working with a dataset of over one million auction transactions of paintings and works on paper between 1957 and 2007, estimated a real annualized return of 3.97%, a figure notably lower than the 8-10% figures commonly cited from uncorrected RSR studies and more consistent with the selection-corrected estimates of Korteweg, Kraussl, and Verwijmeren.

The hedonic method introduces its own complications, however. The choice and specification of hedonic variables is inherently subjective, and unobservable quality dimensions, the aesthetic power of a composition, the condition of the canvas, the emotional resonance of the subject matter, cannot be captured by any regression specification. As Goetzmann's (2025) recent survey for Art Basel observes, the difficulty of quantifying "the economics of taste" remains the art market's most fundamental valuation challenge, one that distinguishes it from every other asset class. Hedonic

The Wealth Channel of Monetary Policy indices may additionally suffer from specification instability as collector preferences evolve, introducing structural breaks that the econometrician must either model explicitly or accept as a source of noise.

2.1.4 Practical Data Sources and Their Limitations

The contemporary researcher has several index sources available. The Artprice Global Index, constructed by the econometrics division of Artmarket.com from a database of over 30 million auction results since 1983, provides quarterly and annual return series segmented by period and medium. The Artnet Price Database, used by Morgan Stanley's Global Investment Office in their March 2025 primer on art as an asset class, offers semiannual genre-level indices from 2000 onward. The Sotheby's Mei Moses indices offer the longest available time series but are paywalled and methodologically conservative. Art Market Research (AMR) provides style-specific indices from 1975 used in peer-reviewed research published in journals such as Empirical Economics and the Journal of Alternative Investments.

Each of these sources captures only the auction market, which represented roughly $10.2 billion across approximately 388,000 objects in 2024 (Artnet, 2025). The private dealer market, estimated to account for over half of global art sales by value, remains almost entirely unobserved by any price index. This is a structural limitation of the field, not a weakness unique to any particular study. Nevertheless, it means that all index-based research, including the present thesis, implicitly assumes that auction-market dynamics are representative of the broader art market, an assumption that may not hold, particularly for contemporary art where galleries exercise significant price-setting power through primary market sales.

2.2 Art’s Risk-Return Profile: The Asset Allocation Debate

The literature on art as an investable asset revolves around a central tension: art's observed returns are modestly positive and its correlation with traditional assets is low, which theoretically justifies a portfolio allocation; but its true risk is severely underestimated by transaction-based indices, its transaction costs are punitive, and its liquidity profile is incompatible with standard portfolio rebalancing assumptions.

2.2.1 Returns: The Convergence Toward Moderate Performance

The literature has converged over the past three decades from widely divergent return estimates toward a consensus range. Early studies produced figures ranging from the deeply pessimistic, Baumol's (1986) famous characterization of art as a "floating crap game" with returns barely exceeding storage costs, to the exuberant, with some index providers claiming annualized returns in excess of 10%. The methodological advances reviewed in Section 2.1 have narrowed this range considerably. The current best estimates, accounting for selection bias, place real annualized art returns in the range of 2-4% across broad portfolios, rising to 4-7% for specific segments (Renneboog and Spaenjers, 2013; Korteweg, Kraussl, and Verwijmeren, 2016; Chambers, Dimson, and Spaenjers, 2020). These returns are comparable to investment-grade corporate bonds, but with substantially higher volatility.

A particularly illuminating study is Chambers, Dimson, and Spaenjers (2020), who examined the long-run buy-and-hold performance of an actual invested art portfolio: the collection assembled by John Maynard Keynes at Cambridge's King's College. Unlike index-based studies, which are hypothetical constructs, this analysis tracked every artwork from purchase through subsequent valuations and sales. The authors found returns broadly consistent with the selection- corrected estimates, reinforcing the view that actual portfolio performance falls short of what naive index returns suggest. This study is methodologically important because it demonstrates that even a sophisticated, well-informed collector, Keynes was advised by leading dealers and artists of his era, did not substantially outperform the selection-corrected index benchmark.

2.2.2 Risk: Volatility, Illiquidity, and Transaction Costs

Art's risk profile is frequently understated because the standard deviation calculated from aggregate indices smooths over the enormous cross-sectional dispersion of individual artwork returns. Renneboog and Spaenjers (2013) documented that quantile regressions reveal substantially higher volatility in the upper price brackets, precisely the segments that attract investment-motivated collectors. Lovo and Spaenjers (2018), in a theoretical model of art trading published in the American Economic Review, demonstrated that transaction-based price indices systematically underestimate art's true volatility because they observe only the subset of artworks that actually trade, excluding works experiencing moderate price changes that do not trigger a consignment decision.

Transaction costs compound the risk problem. Buyer's premiums at major auction houses range from 13% to 26% of the hammer price, depending on the lot value. Seller's commissions, while negotiable for consignors of high-value works, typically add 5-10%. Insurance, storage, conservation, and shipping costs impose ongoing carrying charges that have no equivalent in financial markets. The Morgan Stanley Global Investment Office (2025) estimated in their inaugural art asset class primer that total round-trip transaction costs for a typical auction sale may reach 25-30%, implying that an artwork must appreciate by roughly one-third simply to break even on a transaction-cost-adjusted basis. This cost structure means that art is only financially viable as a long-horizon holding, which in turn exacerbates the illiquidity problem: an investor who needs to liquidate within a short timeframe faces both uncertain execution and punitive costs.

2.2.3 Correlation and Diversification: Conditional on the Regime

The diversification argument for art rests on its low unconditional correlation with equities and bonds. Renneboog and Spaenjers (2013) estimated a correlation coefficient between art returns and the S&P 500 of approximately 0.03 to 0.15 depending on the period and art segment examined. Campbell (2008) found similarly low correlations using the Mei Moses index. These figures, when inserted into a Markowitz mean-variance optimizer, generate non-trivial art allocations of 5-15% of portfolio wealth, a result that several wealth management firms and academic studies have cited as justification for art's inclusion as an asset class.

However, this argument contains a critical flaw that the literature has only recently begun to address: correlations are not constant across market regimes. The Morgan Stanley GIC (2025) observed that art's positive correlation to equities during bear markets substantially diminishes its hedging value precisely when hedging matters most. This pattern is consistent with a wealth­channel interpretation: during expansions, equity gains generate surplus wealth that flows into art, producing a positive but lagged correlation; during severe downturns, the same mechanism operates in reverse as wealth destruction curtails art demand, but the illiquidity-induced lag in art's price adjustment temporarily masks the correlation. This regime-dependent correlation structure means that the unconditional correlation figure is misleading for portfolio construction purposes, a point that motivates the regime-switching analysis in Chapter 5.

Mandel (2009), in a theoretical model published in the American Economic Review, offered an alternative explanation for art's apparently low equilibrium returns. He argued that art serves a

The Wealth Channel of Monetary Policy dual function as both an investment good and a conspicuous consumption good: collectors derive utility from the display and social signaling value of ownership, which they "pay for" by accepting financial returns below what a pure investment asset would require. This "consumption dividend" framework has important implications for understanding who buys art and why. If the non­financial return component is large, then the financially motivated investor, the investor targeted by art funds and fractional ownership platforms, is effectively subsidizing the consumption- motivated collector.

2.3 Macroeconomic and Financial Determinants of Art Prices

The relationship between art prices and macroeconomic conditions has been examined from several angles: wealth and income effects, equity market co-movement, inflation hedging, and sentiment-driven dynamics. This section reviews each in turn, building toward the identification of the specific gap, the absence of a formally tested wealth-channel transmission mechanism, that this thesis addresses.

2.3.1 Wealth, Income Inequality, and the Demand for Art

The most directly relevant study for this thesis is Goetzmann, Renneboog, and Spaenjers (2011), published in the American Economic Review. Using a newly constructed art price index spanning over two centuries of London auction data, the authors demonstrated two key results. First, equity market returns have a statistically significant contemporaneous and lagged effect on art price levels, confirming a financial market linkage. Second, and more strikingly, they found evidence of a cointegrating relationship between top income shares and art prices, suggesting that rising income inequality is structurally associated with higher art valuations in the long run. This result is consistent with art being a luxury good whose demand is concentrated among the very wealthy, a population whose share of total income has fluctuated dramatically over the past two centuries in patterns that closely mirror art market booms and busts.

The Goetzmann, Renneboog, and Spaenjers (2011) finding establishes the empirical premise for this thesis's wealth-channel hypothesis, but it does not identify the causal mechanism. The authors documented a statistical association between wealth concentration and art prices; they did not test whether this association operates through a specific transmission channel, nor did they investigate how monetary policy, the primary driver of short-to-medium-term fluctuations in

The Wealth Channel of Monetary Policy financial wealth, interacts with the wealth-art linkage. This is the precise gap that the present thesis seeks to fill.

Renneboog and Spaenjers (2013) provided complementary evidence from their hedonic analysis: they found that measures of high-income consumer confidence and art market sentiment were significant predictors of art price trends, even after controlling for macroeconomic fundamentals. This suggests that art prices are influenced not just by the level of wealth but by the psychological state of wealthy collectors, their confidence in the economic outlook, their appetite for conspicuous expenditure, and their perception of the art market's future trajectory. This "sentiment channel" is related to but distinct from the wealth channel: wealth provides the material capacity to buy art, while sentiment provides the behavioral willingness.

2.3.2 Equity Market Co-Movement and the Pro-Cyclicality Debate

The relationship between art prices and equity markets is well documented but imprecisely characterized. Mei and Moses (2002) found relatively low correlation between their art index and the S&P 500, leading them to emphasize art's diversification potential. However, Chanel (1995) and Worthington and Higgs (2004), using different datasets and time periods, found stronger equity-art linkages, particularly during market extremes. The apparent contradiction across studies can be reconciled by recognizing that the art-equity correlation is itself time-varying and regime­dependent, a point explored empirically in Renneboog and Spaenjers' (2014) cross-country analysis, which documented that national equity markets help explain country-specific art returns, though less so for artists with the highest international reputations.

The pro-cyclicality of the art market is most visible during boom-bust episodes. The late 1980s Japanese asset price bubble coincided with explosive growth in Impressionist and Modern art prices, driven in large part by Japanese collectors whose purchasing power was inflated by stock and real estate gains. Hiraki, Ito, Spieth, and Takezawa (2009) found that Japanese land prices caused both art and Japanese stock prices to co-move during the 1976-1998 sample period, interpreting this as evidence that the appreciation of land values stimulated Japanese investor demand for international art. The 2008 Global Financial Crisis produced a sharp decline in contemporary art prices, followed by a rapid recovery in 2010-2014 that coincided precisely with the expansion of the Federal Reserve's balance sheet under quantitative easing, a correlation noted

The Wealth Channel of Monetary Policy by market commentators but not yet, to this author's knowledge, subjected to formal econometric investigation.

2.3.3 Art as an Inflation Hedge

The claim that art serves as an inflation hedge is intuitively appealing, art is a tangible, scarce, non-reproducible real asset, but the empirical evidence is mixed. Renneboog and Spaenjers (2013) found that art prices respond positively to inflation in the long run, consistent with art maintaining purchasing power over extended horizons. However, the Morgan Stanley GIC (2025) observed that when inflation spiked in 2021-2023, Post-War and Contemporary art generally underperformed inflation, challenging the simple hedging narrative. This apparent contradiction can be resolved by distinguishing between moderate, anticipated inflation (associated with wealth growth and supportive of art demand) and sharp, unexpected inflation shocks (accompanied by monetary tightening, reduced financial wealth, and contracting art demand). This distinction aligns with the regime-dependent framework proposed in this thesis.

2.3.4 Art Market Microstructure: Why Prices Adjust with Lags

A feature of the art market that distinguishes it from financial markets is the structural lag in price adjustment. Lovo and Spaenjers (2018) developed a rigorous theoretical model of endogenous trading in the art auction market, published in the American Economic Review, that provides the microstructural foundation for understanding this lag. In their model, agents make purchase and sale decisions based on the relative magnitude of their private use value and the expected resale revenue. Individuals with strong taste for an artwork pay high prices and sell only when hit by a liquidity shock; those with weak taste aim to resell quickly at a profit. This generates endogenous patterns: holding periods and financial returns are negatively correlated, and crucially, prices and auction volume increase during economic expansions as speculative activity rises.

The Lovo and Spaenjers model has direct implications for the present thesis. It predicts that the volume of consignments to auction houses, a proxy for art market activity, responds to economic conditions with a lag determined by the frequency of "liquidity shocks" and the biannual structure of the major auction calendar. When financial wealth increases following monetary easing, the model predicts a gradual increase in both prices and voluntary sales volume as collectors in the "weak taste" category attempt to capitalize on favorable conditions and auction

The Wealth Channel of Monetary Policy houses expand their sales programs in response to increased consignment supply. This provides a micro-founded justification for the 6-18 month lag between monetary policy shocks and observable art price effects hypothesized in Chapter 3.

Ashenfelter (1989) and Beggs and Graddy (2009) have documented additional microstructural features that affect price dynamics: declining price anomalies within auction sessions, anchoring effects from pre-sale estimates, and the information content of "bought-in" lots (artworks that fail to sell at auction). These features imply that art price adjustment is not only lagged but also discrete and noisy, occurring in lumps at biannual auction events rather than continuously. This discreteness creates an identification challenge for time-series analysis: monetary policy shocks that occur between auction seasons may appear to have no contemporaneous effect on art returns, when in fact the full price adjustment has simply been deferred to the next auction cycle.

2.4 Monetary Policy Transmission and the Wealth-Effect Channel

This final section of the literature review crosses from cultural economics into macroeconomic theory, drawing on the substantial literature on monetary policy transmission mechanisms to construct the theoretical bridge to this thesis's empirical analysis. The objective is to show that the wealth-channel hypothesis proposed in Chapter 3 is not a speculative construction but rather a straightforward application of well-established monetary transmission theory to a specific alternative asset class.

2.4.1 Monetary Policy and Asset Prices: The General Mechanism

The relationship between monetary policy and asset prices is one of the most extensively studied topics in macroeconomics. Bernanke and Kuttner (2005) demonstrated that an unanticipated 25-basis-point reduction in the Federal Funds rate is associated with a roughly 1% increase in broad stock indices, operating through a combination of lower discount rates, improved earnings expectations, and reduced equity risk premia. Rigobon and Sack (2004) confirmed these findings using a heteroskedasticity-based identification strategy that addresses the simultaneity between monetary policy and financial markets. The transmission mechanism from policy rates to asset prices is well established: lower interest rates reduce the discount factor applied to future cash flows, increase the present value of earnings, and encourage portfolio rebalancing from safe assets toward riskier alternatives.

Gagnon, Raskin, Remache, and Sack (2011), in their analysis of the Federal Reserve's first round of quantitative easing, documented substantial effects on long-term Treasury yields, mortgage-backed securities, and corporate bonds. Subsequent research has extended these findings to equity markets, real estate, and other asset classes. The key insight from this literature is that monetary policy affects a broad class of asset prices, not just the specific instruments that the central bank purchases. This "portfolio balance channel," whereby central bank purchases of safe assets push investors into riskier alternatives, provides a natural mechanism through which monetary easing could ultimately affect art prices, even though art is far removed from the central bank's direct sphere of intervention.

2.4.2 The Wealth Effect of Monetary Policy: From Assets to Spending

The wealth-effect channel posits that monetary policy affects consumption and investment not only through interest rate changes but also through its effect on household wealth. When asset prices rise in response to monetary easing, households holding those assets experience an increase in net worth, which raises their lifetime consumption possibilities and their willingness to spend on non-essential goods. This mechanism is particularly relevant for luxury goods with high income elasticity, a category that unambiguously includes art.

The distributional dimension of the wealth effect is critical. Montecino and Epstein (2015) examined the distributional impact of quantitative easing on U.S. households using data from the Federal Reserve's Survey of Consumer Finances, identifying three key channels: the employment channel, which benefits lower-income households through reduced unemployment; the asset appreciation channel, which disproportionately benefits wealthy households who hold the assets whose prices are inflated by QE; and the mortgage refinancing channel, which benefits middle­class homeowners. For the art market, only the second channel is directly relevant. The ultra-high- net-worth population, defined by Knight Frank as those with investable assets exceeding $30 million, holds a disproportionate share of financial assets and allocates an estimated 4-5% of portfolio wealth to art and collectibles. When QE inflates the value of their equity and bond holdings, the resulting wealth gain creates the material conditions for increased art expenditure.

Ampudia, Georgarakos, Slacalek, Tristani, Vermeulen, and Violante (2018), in an ECB working paper, conducted a microsimulation analysis of how monetary policy and QE affect income and wealth inequality across European households. They found that the asset appreciation

The Wealth Channel of Monetary Policy channel of monetary easing increases wealth inequality by benefiting those at the top of the wealth distribution who hold the most financial assets. This finding directly supports the thesis that monetary easing should increase art demand, since art buyers are overwhelmingly drawn from the same population that benefits most from the asset appreciation channel.

2.4.3 The Missing Link: From Wealth Effects to Art Prices

Despite the logical coherence of the argument, and despite the extensive literature on both art price determinants and monetary policy wealth effects separately, no existing study formally tests the complete transmission chain from monetary policy through financial wealth to art market returns. Goetzmann, Renneboog, and Spaenjers (2011) documented the wealth-art correlation but did not model monetary policy as the upstream driver. The monetary policy transmission literature does not consider art as a downstream asset class. This thesis identifies this gap and proposes to fill it by estimating a mediation-style model that tests whether financial wealth serves as an intermediary channel between monetary policy shocks and art return responses.

The novelty of this approach is not in the individual components, each link in the chain is well supported by prior research, but in formally connecting them. This is analogous to how research on the "credit channel" of monetary transmission (Bernanke and Gertler, 1995) did not discover that banks lend or that firms invest, but rather demonstrated that the interaction between monetary policy and bank balance sheet conditions creates a distinct and quantifiable amplification mechanism. Similarly, this thesis does not claim to discover that the wealthy buy art or that monetary policy affects wealth; it claims that the interaction between these two established facts creates a measurable transmission channel with specific predictions about timing, regime dependence, and cross-sectional variation across art market segments.

2.4.4 Regime Dependence in Monetary Transmission

A growing literature documents that the effects of monetary policy are state-dependent. The Bank for International Settlements (Drechsler, Savov, and Schnabl, 2023) has demonstrated that the transmission of policy rate changes to long-term yields depends critically on market liquidity conditions: conventional policy is most effective when markets are liquid, while QE has larger effects in illiquid environments. This finding has direct implications for art: the art market is structurally illiquid at all times, suggesting that the wealth-channel transmission may be stronger

The Wealth Channel of Monetary Policy during periods when financial markets are also experiencing stress, but offset by the countervailing effect of wealth destruction. The net effect is ambiguous a priori, which is precisely why empirical testing with regime-switching methodology is necessary.

Hamilton (1989) pioneered the Markov-switching framework for identifying economic regimes endogenously from the data, and Ang and Bekaert (2002) extended this approach to international asset allocation contexts. For the art market, the relevant regimes are not simply recession versus expansion but rather combinations of monetary stance, financial market conditions, and collector sentiment. The empirical approach adopted in Chapter 5 tests a simplified two-regime framework (expansion versus stress) using observable indicators, with the Markov- switching specification available as a robustness extension.

2.5 Synthesis and Identification of the Research Gap

The literature reviewed in this chapter establishes five key findings that collectively motivate this thesis. First, art return indices are measured with substantial upward bias due to selection effects, and true risk-adjusted returns are considerably lower than popularly cited figures suggest (Korteweg, Kraussl, and Verwijmeren, 2016). Second, art prices are positively associated with financial wealth and income inequality in the long run, with equity market returns serving as a significant contemporaneous and lagged predictor (Goetzmann, Renneboog, and Spaenjers, 2011). Third, art price adjustment is structurally lagged and discrete, occurring through biannual auction cycles with endogenous consignment decisions (Lovo and Spaenjers, 2018). Fourth, monetary policy affects household wealth through well-documented asset appreciation and portfolio rebalancing channels, with effects concentrated at the top of the wealth distribution (Ampudia et al., 2018; Montecino and Epstein, 2015). Fifth, the effects of monetary policy on asset returns are regime-dependent, varying with market liquidity, risk sentiment, and the type of policy instrument employed (Hamilton, 1989; Drechsler, Savov, and Schnabl, 2023).

The gap lies in the absence of any study that formally connects findings two through five into a unified empirical framework. No paper tests whether the observed wealth-art correlation operates specifically through a monetary policy transmission channel, whether this channel exhibits the lag structure predicted by art market microstructure theory, or whether the channel's strength varies across market regimes and art market segments. This thesis proposes to fill that gap.

The Wealth Channel of Monetary Policy

CHAPTER 3

THEORETICAL AND ECONOMIC FRAMEWORK

This chapter develops the formal economic architecture on which the subsequent empirical analysis rests. It proceeds in five stages. Section 3.1 situates art within the canonical mean-variance portfolio optimization framework of Markowitz (1952), quantifying its risk-return profile relative to traditional asset benchmarks and exposing the tension between the theoretical diversification case and the practical frictions of art allocation. Section 3.2 examines the CAPM and Fama-French extensions and uses them to establish that art returns are only weakly explained by conventional systematic-risk factors, a pattern that motivates the search for alternative transmission mechanisms. Section 3.3 introduces the thesis's central theoretical contribution: the wealth-channel hypothesis, a formally specified three-stage transmission model linking monetary policy to art prices through the intermediary of ultra-high-net-worth financial wealth. Section 3.4 extends the framework to a regime-dependent specification, arguing that the wealth channel operates asymmetrically across macroeconomic states. Section 3.5 confronts the selection-bias and illiquidity problems that any honest analysis of art returns must address. Throughout, models are presented with complete mathematical derivation, parameters are grounded in published empirical estimates, and numerical illustrations are traceable.

3.1 Art in Portfolio Theory: The Markowitz Foundation

The starting point for any analysis of art as an investable asset is the mean-variance optimization framework of Markowitz (1952). The framework asks a deceptively simple question: given a universe of assets with known expected returns, variances, and covariances, what portfolio weights minimize risk for a given target return? Consider an investor allocating wealth across n assets, with w the nl vector of portfolio weights, S the variance-covariance matrix, and p the expected-return vector. The investor solves:

Abb. in Leseprobe nicht enthalten

The Lagrangian solution yields the efficient frontier, the set of portfolios offering the maximum expected return for each level of variance. To evaluate art's role in this framework, three quantities are required: its expected return, its volatility, and its correlation with traditional assets.

3.1.1 Art’s Risk-Return Profile

Table 3.1 synthesizes consensus estimates from the leading studies, all expressed as annualized real USD figures.

Abb. in Leseprobe nicht enthalten

Table 3.1: Cross-Asset Risk-Return Summary. Sources: Renneboog & Spaenjers (2013); Korteweg, Kräussl & Verwijmeren (2016); Li, Ma & Renneboog (2022); UBS Art Basel Report (2025); CAIA (2024); Damodaran historical returns database. All figures annualized, real USD.

Sharpe ratios are reported as published in the source studies and are computed against the contemporaneous real risk-free rate of each sample window (mean U.S. real T-bill yield 1972­2010 was approximately 1.7-5.7 percent, depending on study sub-window); they are not directly recomputable from the return-and-volatility columns above without the original Rf series. Volatility for the selection-corrected art series is reported as 15.30 percent following Korteweg, Kraussl and Verwijmeren (2016, Table V), which is the figure cited verbatim in §2.1.2 and §3.5.

3.1.2 Worked Example: The Diversification Benefit of Art

To demonstrate art's diversification potential quantitatively, consider a simplified two-asset portfolio comprising U.S. equities and art, with p E = 7.10%, p A = 3.97%, o E = 16.55%, aA = 15.21%, and p_EA = 0.10. The portfolio variance for weight w allocated to art is:

Abb. in Leseprobe nicht enthalten

Evaluating at w = 0.10 gives o p = 16.42% and p p = 6.79%. Assuming R_f = 2.00%, the Sharpe ratios compare as follows:

Abb. in Leseprobe nicht enthalten

Table 3.2: Two-Asset Portfolio Optimization Results.

The volatility reduction of 13 basis points from a 10% art allocation is real but modest, a consequence of the genuinely low equity-art correlation. The Sharpe ratio nonetheless declines because art's lower expected return dominates the marginal variance reduction. In practice, four additional frictions further erode the theoretical case. Round-trip auction transaction costs of 25­36% impose, at a seven-year average holding period, an annual drag of approximately 360-510 basis points, reducing net real returns to between -0.5% and +1.5%. Illiquidity imposes a further 120-200 bps annualized discount (Korteweg, Kraussl and Verwijmeren, 2016). Individual blue­chip artworks cost $5-50 million, making diversification within art prohibitively expensive for all but the largest family offices. And selection bias, the subject of Section 3.5, reduces the observed Sharpe ratio from 0.27 to 0.11 once corrected (Korteweg, Kraussl, and Verwijmeren, 2016).

The key takeaway for the thesis is that the Markowitz framework alone does not justify art allocation for a return-maximizing investor once frictions are incorporated. This tension motivates the central question: if pure return optimization does not explain the $1.7 trillion allocated globally to art and collectibles (Deloitte, 2024), what does? The answer lies in the wealth channel: a demand-side mechanism driven not by risk-return optimization but by the spending behavior of ultra-high-net-worth individuals whose financial wealth fluctuates with monetary policy.

3.2 The CAPM and Multi-Factor Extensions for Art

Before introducing the wealth channel, it is essential to establish whether art returns are explained by exposure to systematic risk factors. If they are, art is simply another risky asset and the standard factor model suffices. If a persistent unexplained alpha remains, this constitutes evidence of market segmentation, and justifies the search for alternative transmission mechanisms. Following Sharpe (1964), the excess return on art is modeled as a linear function of the equity market risk premium:

Abb. in Leseprobe nicht enthalten

Based on the existing empirical literature, three parameter expectations emerge. First, art's beta is expected to be low (0 ~ 0.15-0.40): Goetzmann (1993) documents betas in this range for aggregate art indices, confirming that art is not a leveraged equity play but is driven primarily by idiosyncratic and demand-side factors. Second, alpha is generally small (a ~ 0.5-2.5% annually) and often statistically insignificant at annual frequency. Third, and most important, the R-squared is extremely low (~0.03-0.15), confirming that 85-97% of art return variation is not explained by equity market risk. This is the empirical foundation for the claim that art is a partially segmented asset class.

The Fama-French extension augmented with Carhart's (1997) momentum factor generalizes the specification:

Abb. in Leseprobe nicht enthalten

Each loading carries an economic interpretation: the MKT loading captures wealth-channel co-movement with equity markets; the SMB loading reflects art's shared illiquidity and high idiosyncratic risk with small-cap equities; the HML loading reflects art's character as a tangible no-cash-flow "value" asset; the MOM loading tests for trend-following in art markets. The critical prediction is that even after including four factors, the R[2] should remain below 0.15, providing formal evidence that standard asset-pricing models are insufficient to explain art returns and creating the econometric justification for introducing the wealth channel as a supplementary, and potentially dominant, explanatory mechanism. This prediction is borne out in Chapter 5.

3.3 The Wealth-Channel Hypothesis: The Novel Contribution

The argument advanced in this thesis is constructed in three formally specified stages, each representing a link in the causal chain from monetary policy to art prices. Unlike prior studies that document correlations between macro variables and art returns, this framework specifies the mechanism through which monetary policy transmits to art demand, and generates testable empirical predictions that distinguish it from competing explanations.

3.3.1 Stage 1: Monetary Policy ^ Financial Wealth

The first link is well established in the macroeconomics literature. Expansionary monetary policy inflates financial asset values through three concurrent channels. The discount-rate channel mechanically raises the present value of equities and bonds when policy rates fall; Bernanke and Kuttner (2005) estimate that an unanticipated 25 bps Fed funds rate cut raises the S&P 500 by approximately 1.0-1.3% on the announcement day. The portfolio-rebalancing channel forces investors out of safe assets into riskier alternatives; Gagnon et al. (2011) document that QE1 reduced the 10-year Treasury yield by 58-91 bps, triggering substantial rebalancing into equities and alternative assets. The wealth-distribution channel, documented by Montecino and Epstein (2015) and Ampudia et al. (2018), demonstrates that the resulting wealth gains accrue disproportionately to the top 1% precisely because this group holds a larger share of financial assets.

The monetary-policy-to-wealth link is specified as:

Abb. in Leseprobe nicht enthalten

where AW UHNW,t is the change in aggregate UHNWfinancial wealth (proxied by household net worth from FRED Z.1), Ar t is the change in the effective Federal Funds Rate (or Wu-Xia shadow rate during ZLB periods), AQE t is the change in the Federal Reserve balance sheet, and Z t is a vector of controls. Standard New Keynesian transmission theory predicts Si < 0 (rate cuts increase wealth) and S2 > 0 (balance sheet expansion increases wealth); both are confirmed in the macro literature.

3.3.2 Stage 2: Financial Wealth ^ Art Demand

The second link formalizes the relationship between UHNW financial wealth and art demand at auction. Art is a luxury good with income (and wealth) elasticity substantially greater than one. Art demand is modeled as D art,t = /(W UHNW,t, P art,t, Z t), with dD/dW > 0 and wealth elasticity s_W > 1 (expected range 1.5-3.0). The wealth elasticity can be empirically calibrated using aggregate market data over 2009-2021: the global UHNW population grew from approximately 78,000 to 295,450; U.S. household net worth of the top quintile rose from $42.5T to $110.4T (a +159.8% change); and global art auction sales rose from $4.6B to $17.1B (a +271.7% change). The implied wealth elasticity is therefore:

Abb. in Leseprobe nicht enthalten

A wealth elasticity of 1.70 is consistent with art being a luxury good in the formal economic sense: a 1% increase in UHNW financial wealth is associated with a 1.70% increase in global art auction sales. This is consistent with the Ait-Sahalia, Parker and Yogo (2004) framework, which predicts that luxury consumption goods exhibit elasticities in the range of 1.5-3.5 depending on the good’s exclusivity and the concentration of its buyer base. The figure of 1.70 is presented as an illustrative back-of-envelope calibration rather than as an econometrically estimated elasticity: it is computed as the ratio of two cumulative growth rates over a single twelve-year window, both of which are dominated by trend, and it does not control for the price level, transaction costs, or the lagged adjustment dynamics that the empirical chapter formalizes. The figure is included to motivate the wealth-elasticity assumption used in the structural Chapter 3 framework; it is not itself a test of the wealth-channel hypothesis. The proper econometric estimate of the wealth coefficient is the one reported in Chapter 5, with appropriate small-sample inference.

3.3.3 Stage 3: The Lag Structure

A critical feature distinguishing art from financial assets is that art prices do not adjust instantaneously to monetary shocks. While equities process new information within minutes, art prices adjust with a lag of approximately 6-18 months. This lag is not arbitrary but arises structurally from three features of the art market. First, the discrete auction calendar concentrates major evening sales in fixed biannual windows (May and November in New York; February/March and June/October in London); a monetary shock in January cannot affect auction

The Wealth Channel of Monetary Policy prices until the May sales at the earliest, imposing a structural minimum lag of 4-5 months. Second, the consignment decision lag adds 3-6 months between a wealth shock and the moment at which a collector consigns works through the auction house process; Lovo and Spaenjers (2018) model this as an optimal timing problem under uncertainty about future price levels. Third, sequential price discovery in thin markets requires that several comparable works sell at elevated prices before the market "confirms" a new price level; this process can take 2-3 auction seasons (12-18 months).

Practitioner testimony reinforces the structural character of the lag. In an interview conducted for this thesis with Tom Legh, Head of Valuations at Christie's, the auction calendar and consignment decision microstructure were described in terms that closely match the formal model: "Art prices don’t move with the first whisper of a rate cut. People need eighteen months, sometimes two years, before a change in their financial position really shows up in how they bid. The decision to consign a serious work is emotional as much as financial, and the calendar of our sales imposes its own rhythm, nobody consigns a Cézanne in March because the Fed pivoted in February" (T. Legh, personal communication, March 16, 2025). This statement, from a senior practitioner with first-hand knowledge of consignment flow at the world's largest auction house, supports the interpretation that the observed 6-18-month lag is a structural feature of the market microstructure rather than an econometric artifact of low-frequency data.

The lag structure is formalized as a distributed-lag specification following Pesaran, Shin and Smith (2001):

Abb. in Leseprobe nicht enthalten

where L is the maximum lag length selected by information criteria. The cumulative long-run multiplier, 0 LR = Xk 0_k, gives the total cumulative effect of a one-unit monetary policy shock on art returns over all lag horizons and is the key parameter for testing H1.

Abb. in Leseprobe nicht enthalten

Figure 3.1: The Wealth-Channel Transmission Chain from Monetary Policy to Art Prices.

3.3.4 The Mediation Test: Empirical Identification Strategy

To formally test the wealth-channel hypothesis, this thesis employs a mediation framework inspired by Baron and Kenny (1986). The test requires three regressions run sequentially: the total effect of monetary policy on art returns, the effect of monetary policy on the mediator (wealth), and art returns on both monetary policy and the mediator simultaneously:

Abb. in Leseprobe nicht enthalten

Abb. in Leseprobe nicht enthalten

Abb. in Leseprobe nicht enthalten

Mediation is established if four conditions hold: 0 is significant in (a) (monetary policy affects art returns); 5 is significant in (b) (monetary policy affects wealth); y is significant in (c) (wealth affects art returns); and |0‘| < |0| (the direct effect shrinks when wealth is included). The indirect (mediated) effect 67 can be tested via the Sobel (1982) z-statistic or via bootstrap confidence intervals. Full mediation occurs if 0‘ becomes statistically insignificant; partial mediation occurs

The Wealth Channel of Monetary Policy if |P'| < |P| but P‘ remains significant. For a thesis at annual frequency with limited degrees of freedom, partial mediation is the most likely finding and is sufficient to support the hypothesis. (As Chapter 5 will show, the formal Baron-Kenny test in fact fails in this sample for reasons related to the well-known "controlling-away-the-mechanism" problem, a matter to which Chapter 6 returns.)

3.4 Regime Dependence: When the Wealth Channel Switches Off

The wealth-channel hypothesis, as stated in Section 3.3, implicitly assumes a stable relationship between monetary policy, financial wealth, and art prices. This section relaxes that assumption and argues that the transmission mechanism is regime-dependent. During financial stress, four mechanisms attenuate or reverse the wealth channel. Wealth destruction: the S&P 500 fell 56.8% peak-to-trough during the 2007-2009 crisis, and global art auction sales fell 36% in 2009 (Art Basel/UBS). Risk aversion and flight to liquidity: during crises, even wealthy collectors shift toward liquid safe-haven assets. Forced selling: collectors facing margin calls or estate obligations consign works at inopportune times, producing supply shocks that depress prices. Credit-channel impairment: art-secured lending contracts during crises as lenders tighten underwriting, removing a source of leveraged art buying.

The regime-dependent model takes the general form R art,t = a s + P s'X t + s_t with s G {expansion, stress}. The thesis employs three approaches to regime identification. The primary approach is a threshold regression using the VIX, with the threshold set at VIX = 25 following the convention in the macro-finance literature:

Abb. in Leseprobe nicht enthalten

where D t = 1(VIX t > 25). A Wald test of Ho: Pi = yi = 0 tests the joint hypothesis that the relationships are regime-independent. Approach B uses NBER recession dating as an exogenously determined robustness check; Approach C employs a two-state Markov-switching model following Hamilton (1989) for endogenous regime identification. The regime-dependent framework generates testable cross-sectional predictions across art market segments: Contemporary and Post-War art, whose buyer base consists disproportionately of newly wealthy UHNW collectors, should exhibit stronger pro-cyclical wealth-channel sensitivity than Old Masters and Impressionists, whose buyer base is more heavily composed of established collectors and museum-adjacent institutions.

Abb. in Leseprobe nicht enthalten

Table 3.3: Segment-Level Wealth-Channel Predictions.

3.4.1 Microfoundation: Loss Aversion and the Asymmetric Wealth Elasticity

The regime-dependent specification above is reduced-form: it treats the asymmetry between expansion- and stress-regime wealth coefficients as an empirical feature to be tested. To strengthen the theoretical case it is useful to derive the asymmetry from a primitive utility-theoretic foundation. Let the representative ultra-high-net-worth collector have the Kahneman-Tversky (1992) value function defined over wealth changes AW relative to a reference point W:

Abb. in Leseprobe nicht enthalten

with curvature parameter a ~ 0.88 and loss-aversion parameter X ~ 2.25, both estimated from large-scale experimental data and replicated in dozens of subsequent studies. Art expenditure is a discretionary luxury outlay whose first-order condition equates the marginal utility of an additional artwork with the shadow price of capital diverted from other uses. Under this setup, the wealth elasticity of art expenditure is asymmetric by construction: a one-unit wealth gain generates an art­expenditure response proportional to a, while a one-unit wealth loss generates a response proportional to k-a. The ratio of stress-regime to expansion-regime wealth elasticity for a single representative collector is therefore exactly X ~ 2.25.

A confounding force runs in the opposite direction at the aggregate level. During downturns, forced sellers are rarer in art than in other asset classes because high-end collections are overwhelmingly held debt-free; when forced selling does occur, the selling pressure dominates the demand contraction. Let f_forced denote the fraction of marginal market activity driven by forced

The Wealth Channel of Monetary Policy selling rather than discretionary trading. The net regime asymmetry observed in aggregate data is approximately:

Abb. in Leseprobe nicht enthalten

where £, >1 captures the price-pressure intensity of forced selling. Practitioner testimony (Legh, 2025) indicates fforced ~ 0 for Old Masters and forced ~ 0.10-0.20 for Contemporary art during acute stress. Substituting these values yields predicted asymmetries of approximately 2.25 for Old Masters and 1.8-2.0 for Contemporary, both consistent with the empirical findings reported in Chapter 5. Equation 3.11 also generates three sharper predictions tested in Chapter 5: (i) the wealth coefficient is larger in magnitude during expansions than during contractions; (ii) the asymmetry is attenuated in segments with more forced selling; and (iii) the asymmetry is amplified in segments with no forced selling.

3.4.2 Formalized Segment Theory: Buyer-Type Decomposition

The cross-sectional prediction that wealth sensitivity should differ across art-historical segments was stated in Section 3.4 as a stylized claim grounded in differing buyer composition. It is useful to derive the prediction more formally. Let total demand for art in segment s at time t be the integral of individual demand q(w, s, t) over the wealth distribution F_s(w) of buyers active in that segment:

Abb. in Leseprobe nicht enthalten

Under a standard iso-elastic specification q(w, s, t) = c • wA{n s} • X tA{P_s}, the aggregate wealth elasticity of segment s is the wealth-weighted average:

Abb. in Leseprobe nicht enthalten

Equation 3.13 reduces to the constant n when F s is degenerate and rises above n as F s becomes more concentrated in the right tail. The prediction follows: the more heavily a segment's buyer base is concentrated among the ultra-wealthy, whose financial-asset positions are most sensitive to monetary policy, the higher the segment's reduced-form wealth coefficient. Practitioner testimony provides an ordering consistent with this prediction: "The Old Masters market is insulated in a way the Contemporary market is not. These works usually sit in collections that have existed for three or four generations, and the owners are almost never forced sellers. In a recession, a hedge-fund buyer may need to liquidate a Basquiat; a family that has owned a Rubens

The Wealth Channel of Monetary Policy since the 1920s will not. That asymmetry is why the old categories look so much less cyclical in the data, and it is not because the macroeconomy has stopped mattering, it is because the supply side does not respond" (Legh, 2025).

Two complications must be acknowledged before this framework is taken to data. First, the relevant wealth proxy differs across segments: Old Masters buyers skew toward old-money families whose wealth is partly in real estate, closely held businesses, and inherited assets that move imperfectly with public markets, while Contemporary buyers skew toward newly wealthy hedge-fund and tech-founder collectors whose wealth is heavily in public equities and therefore co-moves more tightly with the Federal Reserve's Distributional Financial Accounts top-1% measure used in this thesis. The estimated wealth coefficient using a single wealth proxy will therefore understate Old Masters' true sensitivity to old-money wealth and overstate the gap between the two segments' true elasticities. Second, the supply-side mechanism described by Legh (2025) operates orthogonally to the demand-side wealth elasticity: even if Old Masters demand is wealth-elastic, the segment's return may be insensitive to wealth shocks because the supply curve is nearly vertical when forced selling is absent. The empirical chapter must therefore interpret a flat or non-monotonic segment gradient with care: the absence of a wealth gradient in observed returns does not imply the absence of a wealth elasticity in latent demand.

3.5 Selection Bias and the Illiquidity Problem

No theoretical framework for art as an investable asset is complete without confronting the most important data limitation in the field: selection bias in observed art returns. Art return indices are constructed exclusively from observed auction transactions, but the decision to sell is endogenous to returns. Artworks that have appreciated substantially are disproportionately likely to be resold; artworks that have depreciated are more likely to be withdrawn or sold privately. This creates a systematic upward bias in observed returns. Using a Bayesian MCMC selection model on 20,538 repeat-sales pairs over 1972-2010, Korteweg, Kraussl and Verwijmeren (2016) demonstrate that correcting for selection bias reduces annualized real returns from 8.70% to 6.30%, volatility from 19.80% to 15.30%, and Sharpe ratios from 0.27 to 0.11, a deflation of 59%.

This thesis does not attempt to implement the full Korteweg MCMC correction, which requires artwork-level transaction data beyond the scope of a bachelor's thesis. Instead, three mitigation

The Wealth Channel of Monetary Policy strategies are adopted. First, transparency: all results are presented with the explicit caveat that art return indices suffer from selection bias and that true returns are likely lower than observed. Second, focus on relative rather than absolute effects: the core contribution, the wealth-channel mediation test and regime-dependent analysis, does not depend on the absolute level of art returns but on the relative sensitivity of art returns to monetary policy and wealth across time and regimes. Provided the selection mechanism is approximately stable over time, it inflates the level of returns but is unlikely to create spurious time-series co-movement patterns. Third, robustness via multiple indices: testing results across the Artprice Global Index and the Artprice100 blue-chip index, each with its own selection properties, assesses robustness to construction methodology.

3.6 Synthesis: From Theory to Testable Predictions

The theoretical framework developed in this chapter generates a set of precise, testable predictions summarized in Table 3.4.

Abb. in Leseprobe nicht enthalten

Table 3.4: Master Hypothesis Summary: Theoretical Predictions and Empirical Tests.

The framework is deliberately modular. The baseline CAPM analysis establishes the empirical premise of market segmentation. The wealth-channel mediation test constitutes the novel contribution. The regime-dependent analysis extends the hypothesis into a richer, more realistic framework. The segment-level decomposition generates cross-sectional predictions that sharpen explanatory power. At each stage, the analysis is self-contained: even if the more advanced specifications prove infeasible due to data constraints, the core argument remains testable. Chapter 4 operationalizes these models; Chapter 5 reports the empirical results.

The Wealth Channel of Monetary Policy

CHAPTER 4

DATA AND METHODOLOGY

This chapter documents the data architecture and econometric methodology used to test the wealth-channel hypothesis. The empirical strategy is designed as a five-layer architecture: baseline OLS with HAC standard errors; an ARDL specification to capture the 6-18 month transmission lag; a Baron-Kenny mediation test; a threshold regression for regime dependence; and a segment­level decomposition. The modular design ensures that the core argument remains defensible even if the more advanced layers prove infeasible under small-sample constraints.

The sample period runs from 2000 to 2024 (T = 25 annual observations), chosen to align with the availability of consistent Artprice index data while capturing three distinct monetary regimes: the post-dot-com easing cycle (2001-2004), the Global Financial Crisis and its aftermath including three rounds of quantitative easing (2007-2015), and the COVID-era ultra-accommodative stance followed by the most aggressive tightening cycle in four decades (2020-2024). A semiannual robustness specification (T = 50) is also estimated where data permit.

4.1 Art Return Data

The primary dependent variable is the Artprice Global Index, a repeat-sales regression index published by Artmarket.com, denominated in U.S. dollars with base 100 in January 2000. Coverage extends to over 804,000 lots sold annually across more than 6,300 auction houses in 72 countries. The index encompasses paintings, sculpture, drawings, photography, prints, and mixed media. The index is constructed using hedonic repeat-sales methodology, the industry standard for art return measurement, with the underlying methodology peer-reviewed in academic literature.

Annual index levels are extracted from Artprice's publicly released annual reports, cross­verified against the official Artprice press releases and third-party financial media. Table 4.1 reports the year-by-year index values and computed returns used in the primary specification.

Abb. in Leseprobe nicht enthalten

Table 4.1: Artprice Global Index: Year-End Levels and Annual Returns (Base 100, January 2000).

A secondary dependent variable, the Artprice100 blue-chip index, simulates investment in the 100 top-selling artists weighted by five-year trailing auction revenue. Launched in 2018 with retrospective calculation back to January 2000, the Artprice100 grew +405% from 2000 to 2020 (8.2% annualized) and reached +609% by January 2023 (8.9% annualized). Because it excludes illiquid artists and mid-market lots, the Artprice100 isolates the high-end market segment most

The Wealth Channel of Monetary Policy likely to be influenced by UHNW collector wealth. An important methodological clarification is in order. The Artprice100 is constructed from artists who are themselves identified by past auction revenue, which embeds survivorship by construction. The thesis does not construct any artist-level tier or ranking of its own; the only ranked-artist series used is the Artprice100, which is a pre­existing index from Artmarket.com with its own published methodology. Results using this index are reported as a robustness specification only and are interpreted as applying to the high-end market segment rather than to the broad art market. The macro specifications use the Artprice Global Index as the primary dependent variable; this index is constructed from aggregate auction­market data and does not condition on any ex-post artist-ranking variable.

All repeat-sales art indices suffer from well-documented selection, survivorship, and aggregation biases (Korteweg, Kraussl and Verwijmeren, 2016). These inflate the level of measured returns by approximately 240 basis points per year but, provided the selection mechanism is approximately stable, are unlikely to distort the time-series co-movement with macro variables that is the focus of this thesis. Section 5.10 implements a Korteweg-Kraussl- Verwijmeren sensitivity analysis that rescales the wealth coefficient under five assumed selection­bias factors; the economic conclusion is robust at the benchmark factor of 0.70, where the wealth coefficient remains +4.85 and is of the same order of magnitude as the unadjusted estimate. The implication is that selection bias affects the level of measured returns substantially but is unlikely to affect the time-series co-movement structure on which the wealth-channel inference depends. Log returns on the art index are computed as r art,t = ln(P t) - ln(P_{t-1}). Log returns are preferred over arithmetic returns because they are time-additive and approximately normally distributed for moderate magnitudes (Campbell, Lo and MacKinlay, 1997).

4.2 Macro-Financial Data

All macro-financial variables are sourced from publicly available institutional databases. No proprietary or paywalled data are required. The effective Federal Funds Rate (FRED: FEDFUNDS) is the primary conventional monetary policy instrument, with December values used for annual specification. The Wu-Xia shadow rate, available from the Federal Reserve Bank of Atlanta, extends the FFR below the zero lower bound using a term-structure model; it reached a trough of approximately -3.0% in mid-2014, reflecting cumulative QE3 accommodation. A

The Wealth Channel of Monetary Policy composite monetary policy rate is constructed by splicing the Wu-Xia shadow rate (for months when the FFR target range is 0-25 bps) with the effective FFR (for all other months). The Federal Reserve balance sheet (FRED: WALCL) captures the quantity of unconventional accommodation, entering the regression in log-difference form Aln(BS).

U.S. household net worth (FRED: BOGZ1FL192090005Q) is the primary wealth proxy: total market value of all financial and real assets held by U.S. households minus total liabilities, published quarterly by the Federal Reserve Board, standing at approximately $156.2 trillion at Q4 2023. The log change Aln(W) is the mediator in the Baron-Kenny test. As a supplementary wealth measure, UHNW population counts are hand-collected from Knight Frank Wealth Reports (2007­2025) and Capgemini World Wealth Reports for pre-2007 data. S&P 500 total returns and Fama- French factor returns are from the Kenneth French Data Library. The VIX (FRED: VIXCLS) measures 30-day implied volatility; a VIX level above 25 demarcates "high-stress" regimes (Whaley, 2009). CPI inflation (FRED: CPIAUCSL), the 10-year Treasury yield (FRED: DGS10), and the yield curve slope (FRED: T10Y2Y) serve as controls and alternative regime indicators.

4.3 Variable Transformations and Descriptive Statistics

Variables that are I(1) in levels enter the regressions in first-differenced or log-differenced form; variables that are I(0) enter in levels. The testing protocol employs three complementary tests (ADF, Phillips-Perron, KPSS) following Elder and Kennedy (2001). This is consistent with the Pesaran and Shin (1999) ARDL framework, which permits a mix of I(0) and I(1) variables provided none is I(2). Table 4.2 reports descriptive statistics for the regression-ready variables.

Abb. in Leseprobe nicht enthalten

Table 4.2: Descriptive Statistics of Key Variables (Annual, 2000-2024, T = 25).

The Artprice Global Index return r(art) has a mean of 1.31% per annum, consistent with the well-documented finding that broad art market returns after transaction costs are modest, substantially below the blue-chip segment mean of 7.29%. The large standard deviation (10.39%) and negative minimum (-19.87% in 2022) confirm that art is a volatile asset class. Aln(BS) is the most volatile series due to unprecedented QE expansions in 2008-2009 and 2020-2021.

4.4 The Five-Layer Estimation Architecture

4.4.1 Identification Strategy

This subsection makes explicit the identification claims this thesis does and does not make, addressing the well-known endogeneity of monetary-policy variables to the macroeconomic cycle. The fundamental econometric concern is that the Federal Reserve does not set its policy rate randomly. Rate cuts are concentrated in periods of weak growth and falling asset prices; rate hikes are concentrated in periods of strong growth and rising asset prices. Any contemporaneous correlation between policy-rate changes and art returns, therefore, confounds the wealth-channel transmission this thesis posits with the broader cyclical co-movement of policy and asset markets. Under this confounding, ordinary least squares applied to equation (4.1) is inconsistent: plim(P) f p. The estimated coefficient on ASR, is therefore best understood as the conditional association between the policy variable and art returns given the included controls, not as a causal effect in the potential-outcomes sense.

The gold standard for identification of monetary-policy effects is high-frequency surprise­shock methodology. Romer and Romer (2004) construct narrative-identified policy shocks; Nakamura and Steinsson (2018) extract interest-rate surprises from intraday Treasury futures around FOMC announcements; Jarocinski and Karadi (2020) decompose surprises into pure policy and information components using sign restrictions on policy and stock-market reactions. Any of these would, in principle, allow recovery of the structural policy parameter free of the cyclical confound. None is feasible in the present empirical strategy: the surprise series exist at daily and intraday frequency, the identification logic requires aggregation to monthly or quarterly horizons,

The Wealth Channel of Monetary Policy and an annual aggregation with T = 25 leaves insufficient degrees of freedom for a meaningful first-stage instrument. The thesis therefore defers an instrumental-variable specification to the quarterly extension flagged in §7.3.

In place of an instrumental-variable strategy, the thesis assembles what may be termed a weak identification bundle. Three pieces of evidence collectively constrain the space of explanations for the observed wealth-art association more tightly than a raw correlation would. First, the Baron- Kenny mediation framework (§4.4.4) decomposes the policy-art association into a direct and an indirect effect operating through the wealth measure; although the formal test is underpowered at T = 26 (§5.5), the structural decomposition itself rules out the simplest alternative explanation in which policy moves art prices through a channel that does not flow through wealth. Second, the lag prediction of equation (3.7), six to eighteen months from policy shock to art-price response, is falsifiable: the data either exhibit the predicted lag pattern or they do not. As reported in §5.4, they do, and the lag pattern is independently corroborated by interview testimony on auction-house consignment microstructure (Legh, 2025). Third, the out-of-sample 2022-2024 prediction (§6.3.3) is a falsifiable test against data the model did not see during estimation; a model whose coefficients reflect coincident cyclical co-movement rather than the wealth-channel mechanism has no reason to deliver an accurate three-year-ahead forecast.

Three confounders are not ruled out by this bundle and are acknowledged explicitly. (a) A common risk-appetite factor, the VIX, the high-yield credit spread, or a latent sentiment index, drives both wealth-share variation and art demand. The thesis tests this in §5.12 and finds the wealth coefficient does not collapse when the VIX is added; this disfavors but does not refute the common-factor story. (b) The Federal Reserve’s Z.1 household net-worth measure includes the market value of household art holdings, creating a small mechanical positive co-movement between the wealth measure and the art return series. This is addressed by adopting the Federal Reserve’s top-1% wealth share, which strictly excludes art holdings, as the primary wealth proxy in §§5.5-5.7; the wealth coefficient is robust to the change of proxy. (c) Selection bias in the Artprice index. This is bounded via the Korteweg-Kraussl-Verwijmeren rescaling exercise in §5.10, which shows the wealth coefficient remains positive and of economically meaningful magnitude at every plausible bias factor.

The interpretive convention adopted from Chapter 5 onward is therefore that all reported coefficients are conditional associations under the specified control set, and that the wealth­channel hypothesis is supported by the convergent evidence of (i) the structural micro-foundation in §3.3, (ii) the lag prediction confirmed in the data and in practitioner testimony, (iii) the regime­dependent sign pattern in §5.6, and (iv) the out-of-sample prediction in §6.3.3, taken together. None of these on its own constitutes causal identification; their convergence is more demanding than raw correlation but weaker than IV. The thesis argues that this is the strongest empirical claim a bachelor-level analysis at annual frequency can defensibly support, and is explicit where it falls short of the gold standard.

Figure 4.1 below visualizes the identification architecture as a directed acyclic graph (DAG). Solid arrows denote pathways the wealth-channel framework posits and the empirical strategy attempts to identify; dashed arrows denote confounding pathways the included controls partially block; the unblocked confounders enumerated in (a)-(c) above are flagged with shaded nodes. The DAG is intended as an interpretive aid, not a formal causal-graph analysis in the Pearl (2009) sense; its purpose is to make the identification claims of the thesis visible at a glance and to acknowledge the residual identification gap relative to an IV strategy.

Abb. in Leseprobe nicht enthalten

Figure 4.1: Directed Acyclic Graph of the Wealth-Channel Identification Architecture. Solid arrows denote the hypothesized causal pathway from monetary policy through financial wealth to art demand and art returns. Dashed arrows denote confounding pathways from risk appetite, the business cycle, and selection bias, all partially blocked by the controls in the empirical specification.

4.4.2 Layer 1: Baseline OLS with HAC Standard Errors

The baseline specification regresses annual art returns on the vector of macro-financial variables:

Abb. in Leseprobe nicht enthalten

OLS standard errors are computed using the Newey-West (1987) HAC estimator with the T/(T-k) finite-sample adjustment factor (Andrews, 1991) and t-distribution inference with (T-k) degrees of freedom. The HAC bandwidth is set to one for the headline specifications, which is the natural choice for annual data with limited residual autocorrelation (the Durbin-Watson and Breusch-Godfrey diagnostics in §5.3 confirm low residual serial correlation). As a robustness check, the same regressions are re-estimated with the automatic bandwidth selection of Newey and West (1994), which chooses the bandwidth as a data-driven function of the residual autocorrelation function; the resulting standard errors are within rounding tolerance of the bandwidth-one specifications, confirming that the inferences are not bandwidth-sensitive. The automatic-bandwidth results are reported in Appendix B. Diagnostics include R[2] and adjusted R[2]; the joint-significance F-statistic; Durbin-Watson; Breusch-Godfrey LM test for higher-order autocorrelation; White’s heteroskedasticity test; Ramsey RESET; and variance inflation factors. The diagnostic suite is extended in the revised draft to include the Breusch-Pagan and Jarque-Bera tests (residual heteroskedasticity and normality), the Ljung-Box Q-statistic (residual autocorrelation), Cook’s Distance and DFBETAS (influence diagnostics), and the maximum hat­value (leverage); all are reported in Table 5.3 in §5.3 in place of the previous diagnostic prose paragraph.

4.4.3 Layer 2: ARDL Model

The ARDL specification (Pesaran, Shin and Smith, 2001) captures both contemporaneous and lagged effects of monetary policy on art returns. Given T = 25 and annual frequency, L max = 2 is the practical upper bound. The cumulative long-run multiplier Xf k measures the total effect of a

The Wealth Channel of Monetary Policy one-percentage-point monetary easing on art returns over the entire transmission horizon. An impulse-response function is constructed by plotting the cumulative sum of p k across horizons, with 95% confidence bands from the HAC covariance matrix.

4.4.4 Layer 3: Wealth-Channel Mediation Test

A formal Baron-Kenny (1986) mediation test is conducted in three steps: total effect, first stage (policy on wealth), and second stage (policy and wealth on art). Mediation is established if the Step 2 coefficient is significant, the Step 3 wealth coefficient is significant controlling for policy, and the direct-effect coefficient attenuates toward zero in Step 3. Significance of the indirect effect axb is assessed using the Sobel z-statistic and, as a small-sample robustness check, a bootstrap mediation test (Preacher and Hayes, 2008) with 5,000 replications.

4.4.5 Layer 4: Regime-Dependent Analysis

Two approaches are employed. The primary approach is a threshold regression with D t = 1 if VIX t > 25: all regressors are interacted with the regime dummy, and a Wald test of Ho: all interaction coefficients = 0 provides the formal test. The threshold of 25 corresponds approximately to the 75th percentile of the VIX distribution and has been used as a stress demarcation in the volatility literature (Whaley, 2009). Approach B uses NBER recession dating as a robustness check. A two-state Markov-switching model, following Hamilton (1989), is specified for completeness but is acknowledged as exploratory given T = 25. A complementary interaction-term specification is also estimated. Where the threshold regression splits the sample by regime and reports separate coefficients for each subsample, the interaction specification estimates the regime asymmetry as a single coefficient on a multiplicative term:

Abb. in Leseprobe nicht enthalten

where D_stress is the stress-regime indicator, Aln(W)_t is the wealth growth term, and X_t is the vector of remaining controls. The coefficient p? is a direct test of the regime asymmetry: under the null of regime independence, p? = 0; under the wealth-channel hypothesis, p? < 0 (the wealth coefficient is smaller in stress than in expansion). Reporting the asymmetry as a single t-test on p? is more transparent than the two-subsample comparison adopted by the threshold regression and the use of interaction terms tests heterogeneity directly. Equation (4.6) is reported alongside Table 5.7 in §5.6 as a robustness check. A second interaction specification, ASR_t x sign(ASR_t),

The Wealth Channel of Monetary Policy separates positive (tightening) and negative (easing) policy shocks and tests for asymmetric monetary-policy transmission, following Tenreyro and Thwaites (2016) and Angrist, Jordà and Kuersteiner (2018). If easing and tightening shocks have asymmetric effects on art returns, the sign-interaction coefficient will be statistically distinguishable from zero; this is reported as a second-order finding in §5.9.

4.4.6 Layer 5: Segment-Level Decomposition

To test H4, the Layer 1-3 specifications are re-estimated separately for each art-historical segment provided directly by the Artprice Econometrics Department: nineteenth-century, Modern (1860-1945), Post-War (1945-1970), and Contemporary (post-1970). The four-segment partition follows Artprice’s official chronological classification, which mirrors the catalog conventions of Sotheby’s, Christie’s, and Phillips for sale organization, and is therefore exogenous to the empirical strategy adopted in this thesis. The cut dates (1860, 1945, 1970) correspond to canonical art-historical transitions: the emergence of modernism, the post-war art market, and the contemporary art market. Section 5.7 reports a robustness specification with a two-segment cut (pre-1945 versus post-1945) to confirm that the qualitative pattern is not an artifact of the four­way classification. The testable prediction is a monotonically increasing pattern of wealth-channel coefficients from the nineteenth century (lowest) to the contemporary (highest). Cross-equation equality of coefficients is tested using a seemingly unrelated regression (SUR) framework.

4.4.7 Specification Parsimony, Information Criteria, and Overfitting Diagnostics

With T = 25 annual observations, parameter parsimony is a binding methodological constraint, not a stylistic preference. Layer 1 is therefore positioned as the primary inferential specification, reported with the minimum regressors needed to identify the wealth and policy channels. Layers 2-5 are auxiliary and interpretive: Layer 2 (ARDL) captures the lag structure that Layer 1 collapses to a contemporaneous specification; Layer 3 (Baron-Kenny) is a structural decomposition rather than a power-driven hypothesis test; Layer 4 (threshold regression) tests regime-dependence; Layer 5 (segment decomposition) tests cross-sectional heterogeneity. Each layer answers a distinct economic question; none is redundant with Layer 1. Model comparison across layers uses Akaike (AIC), Bayesian (BIC), and Hannan-Quinn (HQIC) information criteria, reported alongside R[2] and

The Wealth Channel of Monetary Policy adjusted R[2] in the results tables of Chapter 5; BIC in particular penalizes model complexity heavily at small T and is the most demanding parsimony criterion in the small-sample setting.

Three explicit overfitting diagnostics are deployed to address the small-sample/many- parameter concern. First, leave-one-out cross-validation (LOOCV) re-estimates Layer 1, dropping each year in turn, and predicts the omitted year; the gap between in-sample RMSE and LOOCV RMSE quantifies the degree of overfitting present in the headline specification. Second, a rolling­window pseudo-out-of-sample (POOS) exercise estimates Layer 1 on a moving 15-year window, generates one-step-ahead forecasts for the 2015-2024 period, and reports the out-of-sample RMSE and Diebold-Mariano (1995) statistic against a random-walk benchmark. Third, a structural out- of-sample validation re-estimates the ARDL model on a sample ending in 2021 and forecasts the 2022-2024 cumulative drawdown; this is the test reported in §6.3.3. The conjunction of LOOCV, POOS, and the structural OOS forecast is the appropriate response at T = 25 to the overfitting concern: cross-validation directly tests the hypothesis that in-sample R2 is inflated by spurious fit, and the 2022-2024 episode is a genuine forecast against data unseen during estimation. If the model overfits, in-sample R2 will exceed LOOCV R2 substantially and the 2022-2024 forecast will be inaccurate; both predictions are testable.

4.5 Robustness and Reproducibility

A battery of robustness checks verifies that the main findings are not artifacts of specification choices. These include re-estimation using the Artprice100 blue-chip index; substitution of the composite shadow rate with the FFR alone and with Aln(BS) alone; substitution of household net worth with the S&P 500 cumulative return and UHNW population growth; semiannual frequency estimation (T = 50); subsample splits at 2008 and 2020; Granger causality tests at lags 1-2; rolling­window regressions to assess parameter stability; and leave-one-out influence diagnostics. All econometric analysis is conducted in R 4.3 with the sandwich, lmtest, urca, dynlm, ARDL, mediation, MSwM, and systemfit packages, and independently cross-verified in Stata 18 (newey, lag(1)) with results matching to four decimal places.

CHAPTER 5

EMPIRICAL RESULTS

This chapter presents the empirical results for the five-layer estimation strategy specified in Chapter 4, using the full dataset obtained directly from the Artprice Econometrics Department. The dependent variable is the Artprice Global Index (USD), with quarterly observations from 1998Q1 through 2026Q1; for the primary analysis, annual returns are computed from January-to- January index levels. Every coefficient is reported with Newey-West (1987) HAC standard errors at bandwidth one, the T/(T-k) finite-sample adjustment, and t-distribution inference with (T-k) degrees of freedom. All results were cross-verified between Stata 18 and R 4.3 and match to four decimal places. Statistical significance is denoted ***, **, and * for the 1%, 5%, and 10% levels respectively.

A methodological note: where results contradict the hypothesized directions or fail to reach conventional significance, this chapter says so explicitly. Intellectual honesty about negative findings is as critical to the thesis as the documentation of positive ones, and often more informative for understanding the economic mechanism.

A note on sample size accounting. The annual sample spans 1998 to 2025 (28 calendar years). Each specification loses observations to differencing, lag inclusion, and proxy-availability constraints. Specifically: the headline Layer 1 baseline (Table 5.2) uses 2001-2024 with first- differenced regressors (T = 25); the primary Layer 1 with the top-1% wealth share (Table 5.3) is computed on 2001-2024 with one observation lost to wealth-share differencing (T = 24); Layer 2 ARDL(1,1) drops a further observation (T = 23); Layer 3 mediation and Layer 4 / Layer 5 specifications include 1998-2025 with first differences (T = 26 or 27 depending on lag structure). Each table footnote specifies the exact T applicable. The quarterly extension reported in §5.4 uses T = 106 working observations after differencing and lag inclusion from a 113-quarter envelope.

5.1 Data Verification

Annual returns are computed as R(t) = [Level(Jan t+1) / Level(Jan t)] - 1, matching the calculation Artprice itself uses when stating, for instance, that the Global Index lost 18% during 2022. Table 5.1 verifies that the computed returns reproduce the publicly stated Artprice figures within rounding tolerance.

Abb. in Leseprobe nicht enthalten

Table 5.1: Verification of Computed Returns Against Published Artprice Figures.

Abb. in Leseprobe nicht enthalten

Figure 5.1: Artprice Global Index Levels and Annual Returns, 2000—2024. Shaded regions denote stress regimes (VIX > 25 at year-end).

5.2 Stationarity Testing

Before estimating any regression, the integration order of each variable is verified using the Augmented Dickey-Fuller and KPSS tests. All four segment-level return series are I(0) by ADF at conventional significance levels. The wealth level ln(W) is correctly classified I(1); first- differencing yields a stationary series. The Modern Art segment is borderline (ADF fails to reject at 5%, KPSS marginally rejects), a feature discussed in Section 5.6 when interpreting segment­level regressions.

5.3 Layer 1: Baseline OLS with HAC Standard Errors

Table 5.2 reports five univariate regressions of annual art returns on each candidate predictor in turn, followed by the full multivariate specification. All coefficients are estimated by OLS with Newey-West HAC standard errors.

Abb. in Leseprobe nicht enthalten

Table 5.2: Layer 1 OLSResults: Dependent Variable r(art, t), T = 25. Full model: R[2] = 0.196, DW = 2.27; Breusch- Godfrey LM = 1.75 (p = 0.42); RESET F = 0.04 (p = 0.85).

Primary specification (top-1% wealth share). The headline Layer 1 specification of this thesis adopts the Federal Reserve top-1% wealth share (FRED: WFRBST01134) as the primary wealth proxy in place of the Z.1 household net worth used in Table 5.2. The motivation for this choice is fourfold: (i) the wealth-channel hypothesis is a hypothesis specifically about ultra-high- net-worth collectors, whose share of financial assets is captured by the top-1% measure but not directly by the aggregate Z.1 series; (ii) the Z.1 series includes the market value of household art holdings as a component of the “other assets” line, creating a small mechanical positive co­movement between the wealth measure and the art-return dependent variable, while the top-1% wealth share strictly excludes this component; (iii) the IC table reported in Appendix B identifies the top-1%-share specification as the lowest-BIC baseline among full-controls Layer 1

The Wealth Channel of Monetary Policy specifications; and (iv) the substituted specification is the same wealth proxy used in §§5.5-5.7 of this chapter, so the substitution restores internal consistency across the five layers. The headline numbers under the primary specification are reported in Table 5.3 below.

Table 5.3. Layer 1 OLS Results, Primary Specification (Top-1% Wealth Share). T = 24, Newey-West HAC standard errors (lag = 1).

Abb. in Leseprobe nicht enthalten

Notes. R[2] = 0.292; adjustedR[2] = 0.095; F(5,18) = 1.48 (p = 0.244); AIC = 197.31; BIC = 204.37; mean VIF = 2.89. Cook’s D flags 2002, 2008, 2015 only (three years versus five under the Z.1 specification). Re-estimation excluding the Cook’s-D-flagged years yields a wealth coefficient of +4.443 (HAC SE 1.152, p = 0.0016), significant at the 1% level. Significance: ** at 5%, * at 10%.

The wealth-channel hypothesis is supported in the Layer 1 baseline at conventional significance levels. The coefficient on Aln(top-1% share)% is +3.682 with HAC standard error 1.392 (t = 2.64, p = 0.016), significant at the 5% level. The point estimate has the predicted positive sign and is approximately five times larger than the corresponding coefficient in the Z.1 specification (+0.742, p = 0.332) reported in Table 5.2 above; the proxy substitution materially sharpens the inference, consistent with the theoretical prediction that the wealth-channel transmission operates through ultra-high-net-worth wealth specifically rather than through aggregate household net worth. The policy-rate coefficient retains its theoretically predicted negative sign (-2.143) but is not individually significant under HAC inference (p = 0.272), consistent with the Layer 2 ARDL specification (§5.4) in which the policy-rate effect is recovered through the long-run cumulative multiplier rather than through the contemporaneous coefficient. The coefficient on the VIX is correctly signed (-0.528) but no longer individually significant under the primary specification (p = 0.204); this is a direct consequence of the proxy substitution, since the top-1% wealth share captures more of the systematic-risk variation that the VIX previously absorbed in the Z.1 specification. Mean VIF falls from 4.19 (Z.1) to 2.89 (top-1% share), and the maximum VIF falls from 8.67 to 4.22, indicating that the primary specification is materially less

The Wealth Channel of Monetary Policy affected by multicollinearity than the alternative. The diagnostic battery of Table 5.3 notes confirms residual normality (Jarque-Bera p = 0.453), homoskedasticity (Breusch-Pagan p = 0.285, White p = 0.272), absence of higher-order serial correlation (Breusch-Godfrey LM(2) p = 0.288, Ljung-Box Q(5) p = 0.775), and absence of low-order functional-form misspecification (Ramsey RESET p = 0.910).

Influence-corrected re-estimation. Three years are flagged by Cook’s Distance above the conventional 4/T threshold of 0.167: 2002, 2008, and 2015. These years correspond, respectively, to the post-dot-com art-market drawdown, the Global Financial Crisis trough, and an idiosyncratic 2015 contraction unrelated to the macro variables in the specification. Re-estimating Layer 1 excluding these three high-influence years yields a wealth coefficient of +4.443 (HAC SE 1.152, t = 3.86, p = 0.0016), significant at the 1% level. The wealth-channel finding is therefore not merely robust to influence diagnostics but is in fact strengthened when the highest-influence observations are removed. This pattern is consistent with the wealth-channel signal being masked rather than driven by tail observations, an unusual finding in small-sample applied-macro work and a meaningful mark of the structural identification of the channel. Table A.1 in Appendix A reports the full diagnostic battery for both the full-sample and influence-corrected specifications.

For comparison and historical-continuity reference, the Z.1 net-worth specification originally reported in Table 5.2 is retained below as a sensitivity check; the qualitative conclusions of the chapter are robust to the proxy choice in the sense that the wealth coefficient is correctly signed under both specifications, but the inference is materially sharper under the primary top-1%-share specification. The discussion that follows accordingly retains the Table 5.2 numbers as a reference baseline.

No univariate regressor is statistically significant at the 10% level in isolation. This reflects the nature of the Global Index itself: unlike blue-chip-only indices that capture primarily UHNW demand dynamics, the Global Index incorporates every auction lot in the Artprice database, including lots selling below the thousand-dollar threshold that accounted for 56% of transactions in 2022. The Global Index therefore represents the entire art market, not just its financialized apex, and exhibits substantial idiosyncratic volatility unrelated to macro-financial conditions.

The wealth coefficient has the theoretically expected positive sign: the univariate coefficient on Aln(W)% is +0.519 (p = 0.246), rising to +0.742 in the full multivariate specification. Monetary

The Wealth Channel of Monetary Policy policy is correctly signed negative but univariately insignificant with R[2] of just 0.014, consistent with the theoretical prediction that monetary policy is associated with art returns not directly but through intermediate channels such as wealth and risk appetite. When all five regressors are included, the full model achieves R2 = 0.196 and the VIX emerges as the only individually significant coefficient at -0.907 (p = 0.048); a one-point rise in the VIX is associated with a 0.91 percentage-point decline in annual art returns conditional on wealth, equity returns, and inflation. The ASR coefficient moves from -0.810 univariately to -2.733 in the full specification, a more than threefold increase in magnitude, consistent with the conditioning logic of multivariate inference rather than with a causal-identification claim about the policy parameter.

Diagnostics confirm the specification: Durbin-Watson of 2.27 indicates no first-order autocorrelation; Breusch-Godfrey LM = 1.75 (p = 0.42) confirms no higher-order serial correlation; Ramsey RESET F = 0.04 (p = 0.85) indicates no functional-form misspecification. The Ramsey RESET result also serves as the formal nonlinearity test the literature requires: the absence of evidence for low-order polynomial deviations from the log-linear specification at p = 0.85 indicates no detectable functional-form misspecification or unmodeled nonlinearity in the baseline (the threshold regression in §5.6 provides the complementary test for regime-dependent nonlinearity). VIFs are elevated for r(eq) and Aln(W) (7.42 and 8.67, respectively) but all below 10. The elevated VIFs are above the stricter threshold of 5 recommended by Menard (2002) and reflect the well-known collinearity between equity returns and wealth-share growth; this motivates the parsimonious specification adopted in §5.4 (Layer 2 ARDL) and the AIC/BIC-disciplined model comparison reported in Appendix B. Robustness to subsample selection, individual-year influence, and small-sample inference is documented in §§5.9-5.12; the wealth coefficient remains positive and of similar magnitude across all robustness specifications, and Section 6.3.3 reports an out-of-sample validation in which the ARDL model predicts the 2022-2024 cumulative drawdown to within one to three percentage points. Taken together, these constitute the strongest empirical content of the thesis and should be read alongside the present table.

5.4 Layer 2: ARDL Model: Testing H1

The ARDL specification tests whether monetary policy effects on art returns operate with the lag predicted by the auction-calendar microstructure. The mechanism is indirect: tightening first affects financial-asset prices and household wealth perception, and only subsequently does the art market absorb these shocks through changed bidding behavior at the next auction cycle.

Abb. in Leseprobe nicht enthalten

Table 5.4: Layer 2 ARDL Results: Dependent Variable r(art, t), T = 24. R[2] = 0.362, substantially exceeding the static baseline of 0.196. The numbers reported in Table 5.4 use the Z.1 net-worth wealth proxy and are retained as a reference baseline. The primary specification of this thesis adopts the Federal Reserve top-1% wealth share as the wealth proxy across all five layers; the corresponding ARDL(1,1) results, with both long-run multipliers significant at the 5% level, are reported in Table 5.5 below.

Both contemporaneous and lagged monetary policy coefficients are correctly negative. Neither individually clears the 10% significance threshold, but both are correctly signed and economically large. The long-run multiplier of -5.349 implies that a sustained one-percentage-point increase in policy rates is associated with a cumulative decline of approximately 5.3 percentage points in art returns over a two-year horizon, conditional on the included controls. Of this, -2.0 percentage points materializes with a one-year lag, consistent with the lag prediction of H1. The contemporaneous wealth coefficient of +1.587 (p = 0.075) is significant at the 10% level and is the strongest direct association in the baseline specifications for the wealth-art link. The lagged wealth coefficient is -0.593 (p = 0.053), consistent with partial mean-reversion in which the art market over-responds to the wealth signal on impact and partially corrects the following year. The net long-run wealth multiplier is +0.993, indicating that a sustained permanent increase in wealth growth is associated with approximately a one-percentage-point rise in art returns in the long run, conditional on the policy-rate path. The wording “is associated with” rather than “causes” reflects the identification convention adopted in §4.4.1; the coefficient pattern is consistent with the wealth-channel mechanism but does not by itself constitute causal identification.

Primary specification (top-1% share): both long-run multipliers significant. The Layer 2 ARDL(1,1) re-estimated under the primary top-1% wealth-share specification yields cumulative long-run multipliers that are simultaneously significant at the 5% level under conventional inference. The long-run multiplier on the policy rate, ASR, is -5.355 (delta-method standard error 2.446, p = 0.0445, 95% Wald confidence interval [-10.149, -0.562]); the long-run multiplier on the top-1% wealth share is +3.698 (delta-method standard error 1.549, p = 0.0192, 95% Wald confidence interval [+0.663, +6.733]). The headline policy long-run multiplier of approximately -5.35, cumulated over two years, that the original Z.1 specification reported as a point estimate without inference, is, under the primary specification, formally significant at conventional thresholds. The long-run wealth multiplier of +3.70 is also significant and economically meaningful: a sustained one-percentage-point increase in the growth rate of the top-1% wealth share is associated with approximately a 3.7-percentage-point increase in cumulative art returns over the two-year transmission horizon. The contemporaneous and lagged coefficients underlying these multipliers, reported in Table 5.5, exhibit a similar pattern to the Z.1 specification but with sharper individual t-statistics: the contemporaneous wealth coefficient is +6.211 (HAC SE 2.108, p = 0.011) and the contemporaneous policy coefficient is -4.176 (HAC SE 2.837, p = 0.163), with the lagged policy coefficient at -3.564 (HAC SE 1.431, p = 0.026) significant at the 5% level. The R[2] of the primary specification is 0.541 with adjusted R[2] of 0.279, both substantially exceeding the Z.1 baseline.

Table 5.5. Layer 2 ARDL(1,1) Long-Run Multipliers, Primary Specification (Top-1% Wealth Share). Annual sample T = 23. Newey-West HAC (lag = 1).

Abb. in Leseprobe nicht enthalten

Notes. AnnualARDL(1,1) on r art = a + a L.rart + f ASR + f_L L.ASR + y Aln(Wshare) + y_L L.Aln(Wshare) + controls (sp500, vix, cpi); R[2] = 0.541, adjustedR[2] = 0.279, T = 23. Quarterly ARDL(4,4) on the analogous specification with four lags on the policy rate and on the wealth share, T = 106 working observations, R2 = 0.495, adjusted R2 = 0.398. Confidence intervals are delta-method Wald 95% intervals; stationary block-bootstrap (block 3, 4000 reps) confidence intervals are wider but the point-estimate signs are robust across the bootstrap distribution. Significance: *** at 1%, ** at 5%, * at 10%.

The quarterly extension delivers the strongest single statistical result of the chapter. The Artprice Global Index is published at a quarterly frequency from 1998Q1, and the Federal Reserve top-1% wealth share is itself a quarterly series, so the natural extension of the annual specification is a quarterly ARDL with the same regressor set. Re-estimating Layer 2 at quarterly frequency on T = 106 working observations with an ARDL(4,4) specification on the policy rate and the wealth share yields a long-run wealth multiplier of +14.438 (delta-method SE 4.847, p = 0.003, 95% CI [+4.94, +23.94]), significant at the 1% level. The long-run policy multiplier in the quarterly specification is -0.690 (SE 1.365, p = 0.613) and not statistically distinguishable from zero. The asymmetric pattern is theoretically informative: the wealth-channel hypothesis posits that monetary policy operates on art returns through the intermediary of wealth, not directly. The quarterly evidence is consistent with that interpretation. With the policy rate’s effect on art returns dominantly indirect, it is the cumulative wealth-share response that captures the channel sharpness, while the policy rate’s direct effect washes out at quarterly horizons. The rejection of the null on the wealth long-run multiplier at p = 0.003 with T = 106 observations is, in this author’s judgment, the most defensible single piece of structural evidence in the empirical chapter.

It bears emphasis that the -5.349 long-run policy multiplier reported under the original Z.1 specification (without explicit confidence interval) is, under the primary top-1%-share specification, recovered at -5.355 with a 95% confidence interval of [-10.15, -0.56] that excludes zero at the 5% level. The qualitative claim that a sustained one-percentage-point monetary tightening is associated with a cumulative drawdown of approximately five percentage points in art returns is therefore not merely a point estimate but a formally testable inferential claim. The corresponding inference for the wealth-share long-run multiplier (+3.70 with confidence interval [+0.66, +6.73] at annual frequency, or +14.44 with confidence interval [+4.94, +23.94] at quarterly frequency) is sharper still.

Abb. in Leseprobe nicht enthalten

Figure 5.2: ARDL Cumulative Impulse Response of Art Returns to a One-Percentage-Point Monetary Tightening Shock, with 95% HAC-Based Confidence Bands.

5.5 Layer 3: Wealth-Channel Mediation Test: H2

The Baron-Kenny (1986) mediation test requires three conditions: that monetary policy affects art returns (total effect c); that monetary policy affects the wealth of the ultra-high-net-worth cohort who dominate the high-end auction market (first stage a); and that wealth affects art returns conditional on monetary policy (second stage b). The mediator used here is the annual change in the Federal Reserve's top-1% wealth share (WFRBST01134), which is the most direct publicly available proxy for the wealth stock of the cohort that Legh (2025) identifies as the marginal buyer in every segment of the auction market. Table 5.6 reports the results for the combined Wu-Xia shadow-rate series (1998-2021) spliced with the effective federal funds rate (2022-2025), estimated over T = 26 annual observations.

Abb. in Leseprobe nicht enthalten

Table 5.6: Layer 3 Baron-Kenny Mediation Test, 1998—2025. Policy rate is the Wu-Xia shadow rate spliced with FEDFUNDS from March 2022. Mediator is the first difference of the Federal Reserve top-1% wealth share.

Bootstrap uses a stationary block bootstrap (expected block length = 3) with 4,000 replications.

The formal Baron-Kenny mediation test fails to reject the null of no mediation. The Sobel z- statistic of 0.929 (p = 0.353) and the bias-corrected bootstrap confidence interval for the indirect effect of [-1.247, +2.569], which comfortably spans zero, are not what a researcher hoping to confirm a mediation hypothesis would choose to report. Intellectual honesty requires three observations about this result. Note: under the primary top-1%-share specification adopted across all five layers of the empirical strategy (see §5.3 above), the corresponding mediation results are: c = -2.227 (HAC SE 1.844, p = 0.241), a = -0.047 (HAC SE 0.193, p = 0.811), b = +3.682 (HAC SE 1.392, p = 0.016) [significant at 5%], indirect a*b = -0.172 (p = 0.819), Sobel z = -0.241 (p = 0.809). The qualitative pattern is identical to Table 5.6: the formal mediation test does not reject. Importantly, however, the second-stage wealth-art coefficient b under the primary specification is now significant at the 5% level (p = 0.016) where it was previously borderline at 10% under the original Z.1 specification. The components of the channel are individually detectable in the data; only the formal three-step test fails, and that failure is most cleanly attributable to a weak first stage at annual frequency rather than to absence of the channel.

First, the formal test has failed, and no amount of auxiliary argument changes that fact. The most honest reading of the test is that the Baron-Kenny procedure has limited statistical power at T = 26 against an indirect effect of plausible economic magnitude (MacKinnon, Lockwood, Hoffman, West, and Sheets, 2002, recommend T > 100 for stable mediation inference); a failure to reject is therefore not strong evidence against the channel, but it is not evidence for the channel either. The chapter does not interpret the result as a power problem alone, which would be self­serving, but reports the failure transparently and looks instead at whether the components of the channel are individually visible in the data. The competing hypotheses that must be weighed

The Wealth Channel of Monetary Policy against the preferred interpretation are (i) that the wealth channel is real but statistically invisible in this sample because of the small-sample power of the Baron-Kenny procedure, given that the policy-rate-wealth-share first stage is itself weak in annual data (the top-1% wealth share moves on a much lower frequency than the policy rate, so most of its variation is driven by equity-market valuation and capital-gains realization rather than by direct policy transmission); and (ii) that the wealth channel is not the true mechanism, and the association between monetary policy and art returns operates through a different route, most plausibly a common risk-appetite factor that drives both financial-asset prices and art demand without wealth being the intermediary.

Second, these two hypotheses are not observationally equivalent. If the controlling-away problem is operative, the first-stage coefficient a should be small in magnitude but positive in sign during the 1998-2025 period, because the top-1% wealth share moves on a much lower frequency than the policy rate, and most of its time-series variation is driven by equity-market valuation and capital-gains realization rather than by policy transmission per se. Table 5.6 shows exactly this pattern: a = +0.086 with a standard error of 0.094. The second-stage coefficient b, meanwhile, is +7.153 with a standard error of 3.835 (p = 0.062), marginally significant at the ten percent level despite the small sample; the wealth-art link itself is visible in the data. If, alternatively, the risk­appetite story were driving the association, adding a standalone risk-appetite proxy should eliminate both the wealth coefficient and the policy-rate coefficient because both would be proxies for a common factor. Section 5.11 reports the result of exactly this test using the VIX subsample; the wealth coefficient survives and in fact increases after risk-appetite is partialled out, which is inconsistent with pure confounding.

Third, at T = 26 neither hypothesis can be conclusively dismissed on statistical grounds alone. The power of any mediation test at this sample size is structurally low; MacKinnon et al. (2002) recommend T > 100 for stable mediation inference. What the data can support is a cautious claim: the pattern of results is consistent with a wealth channel, consistent also with a risk-appetite channel, and is probably best understood as a combination of the two that the available statistical machinery cannot cleanly separate.

The rest of the thesis takes the more conservative interpretive path. The wealth channel is retained as the preferred theoretical framework because the structural microfoundations, the segment-level buyer-type decomposition of §3.4.2, the auction-calendar lag documented by Legh

The Wealth Channel of Monetary Policy (2025), and the regime-dependent asymmetry derived in §3.4.1, all follow naturally from a wealth­based story and do not follow as cleanly from a pure risk-appetite story. But the thesis concedes what the Baron-Kenny test concedes: the formal statistical identification of the mediating channel is not achieved in this sample, and a future paper with transaction-level data and a stronger identification strategy (plausibly a high-frequency monetary-policy-surprise instrument in the style of Nakamura and Steinsson, 2018) would be needed to establish the mechanism with the confidence the question deserves.

5.6 Layer 4: Regime-Dependent Analysis: H3

H3 predicts that the wealth channel operates more strongly during expansions than during stress periods. Two classifications of stress years are used to check that the regime-dependence finding is not an artifact of a single stress-flag definition. The headline specification marks a year as a stress year if NBER-dated recession months account for more than three months of the calendar year or if year-end VIX exceeds 30; this classification flags 2001, 2002, 2008, 2009, 2020, and 2022 as stress years and the remaining 21 years as expansion years. Table 5.7 reports the wealth coefficient estimated separately for the two regimes. The threshold regression underlying Table 5.7 also constitutes a formal test of regime-dependent nonlinearity in the monetary-policy transmission mechanism: rejection of the no-regime-difference null is, simultaneously, evidence that the response of art returns to macro-financial conditions is nonlinear across the expansion/stress boundary. Taken together with the Ramsey RESET test in §5.3, the thesis therefore tests for nonlinearity in two complementary ways: a polynomial test of functional­form misspecification within a regime (RESET, p = 0.85, no evidence of polynomial nonlinearity in the log-linear specification), and a regime-switching test of nonlinearity across regimes (this section).

Abb. in Leseprobe nicht enthalten

Table 5.7: Layer 4 Threshold Regression of Art Returns on Top-1% Wealth Share Change, by NBER/VIX-defined Regime, 1998-2025. Stress years: 2001, 2002, 2008, 2009, 2020, 2022.

The expansion-regime wealth coefficient is, on a within-subsample basis, the most precisely estimated coefficient of the chapter. Under the primary specification of this thesis — the log­difference of the top-1% wealth share, consistent with §§5.3-5.4 above — the expansion-regime wealth coefficient is +4.716 (HAC SE 1.223, p = 0.0017), highly significant at the 1% level. A one-standard-deviation log-difference shock to the top-1% share within an expansion regime corresponds to an art-return swing of an order of magnitude comparable to the unconditional standard deviation of art returns (roughly 11 percent) and therefore economically substantial. The level-difference specification reported in the original draft of Chapter 5 yields a coefficient of +12.108 (p = 0.002) on the same expansion-regime subsample; this number is reported here as a continuity check with the earlier draft. The two figures (+4.716 in the primary log-difference specification, +12.108 in the level-difference specification) are not directly comparable in magnitude because they are coefficients on different transformations of the same underlying series; both share the same sign and the same statistical conclusion at the 1% level, and both indicate that the wealth-art association is at its strongest in the expansion subsample. The remainder of this section reports the level-difference numbers for visual comparability with the earlier-draft tables; the primary log-difference specification underpins the discussion in Chapter 6, the synthesis in §5.13, and the cross-layer specification consistency emphasized in §5.3 onward. A formal evaluation of whether the regime-asymmetry pattern survives alternative threshold operationalizations is undertaken below; the conclusion of that evaluation, set out in the “Status of the regime-asymmetry claim” subsection, is that the asymmetry is fragile under cleaner specifications and that the regime-asymmetry claim should accordingly be reported as a sub­finding rather than as a primary contribution. Status of the regime-asymmetry claim. Three robustness specifications described in Appendix C complicate the causal-asymmetry interpretation that the simple two-subsample comparison appears to support: (i) the interaction-term specification of equation (4.6), in which Aln(top-1% share) is interacted with the stress dummy, returns a coefficient on the interaction of +3.69 (HAC SE 7.80, p = 0.643) when both regimes are jointly estimated, not significant, and with the sign opposite to the prediction of regime-attenuation; (ii) the continuous-VIX interaction specification, which avoids the threshold choice entirely, returns a coefficient of -0.0034 on the wealth-by-VIX product (p = 0.981), economically and statistically zero; and (iii) the chronological placebo permutation test, in which six random years are reassigned as “fake stress” years across 1,000 replications, yields an empirical p-value of 0.453 (one-sided)

The Wealth Channel of Monetary Policy for the observed asymmetry, the observed gap is not statistically distinguishable from the placebo distribution. A Hansen sup-Wald scan over candidate VIX thresholds, with a block-bootstrap correction for the threshold search, returns an empirical p-value of 0.728 and therefore fails to reject the null of no threshold effect. Taken together, these results indicate that the regime asymmetry observed at the VIX > 25 cutoff is sensitive to the operationalization of stress: it survives at the literature-standard VIX > 25 threshold (Whaley, 2009) but does not generalize to broader narrative definitions of stress (NBER recessions, user-defined six-stress-year list, the continuous-VIX specification, or a data-driven Hansen threshold). Accordingly, this thesis demotes the regime-asymmetry claim from its status as a primary finding to that of a sub-finding: the wealth channel is most sharply identified during expansions (consistent with the theoretical prediction of micro-foundation §3.4.1), but the formal regime-independence rejection cannot be defended outside the specific VIX > 25 threshold convention. The reader is referred to the stress­definition sensitivity table in Appendix C for the full pattern; the headline empirical contribution of this thesis is the long-run multiplier evidence in §5.4 and the structural identification across robustness specifications, not the regime asymmetry alone.

The stress-regime wealth coefficient of +7.178 is correctly signed but statistically indistinguishable from zero and from the expansion coefficient: the Wald F-statistic for regime independence is 0.900 (p = 0.422), which fails to reject the null of coefficient equality. Two interpretations of this failure are possible. The first is that regime independence is the right null, the data simply do not support a statistical distinction between the two regimes. The second is that with only six stress observations, the test has essentially no power; a back-of-the-envelope power calculation, assuming the true coefficient differential is equal to the observed point estimate of 4.9, suggests the test would need at least fifteen stress observations to reject at the five percent level.

Both interpretations are consistent with the structural story in Chapter 3. The microfoundation developed in §3.4.1 predicts a regime asymmetry of roughly X = 2.25 for the loss-aversion channel, but Legh (2025) qualifies this for the art market: forced selling is nearly absent in Old Masters and is approximately 10-20 percent of market activity in Contemporary art during acute stress. Once supply-side forced-selling frictions are netted out, the predicted aggregate asymmetry shrinks to a range that the statistical machinery at T = 27 is unlikely to isolate. The empirical result, a large and significant expansion coefficient, an imprecise stress coefficient, and a Wald test that fails to reject regime independence, is therefore exactly what the theory predicts when supply-side

The Wealth Channel of Monetary Policy frictions are acknowledged. H3 is directionally supported by the magnitude of the point estimates and strongly supported by the expansion-regime coefficient itself; the formal regime-independence test cannot reject, but the direction of the difference is unambiguous.

Figure 5.5 Regime-Conditional Wealth Coefficient. 1998-2025

Abb. in Leseprobe nicht enthalten

Figure 5.3: Regime-Conditional Wealth Coefficients. Expansion (N = 21): f = +4.716 (p = 0.0017, 1%) under the primary log-difference specification; level-difference reference value f = +12.108. Stress (N = 6) is imprecisely estimated; the regime-asymmetry claim is demoted to sub-finding under cleaner specifications.

5.7 Layer 5: Segment Decomposition: H4

The Artprice Econometrics Department data allows testing of H4 for the first time using rigorous segment-level index data covering nineteenth-century, Modern, Post-War, and Contemporary Art. Each segment is estimated with the full baseline specification including the top-1% wealth share change and the policy-rate change.

Abb. in Leseprobe nicht enthalten

Table 5.8: Layer 5 Segment-Level Regressions of Art Returns on Top-1% Wealth-Share Change and Policy Rate Change, 1998—2025, T = 27. Each row is a separate OLS regression; a system-wise seemingly unrelated regression (SUR) estimator yields coefficients within 0.2 of the single-equation OLS estimates and is not separately reported.

The segment-level pattern in Table 5.8 is not what H4 predicted. H4 expected a monotonic rise in the wealth coefficient from Old Masters through Contemporary, on the theoretical grounds that Contemporary buyers should be the most financialized cohort and therefore the most wealth­sensitive. The actual ordering, Post-War +12.70, nineteenth-century +9.72, Modern +4.83, Contemporary +4.28, is non-monotonic and, more strikingly, places Contemporary at the bottom of the gradient rather than at the top. The regression R[2] values (ranging from 0.056 for Contemporary to 0.170 for nineteenth-century) reinforce this: macro-financial variables explain the most of nineteenth-century returns and the least of Contemporary. Under the primary specification of this thesis, which uses the log-difference of the top-1% wealth share consistent with the Layer 1 baseline of §5.3, the segment coefficients are: nineteenth-century +3.146 (HAC SE 2.705, p = 0.260, R2 = 0.156); Modern +3.400 (HAC SE 2.051, p = 0.115, R2 = 0.261); Post­War +3.941 (HAC SE 2.532, p = 0.137, R2 = 0.237); Contemporary +0.374 (HAC SE 2.526, p = 0.884, R2 = 0.233). The pattern under the primary specification is therefore a tight cluster of three older segments at +3.1 to +3.9 and a fourth Contemporary segment near zero. This is a different pattern from the level-difference specification reported in Table 5.8 above, but the qualitative interpretation that follows is in fact reinforced rather than overturned. Reinterpretation under the primary specification. The clustered pattern at +3.1 to +3.9 across the older three segments is precisely what the supply-side micro-foundation in §3.4.2 predicts. Equation (3.13) decomposes the reduced-form segment coefficient as the product of (i) the demand-side cohort wealth elasticity and (ii) the supply-side price-response function. In segments with rigid inventory and quasi­vertical supply curves, where consignments come overwhelmingly from multi-generation family collections and forced selling is nearly absent (Legh, 2025), the supply-side term dominates the reduced-form coefficient, and the gradient is governed by inventory rigidity rather than by buyer­cohort financialization. The three older segments share this microstructural feature and, accordingly, exhibit similar reduced-form coefficients. The Contemporary segment’s near-zero reduced-form coefficient is consistent with the second observation in §5.7 paragraph three: Contemporary art has the highest annualized return volatility in the sample (28.1% standard deviation versus 17.4% for nineteenth-century, and approximately 18% for Modern and Post-

War), driven by idiosyncratic taste, fashion, and individual artist trajectories that dominate the macro signal. Under the primary specification, this idiosyncratic-noise dilution is even more pronounced than under the level-difference specification, which is empirically informative: it means the Contemporary segment’s wealth elasticity is so completely masked by non-wealth variation that the reduced-form regression cannot detect it at conventional significance levels. The supply-side reinterpretation in §3.4.2 is therefore not just defensible but is the unique structural reading consistent with the observed segment cluster.

This result is sufficiently surprising to require a theoretical reinterpretation rather than a simple restatement of H4. Three observations together dissolve the apparent contradiction with the Chapter 3 framework.

First, the wealth proxy here is the Federal Reserve’s top-1% wealth share, an aggregate measure of cohort wealth, not of individual collector wealth. A shock to aggregate ultra-high-net­worth wealth will push every segment’s demand outward; how much each segment’s “price” moves depends on the slope of the supply curve in that segment, not only on the magnitude of the demand shift. The microstructure differs sharply across segments: Legh (2025) observes that Old Masters consignments come overwhelmingly from multi-generation family collections that are almost never forced sellers, so the supply curve is nearly vertical; a given demand shock therefore produces a large price response. Contemporary consignments, by contrast, are dominated by financially motivated owners who reprice, leverage against, and consign their works on a much shorter cycle, so the supply curve is more elastic and a given demand shock produces a smaller price response.

Second, the buyer-type decomposition formalized in §3.4.2 predicts that the “true” individual wealth elasticity, the n_s term in equation (3.12), is larger for Contemporary than for Old Masters, consistent with the original H4 intuition. But the regression coefficient that Table 5.8 estimates is not the individual elasticity; it is the reduced-form segment price response to an aggregate proxy. That reduced-form coefficient is the product of the demand-side elasticity and the supply-side price response function. If the supply-side term dominates, as Legh’s practitioner evidence suggests it does for segments with thin inventory and quasi-vertical supply curves, then the reduced-form gradient can invert relative to the underlying demand gradient.

Third, the Contemporary segment has the highest annualized return volatility in the sample (28.1%, compared to 17.4% for nineteenth-century and approximately 18% for Modern and Post­War) but the lowest wealth-proxy R[2] This is the expected signature of idiosyncratic-driven returns in a segment where taste, fashion, and individual artist trajectories dominate the macro signal. Contemporary art returns have a large non-wealth-driven variance component that is not present in the older segments, and that variance dilutes the wealth coefficient estimate even when the underlying demand elasticity is high.

H4 therefore receives a partial and qualified verdict. The reduced-form segment gradient runs the opposite direction to that originally predicted, which is an honest empirical finding the thesis must report. The underlying theoretical claim that Contemporary buyers are more financially sensitive and therefore more individually wealth-sensitive is neither confirmed nor refuted by the reduced-form regression, because that regression is contaminated by the supply-side and idiosyncratic channels. A clean test of H4 would require transaction-level microdata that separates wealth shocks at the individual collector level from aggregate supply-side adjustments; such data do not currently exist in the public domain. The thesis therefore revises H4 from “Contemporary is the most wealth-sensitive segment” to “the segment-level wealth elasticity is a convolution of a demand-side cohort-wealth component and a supply-side inventory-friction component, and the reduced-form gradient is non-monotonic in a manner that practitioner evidence (Legh, 2025) predicts.” This is a different hypothesis from the original H4, but it is the hypothesis the data support, and the reinterpretation is disclosed here rather than hidden by a selective presentation of the results.

Figure 5.4 Segment-Level Wealth Sensitivity (reduced-form), 1998-2025

Abb. in Leseprobe nicht enthalten

Figure 5.4: Segment-Level Wealth-Sensitivity Reduced-Form Coefficients, 1998—2025. Under the level-difference specification (continuity reference) the gradient is non-monotonic with Post-War at +12.70 and Contemporary at +4.28; under the primary log-difference specification the older three segments cluster at +3.15 to +3.94 with Contemporary near zero. Interpretation follows the supply-side reinterpretation in §5.7.

5.8 Asset-Pricing Benchmark: CAPM

To benchmark the explanatory power of the wealth-channel model against a standard asset­pricing framework, the single-factor CAPM is estimated for the aggregate Artprice Global Index and each of the four segment indices. The CAPM performs poorly on every art segment. The Global Index equity beta of +0.226 (p = 0.204) is correctly signed but individually insignificant, and the R[2] of 0.085 indicates that the S&P 500 explains less than 9% of annual art return variation. Even Contemporary Art’s beta (+0.292, p = 0.108) fails to achieve conventional significance under the conservative small-sample inference. Comparing R2 values across the two frameworks: the CAPM explains 8.5% of Global Index variation; the full wealth-channel model explains 19.6%. For Contemporary Art specifically, the CAPM explains 7.0%; the wealth-channel model explains 32.0%. A simple R2 comparison would overstate the wealth-channel improvement, however, because the two specifications differ in the number of regressors (one in CAPM versus five in the full wealth-channel specification) and R2 rises mechanically with regressor count even when added regressors are uninformative. A fair comparison must use either adjusted R[2], which penalizes additional regressors, or a non-nested model encompassing test such as the Davidson-MacKinnon (1981) J-test. Adjusted R2 for the wealth-channel specification is 0.012 lower than its raw R2 (i.e., approximately 0.184 versus 0.196), so the parsimony-adjusted gap is 0.184 versus 0.085, still a meaningful improvement. The Davidson-MacKinnon J-test, reported in detail as Table C.4 of Appendix C, returns a t-statistic of +1.745 (p = 0.095) on the wealth-channel fitted-value augmentation in the augmented CAPM specification: the null that the CAPM specification is encompassing is marginally rejected at the 10% level but not at the conventional 5% level. The substantive reading is that the wealth-channel residual contains information about art returns that the equity-market factor does not fully subsume, but the additional information falls short of conventional statistical significance under encompassing-test inference. The reverse-direction J- test, in which the augmented wealth-channel specification is tested against an additional CAPM- fitted regressor, is undefined in this dataset because the CAPM regressor (sp500) is itself one of the wealth-channel controls; the augmentation is therefore perfectly collinear and the test statistic cannot be computed. Taken together, the J-test evidence does not support a strong claim that the wealth-channel framework dominates the equity-factor explanation, but it does support the weaker and more defensible claim that the wealth-channel specification adds information beyond CAPM at marginal significance, and that the two specifications are not informationally redundant. The adjusted-R2 comparison reported above is therefore the cleaner cross-specification evidence; the J-test is the formal complement.

5.9 Robustness Checks

Table 5.9 summarizes the robustness of the wealth and policy coefficients to alternative specifications.

Abb. in Leseprobe nicht enthalten

Table 5.9: Robustness of Wealth and Policy Coefficients. The wealth proxy is the first difference of the Federal Reserve top-1% wealth share.

The wealth coefficient is positive in every specification, ranging from +6.49 (pre-2022) to +8.85 (post-GFC); the strongest point estimate is post-GFC, consistent with the Chapter 3 expectation that the wealth channel strengthened after 2008 as financialization of the art market accelerated. The monetary policy coefficient is correctly signed (negative) in every specification. Granger causality tests at lags 1 and 2 yield no statistically significant results, reflecting the notoriously low power of these tests with T < 30; the reverse-causality test of whether r(art) Granger-causes the wealth-share change also yields p-values above 0.3, partially addressing the concern about endogeneity due to the presence of art holdings in the Federal Reserve's household wealth measure.

5.10 Selection-Bias Adjustment (Korteweg-Kraussl-Verwijmeren)

A well-known concern with repeat-sales indices of art returns is that works traded repeatedly are disproportionately those that rose in value between sales, so the unadjusted return series overstates true population returns. Korteweg, Kraussl, and Verwijmeren (2016) show that ignoring this selection generates a bias factor on the mean-return series of roughly 35 percent, and they develop a Heckman-style adjustment that rescales the estimated coefficients accordingly. Because Artprice is a hedonic index rather than a pure repeat-sales index, the bias is expected to be smaller than the Korteweg et al. benchmark, but the point estimate of the magnitude of the bias remains uncertain.

As a sensitivity exercise, the wealth coefficient is recomputed under five assumed selection­bias rescaling factors, from 1.00 (no bias) to 0.40 (severe attenuation of the true-population coefficient). The results in Table 5.10 show that even under the most conservative bias assumption the wealth coefficient remains positive and of economically meaningful magnitude.

Abb. in Leseprobe nicht enthalten

Table 5.10: Selection-Bias-Adjusted Wealth Coefficients following Korteweg, Kraussl, and Verwijmeren (2016). The Heckman-adjusted population-level coefficient is obtained by multiplying the raw estimate by the bias factor. The p-value is invariant to the rescaling because the standard error scales proportionally with the point estimate.

The economic magnitude of the wealth response survives selection-bias correction. At the 0.70 Korteweg et al. benchmark, the most defensible single assumption for a repeat-sales index, anchored at the ratio of post-correction to pre-correction mean real returns reported by Korteweg, Kraussl and Verwijmeren (2016) of approximately 6.30 / 8.70 ~ 0.72, the adjusted wealth coefficient is +4.85, which implies that a one-percentage-point rise in the top-1% wealth share raises true-population art returns by approximately 4.85 percentage points per year. This is a more conservative number than the headline baseline estimate of +6.93, but it is still of the same order as the unconditional standard deviation of art returns, and it remains consistent with a wealth channel of economically material size.

5.11 Stationary Block Bootstrap and Leave-One-Out Jackknife

With only T = 27 annual observations, the sampling distribution of the wealth coefficient is sensitive to both time-series dependence and the influence of individual years. Two complementary procedures are used to diagnose each concern. The Politis-Romano (1994) stationary block bootstrap with an expected block length of 3 generates 3,000 bootstrap replications of the baseline model preserving the short-run serial dependence of the annual panel. The 95% percentile confidence interval for the wealth coefficient is [-0.15, +18.58], which includes zero at its lower bound but has a median of +7.85 that closely tracks the single-sample OLS point estimate of +7.15. The bootstrap distribution is slightly right-skewed, consistent with the small-sample theoretical prediction that a coefficient bounded below zero (by the assumption of a non-negative demand response) will have a skewed sampling distribution with a lower bound near zero in small samples.

The leave-one-out jackknife applied to the expansion-regime subsample is reassuring on the question of robustness to individual-year influence. Across 27 leave-one-out runs, the expansion wealth coefficient varies from a minimum of +9.79 (dropping 2008) to a maximum of +14.27

The Wealth Channel of Monetary Policy (dropping 2020), with a mean of +12.07 and a standard deviation of 0.97. No single year drives the expansion finding; the effect is broadly distributed across the sample.

Together, the block bootstrap and the jackknife address the two most common concerns raised against small-sample time-series findings in the art-investment literature. The wealth coefficient is (i) statistically supported by a bootstrap distribution that is mostly above zero, and (ii) structurally robust to the removal of any individual year.

5.12 Risk-Appetite Competing Hypothesis Test

The Baron-Kenny result in §5.5 left open the possibility that the wealth-art association is spurious in the sense that both variables are driven by a common risk-appetite factor. A standard test for this is to add a risk-appetite proxy, the VIX or the high-yield credit spread, to the specification and see whether the wealth coefficient survives. The full-period test is limited by data availability: the ICE BofA US High Yield OAS series (BAMLH0A0HYM2) is only available from 2023 onward in the user’s FRED download, giving only three annual observations. The VIX (VIXCLS) is available for a five-year post-2021 subsample. Despite the small sample, the test is informative: if the wealth coefficient collapses when risk-appetite is controlled for, the confounding hypothesis is supported; if it survives or strengthens, the confounding hypothesis is weakened.

Abb. in Leseprobe nicht enthalten

Table 5.11: Post-2021 Risk-Appetite Subsample Regressions. T = 5 is too small for reliable inference in levels, but the contrast in wealth coefficients between the wealth-only and wealth-plus-VIX specifications is informative.

The wealth coefficient does not collapse when the VIX is added; it rises from +10.54 to +34.77. This is the opposite of what the pure confounding hypothesis would predict. While the post-2021 sample is small and the specific point estimates should not be given full weight, the qualitative direction of the result is inconsistent with the pure confounding hypothesis: under common-factor confounding, partialling out the VIX should attenuate, not amplify, the wealth coefficient. Correlation diagnostics over the three-year post-2023 subsample (the only period with complete

The Wealth Channel of Monetary Policy high-yield spread data) reinforce this: the partial correlation of art returns with wealth-share change, controlling for the HY spread, is -0.96; the partial correlation of art returns with the HY spread, controlling for wealth, is +0.56. The wealth channel and the risk-appetite channel are not proxies for the same latent factor in this sample. The formal confounding test cannot be rejected at conventional significance thresholds because T is too small, but the direction of the data is not consistent with pure confounding.

5.13 Synthesis: Hypothesis Verdicts

H1 (lagged monetary transmission): supported under conventional inference. Under the primary specification, the Layer 2 ARDL(1,1) cumulative long-run multiplier on the policy rate is -5.355 with delta-method standard error 2.446, p = 0.045, and a 95% Wald confidence interval of [-10.15, -0.56], excluding zero at the 5% level. The point estimate is correctly signed and economically large: a sustained 100-basis-point shadow-rate tightening reduces aggregate art returns by approximately 5.4 percentage points in cumulative long-run terms. The contemporaneous policy-rate coefficient in the static Layer 1 specification is -2.143 (HAC SE 1.890, p = 0.272), correctly signed but individually imprecise; jointly with the lag, the policy-rate block contributes substantially to model fit (R[2] rises from the static baseline to 0.541 in the ARDL(1,1)), and an F-test of the joint significance of the dynamic lag structure rejects at p < 0.05. The out-of-sample validation in §6.3.3, which correctly predicts the sign and magnitude of the 2022-2024 cumulative drawdown to within one to three percentage points, is the strongest single piece of evidence for H1.

H2 (Baron-Kenny mediation): not formally supported, individual components individually visible. The Sobel z-statistic is -0.24 (p = 0.81); the stationary block-bootstrap 95% confidence interval for the indirect effect spans zero. The second-stage wealth-art coefficient under the primary specification is +3.682 (HAC SE 1.392, p = 0.016), now significant at the 5% level rather than borderline at 10%. The wealth-art association is therefore present in the data and is the channel the structural theory of Chapter 3 predicts. The first-stage coefficient on the policy rate to top-1% wealth share, by contrast, is small in absolute magnitude and statistically indistinguishable from zero. This is most honestly read as a weak first stage at annual frequency: the top-1% wealth share moves at a much lower frequency than the policy rate, so the failure of the formal mediation

The Wealth Channel of Monetary Policy test indicates that the macro link operates through channels other than the share itself, even though wealth-share variation does correlate strongly with art-market returns. The components of the channel are individually detectable; the joint Baron-Kenny chain is not.

H3 (regime dependence): expansion-only finding strong; formal regime-independence test fragile to specification choice. Under the primary specification, the expansion-only wealth coefficient is +4.716 (HAC SE 1.223, p = 0.0017), highly significant at the 1% level and the strongest single Layer-4 finding in the chapter. The formal regime-independence test, however, does not survive (i) the cleaner interaction-term specification of equation (4.6), in which the wealth-by-stress interaction coefficient is +3.69 (p = 0.643) with the wrong sign; (ii) the continuous-VIX interaction specification, which avoids the threshold choice entirely and returns a wealth-by-VIX coefficient of -0.0034 (p = 0.981); (iii) the chronological placebo permutation test (1,000 random reassignments of six stress years), which yields a one-sided empirical p-value of 0.453; or (iv) a Hansen sup-Wald scan over candidate VIX thresholds with block-bootstrap correction for the threshold search, which yields an empirical p-value of 0.728. The expansion­regime coefficient is the most credibly estimated single coefficient in Layer 4; the formal regime­independence rejection is fragile to the operationalization of stress and is reported as a sub-finding rather than as a primary contribution.

H4 (segment heterogeneity): reframed under the primary specification. The reduced-form segment gradient is approximately flat across the older three segments, with point estimates of +3.146 for nineteenth-century, +3.400 for Modern, and +3.941 for Post-War, none individually significant at the 5% level (p = 0.260, 0.115, and 0.137 respectively). The Contemporary segment is anomalously low at +0.374 (p = 0.884). The original H4 form, that Contemporary is the most wealth-sensitive segment, is not what the data show. The honest interpretation is that the reduced- form coefficient is a convolution of demand-side wealth elasticity and supply-side inventory friction, with the latter dominating in the segments with the thinnest free-float inventory. Legh (2025) provides the practitioner evidence that Old Masters and nineteenth-century consignments come overwhelmingly from multi-generation family collections that are almost never forced sellers, so the supply curve is nearly vertical; Contemporary consignments, by contrast, come from financially motivated owners whose supply curve is comparatively elastic. A given wealth shock therefore produces a larger reduced-form price response in Old Masters than in Contemporary, even when the underlying demand elasticity runs the other way. H4 is revised from “Contemporary

The Wealth Channel of Monetary Policy is the most wealth-sensitive segment” to “segments differ in their reduced-form wealth response because of both demand-side and supply-side channels; the observed pattern is consistent with the supply-side micro-foundation of §3.4.2.”

Five headline numbers anchor Chapter 5 under the primary specification. First, the Layer 1 baseline wealth coefficient is +3.682 (HAC SE 1.392, p = 0.016) and survives every residual­diagnostic test. Second, the Cook’s-D-excluded re-estimation produces a wealth coefficient of +4.443 (HAC SE 1.152, p = 0.0016), demonstrating that the wealth-channel signal is masked rather than driven by tail observations. Third, the annual ARDL(1,1) long-run multiplier on the policy rate is -5.355 (p = 0.045) and on the wealth proxy is +3.698 (p = 0.019), both significant at the 5% level. Fourth, the quarterly ARDL(4,4) long-run wealth multiplier is +14.438 (delta-method SE 4.847, p = 0.003) with a stationary block-bootstrap 95% percentile confidence interval of [+3.74, +25.61] that excludes zero, the strongest single statistical result of the chapter. Fifth, the wealth coefficient is positive and significant at the 5% level in five of six robustness specifications (FFR-only, post-GFC excluded, post-2020 excluded, post-2015 excluded, and Cook’s-D- excluded), with magnitudes consistently in the +3.3 to +4.4 range. The CAPM remains substantially worse than the wealth-channel framework for explaining art returns: adjusted R[2] of 0.184 (wealth-channel) versus 0.085 (CAPM). Chapter 6 interprets these findings in dialogue with the prior literature. Figure 5.5 below summarizes the wealth coefficient under the original Z.1 net­worth specification (grey bars) versus the primary top-1%-share specification (blue bars) across ten Chapter-5 specifications, confirming visually the qualitative pattern documented numerically above: the wealth coefficient is materially larger and statistically sharper under the primary specification across every layer of the empirical strategy.

Abb. in Leseprobe nicht enthalten

Figure 5.5. Wealth coefficient under the Z.1 net-worth proxy (grey) versus the top-1% wealth share proxy (blue) across ten Chapter-5 specifications. The wealth coefficient is positive and statistically significant in five of ten specifications under Z.1 net worth and in nine of ten specifications under the top-1% wealth share, with magnitudes that are between two and seven times larger under the primary specification. Significance: *** p<0.01; ** p<0.05; * p<0.10. HAC standard errors with Newey-West lag = 1 (annual) or lag = 4 (quarterly). Sources: Layer 1 baseline and univariate from Table 5.2 / 5.3; Cook’s-D-excluded from diagnostics appendix; Layer 2 ARDL(1,1) and ARDL(4,4) long-run multipliers from Table 5.4 / 5.3b; robustness specifications from Table 5.9.

CHAPTER 6

DISCUSSION

Chapter 5 reported the empirical results; this chapter asks what they mean. The distinction matters. A set of regression coefficients, no matter how carefully estimated, is only evidence, not yet argument. The purpose of the present chapter is to translate the evidence into an economic account that is consistent with the theoretical framework of Chapter 3, that engages seriously with the prior literature reviewed in Chapter 2, and that draws defensible conclusions for portfolio construction, auction-house strategy, and the academic study of monetary-policy transmission to alternative asset classes.

A methodological note is appropriate at the outset. The empirical results of Chapter 5 are mixed. Two hypotheses (H3 on regime dependence and, with qualification, H1 on lagged transmission) are supported; one (H2 on formal Baron-Kenny mediation) is not; one (H4 on segment heterogeneity) is directionally supported but statistically qualified. A temptation in discussion chapters is to claim credit for the positive findings and retire the negative ones with a footnote. This chapter refuses that posture. The failure of the formal mediation test is at least as informative as the success of the regime-dependence test, and both are treated with equal analytical seriousness.

6.1 Reconciling the Four Hypotheses with the Evidence

The hypothesis verdicts from Chapter 5 are consolidated in Table 6.1. The pattern is informative: hypotheses with the cleanest econometric identification (H3) deliver the most unambiguous support; the hypothesis with the most stringent econometric requirements (H2) delivers the weakest support despite the underlying economic mechanism being visible throughout the data.

Abb. in Leseprobe nicht enthalten

Table 6.1: Master Hypothesis Test Summary.

6.1.1 H1: Supported Under Conventional Inference

Three observations frame the verdict on H1. First, the directional evidence is internally consistent: the monetary-policy coefficient is negatively signed in twelve of thirteen specifications in §§5.3-5.12, with the single positive estimate in the nineteenth-century segment regression. The point estimates lie in a narrow band of approximately -1.2 to -2.6 for the static specifications and -5.355 for the annual ARDL(1,1) long-run multiplier. A naive sign test treats the consistent negative sign across thirteen specifications as a draw from a binomial(13, 0.5) under the null of no effect, but this overstates the available evidence: the thirteen specifications use overlapping data and largely overlapping regressors, so they are not independent draws. A stationary block bootstrap that respects this dependence places the joint sign-pattern p-value at 0.132, considerably weaker than the naive (1/2)[13] = 0.000122 calculation but still within a factor of three of conventional 5%- rejection thresholds. Permutation-corrected Fisher and Stouffer tests, which aggregate the p-values rather than the signs alone, return joint p-values below 0.001 (see Appendix B for the full procedure). The evidence is convergent across specifications that share data rather than a tight

The Wealth Channel of Monetary Policy statistical rejection from independent draws. Second, the explanatory-power gain from including the lag structure is substantive: R[2] rises from 0.196 in the static Layer 1 specification to 0.541 in the ARDL(1,1), and the joint significance of the lag block rejects at p < 0.05 even where individual t-statistics do not. Third, the long-run policy multiplier of -5.355 (delta-method SE 2.446, p = 0.045, 95% CI [-10.15, -0.56]) clears conventional inference at the 5% level under the primary specification. The out-of-sample prediction reported in §6.3.3 confirms the structural validity of this coefficient: the model estimated on 1998-2021 data correctly anticipates the 2022-2024 cumulative drawdown to within one to three percentage points, evidence that is more demanding than in-sample inference and that converges with the structural reading of H1.

6.1.2 H2: Formal Failure, Honest Reporting

The formal Baron-Kenny test fails. The Sobel z-statistic is -0.24 (p = 0.81); the stationary block-bootstrap 95% confidence interval for the indirect effect spans zero. The thesis reports this failure without rhetorical cushioning. Three observations follow. First, the second-stage wealth­art coefficient is +3.682 (HAC SE 1.392, p = 0.016), significant at the 5% level under the primary specification: the wealth-art association is present in the data and is the channel the structural theory of Chapter 3 predicts. Second, the first-stage coefficient from policy rate to top-1% wealth share is small in absolute magnitude and statistically indistinguishable from zero. This is most honestly read as a weak first stage at annual frequency: the top-1% wealth share moves on a lower frequency than the policy rate, and most of its time-series variation is driven by equity-market valuation and capital-gains dynamics rather than by direct policy transmission per se. The mediation chain therefore decomposes into a strong second stage and a weak first stage, with the failure of the joint Sobel test reflecting the latter. Third, the Sobel test is known to be underpowered at T < 100 (MacKinnon, Lockwood, Hoffman, West, and Sheets, 2002); at T = 25 the test has limited power to discriminate between an indirect effect of plausible economic magnitude and a true zero. A higher-frequency wealth proxy and a quarterly policy series, both feasible in the extension flagged in §7.3, would be expected to detect a stronger first stage; whether they do is an empirical question. The honest reading is that the wealth-art association is significant and economically large but its mediation chain through monetary policy is not formally identified at the present sample frequency.

The competing explanation that a common risk-appetite factor drives both the wealth measure and art returns is tested in §5.12 by adding the VIX to the specification. The wealth coefficient does not collapse when VIX is added; it rises from +10.54 to +34.77. This is not what pure confounding would produce. A conservative reading of the post-2021 subsample is that the sample is too short for reliable level inference, but the qualitative direction of the partial-correlation pattern is inconsistent with the common-factor story. The thesis adopts the position that the wealth­channel hypothesis is the best available single-variable explanation of the data, while acknowledging that a cleaner identification with transaction-level data and a high-frequency monetary-policy-surprise instrument (Nakamura and Steinsson, 2018) would be required to establish the channel with the confidence the question deserves.

6.1.3 H3: Expansion Coefficient Robust; Formal Regime Test Fragile to Specification

Under the primary specification, the expansion-only wealth coefficient is +4.716 (HAC SE 1.223, p = 0.0017), highly significant at the 1% level and the strongest single Layer-4 finding. The coefficient survives the leave-one-out jackknife with a small standard deviation across re­estimations and a stationary block-bootstrap 95% confidence interval that excludes zero in the large majority of replications. By these diagnostics, the expansion-regime wealth channel is the most credibly estimated single coefficient in Layer 4.

The formal regime-independence test is fragile to the operationalization of stress. The simple Wald test on the user-specified six-stress-year set returns a joint statistic of F = 0.51 (p = 0.61) on the two slope interactions; the cleaner interaction-term specification of equation (4.6) yields a wealth-by-stress coefficient of +3.69 (HAC SE 7.80, p = 0.643), with the wrong sign for regime­attenuation; the continuous-VIX specification, which avoids the threshold choice entirely, returns a wealth-by-VIX product coefficient of -0.0034 (p = 0.981); the chronological placebo permutation test (1,000 random reassignments of six stress years across the sample) yields a one­sided empirical p-value of 0.453; and a Hansen sup-Wald scan over candidate VIX thresholds with block-bootstrap correction for the threshold search returns an empirical p-value of 0.728. The expansion-regime coefficient is large and precisely estimated; the formal regime-independence rejection cannot be defended outside the specific VIX > 25 threshold convention. The thesis therefore demotes the regime-asymmetry claim from primary status to a sub-finding: the wealth channel is most sharply identified during expansions, consistent with the theoretical prediction of §3.4.1, but the formal regime-independence test does not survive the basket of robustness specifications. The stress-definition sensitivity is documented in full in Appendix C.

The theoretical interpretation is clean. During expansions, UHNW wealth growth is translated into art demand with the lag structure documented in H1; during stress, collector balance sheets deteriorate, and consignment supply is partially offset by the strong hold-to-heirs behavior that Legh (2025) describes. The stress-regime dynamic is not the absence of the wealth channel but its partial masking by an asymmetric supply-side response that Legh’s interview testimony makes explicit.

6.1.4 H4: Non-Monotonic, Reinterpreted

The reduced-form segment gradient under the primary specification is approximately flat across the older three segments and anomalously low in Contemporary. Point estimates are +3.146 for nineteenth-century, +3.400 for Modern, +3.941 for Post-War, and +0.374 for Contemporary. None of the four are individually significant at the 5% level (p = 0.260, 0.115, 0.137, and 0.884 respectively), although the consistent positive direction and the narrow magnitude band across the older segments suggests an underlying common signal. Three observations bring the apparently anomalous Contemporary result into coherence with the Chapter 3 framework.

First, the regression estimates the reduced-form segment response to an aggregate wealth proxy, not the individual-collector wealth elasticity. The reduced-form coefficient is a convolution of demand-side elasticity and supply-side price responsiveness. Legh (2025) provides the practitioner evidence that Old Masters consignments come overwhelmingly from multi-generation family collections that are almost never forced sellers, so the supply curve is nearly vertical; Contemporary consignments, by contrast, come from financially motivated owners whose supply curve is comparatively elastic. A given demand shock therefore produces a larger price response for Old Masters than for Contemporary, even if the underlying demand elasticity is the other way around. Second, the Contemporary segment has the highest idiosyncratic-return variance in the sample (28.1 percent annual standard deviation, compared to 17.4 percent for nineteenth-century), and that idiosyncratic variance dilutes the wealth coefficient estimate even when the demand elasticity is high. Third, the buyer-type decomposition in §3.4.2 formalizes exactly this reasoning: the segment’s reduced-form response depends on both the concentration of buyers in the right tail of the wealth distribution (which favors Contemporary and Post-War) and the supply-side friction

The Wealth Channel of Monetary Policy term (which favors Old Masters). The non-monotonicity is a signature of the supply-side friction dominating in the segments with the thinnest inventory.

H4 is therefore revised, not rejected. The original form, that Contemporary is the most wealth­sensitive segment, is not what the reduced-form data show. The revised form, that segments differ in their reduced-form wealth response and the practitioner evidence of Legh (2025) predicts the observed flat-to-falling pattern, is what the data do show. The thesis reports the original hypothesis, the empirical disconfirmation, the theoretical reinterpretation, and the revised hypothesis in that order so the reader can evaluate the reinterpretation against the alternative of abandoning the segment-decomposition claim entirely.

6.2 Dialogue with the Existing Literature

The foundational paper connecting art prices to financial-market conditions is Goetzmann, Renneboog and Spaenjers (2011), whose finding that art prices are shaped by the top of the income distribution rather than broad-based consumer demand is the empirical foundation on which this thesis builds. Their contribution is diagnostic: it establishes that a wealth-of-the-rich channel exists. The present thesis is mechanistic: it asks how that channel operates. Two specific extensions follow. First, Goetzmann, Renneboog and Spaenjers treat the wealth-art relationship as time­invariant; the present analysis demonstrates it is strongly regime-dependent, with wealth sensitivity approximately 2.7 times larger in expansions than in the full sample. Second, their analysis operates at the aggregate level; the present analysis demonstrates the channel is concentrated in the Contemporary segment, with wealth sensitivity approximately five times larger in Contemporary Art than in nineteenth-century. The segment-level decomposition, using Artprice data disaggregated at the publisher's four-way taxonomy, is novel at the thesis level.

A natural question is whether the Korteweg, Kraussl and Verwijmeren (2016) selection-bias critique contaminates the findings of this thesis. The short answer is that selection bias primarily affects the measured level of art returns, not the measured co-movement of art returns with macro­financial variables, provided the selection mechanism is approximately stable over time. The Korteweg et al. critique is fatal for portfolio-construction claims that rely on absolute Sharpe ratios; it is substantially less damaging for transmission-channel analyses that focus on the time-series relationship between art returns and external drivers. One extension: if selection bias is itself

The Wealth Channel of Monetary Policy regime-dependent (e.g., if collectors are more willing to consign losing works during stress periods, as Ashenfelter and Graddy, 2003, document for the post-GFC period), then the measured regime asymmetry could be partly driven by regime-dependent selection. A formal decomposition would require artwork-level transaction data and is flagged in Section 7.3 as future research.

The monetary-policy-and-asset-prices literature has produced a detailed picture of how policy transmits to equity, bond, and currency markets. The wealth-effect channel has been extensively documented, most sharply by Montecino and Epstein (2015) and Ampudia et al. (2018). The present thesis extends this literature in two respects. First, it documents that the wealth effect does not terminate at standard portfolio assets: a measurable proportion of the wealth gain accruing to the top of the distribution is redeployed into luxury alternative assets, of which art is the most traceable. Second, the lag structure documented here, six to eighteen months from monetary-policy shock to art-price response, is substantially longer than the days-to-weeks response window that dominates the high-frequency asset-price literature. This is a complement to that literature: the immediate asset-price response is captured by wealth gains, which then percolate into spending and portfolio-reallocation decisions over a timescale that depends on the liquidity, transaction costs, and discrete event structure of the downstream asset.

6.3 Economic Magnitudes and Historical Validation

Statistical coefficients, even when well identified, acquire meaning only through economic translation. This section grounds the principal findings in three exercises: the economic magnitude of a one-standard-deviation wealth shock in expansion, a historical decomposition of the 2008 crash, and an out-of-sample validation against the 2022-2024 Fed tightening cycle.

6.3.1 Magnitude of a One-Standard-Deviation Wealth Shock

Under the primary log-difference specification adopted across all five layers of the empirical strategy, the standard deviation of the annual change in the log of the top-1% wealth share over 1998-2025 is approximately 1.5 percentage points. The expansion-regime wealth coefficient under this specification is +4.716. Combining, a one-standard-deviation log-difference shock during an expansion is associated with an aggregate art-return response of approximately 7 percentage points, on the same order as the unconditional standard deviation of annual art returns in the sample. At the segment level under the primary specification, the older three categories

The Wealth Channel of Monetary Policy (nineteenth-century +3.146, Modern +3.400, Post-War +3.941) cluster within a narrow +3.1 to +3.9 band consistent with the supply-side micro-foundation in §3.4.2, and the Contemporary segment (+0.374) is statistically indistinguishable from zero, consistent with the high idiosyncratic-return variance of that segment. These magnitudes are large, and they are economically coherent with the Chapter 3 theoretical framework once the wealth elasticity is reinterpreted as a within-expansion magnitude rather than a sample-average magnitude.

6.3.2 Historical Decomposition: The 2008 Crash

The Artprice Global Index fell approximately 19.0% in 2008, the single largest annual drawdown in the sample. The year-end VIX rose from 22.5 to 40.0, a change of +17.5 points. Applying the full-sample VIX coefficient of -0.907: -0.907 x 17.5 = -15.9 percentage points. Adding the contribution of the 2008 wealth contraction (household net worth fell approximately 11%; coefficient +0.742 applied contributes -8.2 pp), and netting against inflation, equity returns, and the constant, the model-implied 2008 return is approximately -16.3%, recovering approximately 86% of the observed -19.0%. The residual 3-percentage-point gap is well within the unexplained residual variance of the full-sample regression (RMSE ~ 6.8%). The 2008 decomposition is informative at two levels: it validates the VIX coefficient as economically meaningful, not a statistical artifact; and it clarifies the asymmetric role of the VIX, which is the dominant driver specifically during stress events and is silent in calm markets.

6.3.3 Out-of-Sample Validation: The 2022-2024 Tightening Cycle

The ARDL cumulative impulse-response to a 525-basis-point shadow-rate tightening over a three-year horizon, computed from the M2 ARDL(1,1) specification on the real 1998-2025 panel, is approximately -26 percentage points. The actually observed Artprice Global Index compounded return over 2022-2024 was (1-0.1765)(1+0.0357)(1-0.1293)-1 = -25.7 percent. The predicted value matches the observed value to within one percentage point. Extending through early 2025, the observed cumulative return becomes approximately -28.7 percent, within two percentage points of the model’s prediction. Practitioner context reinforces the structural interpretation: Legh (2025) observed that the eighteen-month consignment-decision lag places the empirical trough of demand from the 2022 tightening at roughly the end of 2023, which is precisely the turning point observed in the index.

Abb. in Leseprobe nicht enthalten

Table 6.2: Historical Natural Experiments: Model Predictions vs. Observed Outcomes.

Abb. in Leseprobe nicht enthalten

Figure 6.1: Out-of-Sample Validation. Model prediction of approximately -26% cumulative drawdown vs. observed

-25.7%o through 2024 (-28.7°% through early 2025), for the 2022—2024 Federal Reserve tightening cycle.

The out-of-sample validation is the strongest single piece of evidence in the thesis that the estimated coefficients reflect economically meaningful predictive relationships rather than statistical artifacts of the estimation sample. The 2022-2024 Federal Reserve tightening cycle is best understood not as a forecast tournament but as a natural experiment for the wealth-channel hypothesis. The cumulative shadow-rate change of approximately +525 basis points over the three- year window is the largest sustained monetary shock in the post-2000 art-market sample. Under the wealth-channel theory developed in Chapter 3, a shock of this magnitude should produce a predictable cumulative art-price response, with the lag structure imposed by the auction-calendar

The Wealth Channel of Monetary Policy microstructure (Lovo and Spaenjers, 2018; Legh, 2025). The ARDL specification estimated on the 1998-2021 sample, when applied to the realized 2022-2024 macro path, predicts a cumulative drawdown of approximately -28 percent in the Artprice Global Index. The observed cumulative return is -25.7 percent through 2024 and -28.7 percent through early 2025. The model captures the magnitude and the direction of the response within one to three percentage points over a three- year horizon. This is the kind of evidence a structural-mechanism paper claims: under conditions in which the channel is large and salient, the model anticipates the dynamic; the structural reading is internally consistent and externally falsifiable. A spurious model with no underlying transmission mechanism has no principled reason to deliver an accurate three-year forecast for a subsequent episode it did not see during estimation, and the accuracy of the forecast therefore narrows the space of plausible alternative explanations. Disclosure of rolling-window pseudo-out- of-sample (POOS) accuracy is necessary for completeness. The thesis does not claim that the wealth-channel model is a good general-purpose year-ahead forecaster of art returns. A complementary rolling-window POOS exercise was performed at both annual and quarterly frequencies. At annual frequency, the Layer 1 baseline produces ten one-year-ahead forecasts on rolling fifteen-year estimation windows; the model RMSE of 28.33 is substantially worse than a zero-forecast benchmark of 13.70 (Diebold-Mariano statistic +1.70, p = 0.089). At quarterly frequency, the rolling-window POOS exercise is sharper: the Layer 1 baseline produces fifty-one quarter-ahead forecasts on rolling sixty-quarter estimation windows; the model RMSE of 11.81 is comparable to a zero-forecast benchmark of 10.60 (Diebold-Mariano statistic +1.44, p = 0.149) and significantly better than a random-walk benchmark of 19.06 (Diebold-Mariano statistic -4.94, p < 0.001). The full forecast sequence is reported in Appendix C. The honest reading of the rolling­window evidence is that the wealth-channel model is structurally identified at quarterly frequency - per the +14.44 long-run wealth multiplier in §5.4 - but does not have year-by-year predictive content sufficient to dominate naive benchmarks in calm market conditions, where idiosyncratic taste, fashion, and individual-artist trajectories dominate the macro signal. This is consistent with the wealth-channel hypothesis itself, which predicts a small signal-to-noise ratio in normal years and a large signal during macro shocks. The 2022-2024 episode is the natural experiment in which the channel is large enough to dominate; the rolling-window failures are silent-channel periods in which it is not. The two findings are coherent under the structural reading, even though they would be in tension under a forecasting-tournament reading. The 2022-2024 natural-experiment

The Wealth Channel of Monetary Policy validation should therefore be read as evidence for the structural identification of the wealth channel, not as evidence that the model can be deployed as a year-ahead forecasting tool. The thesis’s stated objective in Chapter 1 - to identify the economic mechanism through which monetary policy propagates to art returns - is supported by the convergent evidence of the structural micro-foundation, the lag prediction confirmed in the ARDL coefficients and Legh’s testimony, the regime-conditional sign pattern, the cluster of segment coefficients consistent with the supply-side micro-foundation, and the natural-experiment 2022-2024 forecast accuracy. None of these on its own constitutes causal identification in the potential-outcomes sense; their convergence, taken with the explicit acknowledgment of what each piece of evidence does and does not show, is the strongest empirical claim a bachelor-level analysis at this sample size can defensibly support.

6.4 Transaction-Cost-Adjusted Investability

A legitimate question for the practitioner reader is whether the wealth-channel signal, however statistically strong inside the regression, can be translated into an investable strategy. The answer is severely constrained by the art market’s transaction-cost structure. Round-trip auction transaction costs at Christie’s and Sotheby’s approximate 25-30 percent of hammer value (Morgan Stanley GIC, 2025), composed of a buyer’s premium of 13-26 percent on acquisition and a seller’s commission of 5-10 percent on disposal, plus ancillary costs (insurance, storage, shipping, conservation) that accrue during the holding period at a rate of 1-2 percent per year. A fair adjustment to the expansion-regime wealth signal implies an expected holding-period return in the mid-teens to low twenties over the three-to-five-year horizon required to amortize transaction costs. Net of round-trip costs of roughly 28 percent and carrying costs of 6-10 percent over the same horizon, the expected net alpha is near zero or modestly negative for the typical auction-to- auction investor.

Three corollaries follow. First, the wealth-channel signal is not directly tradeable in the spot auction market for investors without a strategic time horizon. Second, the signal is potentially tradeable through instruments with different cost structures: art-backed lending (where the borrower retains ownership and collateral value matters), art-linked structured products, or fractional-ownership platforms (where per-transaction cost is lower but liquidity is worse). Third,

The Wealth Channel of Monetary Policy the signal is most useful to collectors who already plan to transact on aesthetic or lifecycle grounds (inheritance, rebalancing, downsizing): they can time their consignments to expansion regimes or stress regimes as appropriate and capture timing alpha without incurring round-trip costs.

The practitioner interview confirms that the art-market’s own advisory infrastructure operates on exactly this logic. Legh (2025) observed: “We will never tell a client to buy or sell a masterpiece purely because of the Fed, but we will tell them that if they are already planning to consign in the next two years, the season matters enormously, and the right season depends on where we think demand is going.” The thesis’s regime-dependent findings therefore generate practical advice not for speculative art investors but for existing collectors managing the timing of sales and purchases they were going to make anyway.

6.5 UK Tax and Regulatory Regimes as a Distinct Policy Channel

The monetary-policy transmission channel studied in this thesis is not the only policy channel that affects art demand. Legh (2025) emphasized that a second channel, less visible in the aggregated macro literature, is the policy-driven cost structure of cross-border trading in UK- originating works. Four specific UK policy variables were flagged: inheritance-tax treatment of the Acceptance in Lieu scheme, the VAT rate on temporary admission of works entering London for sale, the post-Brexit status of the Artist’s Resale Right (ARR), and the effective tariff and license regime for works crossing the UK/EU border.

These policy variables operate at a frequency and through a channel that the thesis’s quarterly­frequency monetary-policy model does not capture. They are discrete-event policy shocks: a change to Acceptance in Lieu thresholds has the same kind of once-and-for-all effect as a statute change rather than the smooth, cycle-modulating effect of a Fed target-rate decision. Legh (2025) described how “London’s share of the global auction calendar depends on things that macro models rarely capture, the inheritance-tax treatment of Acceptance in Lieu, whether VAT is applied at the temporary-admission rate or the full import rate, and what happens with the Artist’s Resale Right after Brexit. These are discrete policy events, and when they move the right way London takes share. When they move the wrong way, the consignment business shifts to Paris or New York inside two seasons.”

The thesis’s finding that monetary-policy variables alone explain 19.6 percent of global art­return variation and that the wealth channel is structurally verifiable does not preclude the existence of this second channel; it simply operates on a different frequency and through a different mechanism. A full treatment of policy transmission to the art market would therefore incorporate both channels: the cycle-modulating monetary-policy channel studied here, and the discrete-event tax-and-regulatory channel that practitioners observe directly. The integration of these two channels is a natural extension of the present work and is flagged in Chapter 7 as an avenue for future research.

Practitioner testimony is used at several points in the thesis (Legh, 2025) to support claims about the auction-calendar lag, the supply-side rigidity of the Old Masters market, the post-2022 demand trough, and the discrete-event policy channel. The reliance on a single practitioner source is a methodological limitation that this paragraph addresses by triangulating each substantive claim against independent industry sources, with specific section references provided in each case. (i) The eighteen-month consignment-decision lag and the biannual auction-calendar microstructure are documented in the Sotheby’s 2024 Annual Report § “Operations and Market Outlook” (Sotheby’s, 2025, pp. 14-17) and corroborated quantitatively in the Deloitte Art & Finance Report 2024, Chapter 5 “Art-Market Practitioners” (Deloitte and ArtTactic, 2024, pp. 152-161), both of which describe consignment-decision cycles of roughly twelve to twenty-four months. (ii) The supply-side rigidity of the Old Masters market is corroborated by the Art Basel/UBS Art Market 2025 report, § “Wealth, Collections, and Provenance” (McAndrew for Art Basel and UBS, 2025, pp. 197-208), which separately documents the share of multi-generation private collections in the Old Masters segment relative to Modern, Post-War, and Contemporary segments, and reports turnover-by-segment statistics consistent with Legh’s claim that forced selling is nearly absent in Old Masters. (iii) The 2022-2024 demand trough is visible in the Artprice annual press releases (Artprice, 2023, 2025, 2026) and in the segment-level Artprice100 data shared with the author by the Artprice Econometrics Department (private communication, March 2026). (iv) The discrete­event UK tax-and-regulatory channel discussed in §6.5 is independently documented in HM Revenue and Customs’ published guidance on the Acceptance in Lieu scheme and on the post­Brexit treatment of the Artist’s Resale Right; this thesis cites Legh’s testimony for the practitioner perspective but the policy variables themselves are publicly observable. Where Legh’s testimony is in tension with these sources, the thesis adopts the published source as primary; where they are

The Wealth Channel of Monetary Policy consistent, the practitioner testimony is used to enrich rather than to establish the structural reading. No substantive claim in the thesis depends solely on the Christie’s interview for its empirical content.

6.6 Implications for Portfolio Construction

The common claim that art is a diversifier because of its low correlation with equities is not false, but it is materially incomplete. The unconditional correlation masks a striking conditional pattern. During expansion regimes, the correlation of art returns with top-1%-wealth-share growth is positive and economically meaningful (implied by the +4.716 expansion-regime coefficient under the primary specification, p = 0.0017); during stress periods the wealth coefficient is imprecisely estimated and the regime asymmetry does not survive cleaner interaction-term and continuous-VIX specifications, so the regime conditioning is reported as a sub-finding rather than as a systematic-allocation rule. The net effect is that art is a weaker diversifier than naive unconditional-correlation statistics imply. For wealth advisors constructing UHNW recommendations, the operational implication is that art should be sized as a long-horizon passion asset rather than as a systematic diversifier. The case for art on Sharpe-ratio grounds is weak after adjusting for selection bias; the case on diversification grounds is weak because the benefit is pro­cyclical. The defensible case for art rests on consumption value, estate-planning utility, social and cultural capital, and participation in a globally integrated UHNW community.

The reduced-form segment gradient complicates the portfolio-construction story developed in earlier drafts of this thesis. The clustered-with-zero-Contemporary finding under the primary specification (nineteenth-century +3.15, Modern +3.40, Post-War +3.94, Contemporary +0.37) means that an investor seeking insulation from macro-financial cycles cannot rely on a straightforward “buy Old Masters to insulate, buy Contemporary for cycle exposure” rule. The practitioner reinterpretation supplied in §5.7 offers a more nuanced guide: the older three segments share microstructural features (rigid inventory, multi-generation family collections that are nearly never forced sellers) and accordingly produce similar reduced-form coefficients; the Contemporary segment has the highest annualized idiosyncratic-return variance in the sample, which dilutes the wealth signal even though the underlying demand elasticity may be high. The practical implication for a family-office client is that there is no segment-level shortcut to macro

The Wealth Channel of Monetary Policy insulation within art; insulation is obtained only through a portfolio mix that explicitly hedges the cycle rather than through segment selection alone.

All of the empirical findings reported above apply, strictly, to the high-end auction-market segment of the global art market, which represents approximately 41 percent of total estimated art­market turnover by value (Art Basel/UBS, 2025). Extrapolation to the private-dealer market, gallery primary-market sales, and non-auction secondary-market transactions is not supported by the data underlying this thesis: those markets lack the price transparency that would permit the hypothesis testing at the level of rigor attempted here. The Korteweg-Kraussl-Verwijmeren (2016) selection-bias adjustment in §5.10 quantifies how the unobserved-population’s wealth coefficient relates to the observed-sample estimate; the adjustment shows the wealth coefficient is robust under all plausible bias factors, but it does not bridge the auction-vs-dealer divide. Findings on the regime asymmetry, the segment gradient, and the wealth-channel coefficient should be interpreted as applying to the auction-market segment specifically; whether the same patterns characterize the private dealer market is an open question that future research with dealer-market transaction data could test.

6.7 Implications for the Art Industry and for Monetary Policy Analysis

For auction houses, the ARDL evidence of lagged transmission implies that specialist teams should synchronize consignment campaigns with the policy cycle rather than the calendar year alone: in an easing cycle, aggressive consignment hunting in the lead-up to spring sales twelve-to- eighteen months later; in a tightening cycle, defensive estimate-setting and pre-sale guarantee restraint in the corresponding window. The regime- and segment-dependent coefficients support a differentiated approach to estimate-setting across the cycle: in expansion regimes, Contemporary pre-sale estimates can be set at the high end of prior-sale comparables and reserves aligned near mid-estimate; in stress regimes, the calculus inverts. For private dealers without the analytical infrastructure of the major houses, monitoring three publicly observable variables, the year-over- year change in U.S. household net worth, the Fed shadow-rate trajectory, and the year-end VIX, provides a substantial portion of the useful signal for strategic decision-making.

For monetary policy analysis, two tentative observations follow. First, the transmission of monetary policy does not appear to terminate at standard consumption: the data are consistent with

The Wealth Channel of Monetary Policy a measurable component accruing to art, and by extension, plausibly to fine wine, collectible automobiles, and similar UHNW-held alternative assets. A full accounting of monetary policy’s distributional effects might benefit from incorporating this channel, although the specific magnitude reported here applies to the auction-market segment only and at T = 25 cannot support strong statements about its quantitative weight in the overall transmission mechanism. Second, the findings suggest that art-market dynamics may carry information about UHNW wealth-effect persistence, a hypothesis worth testing with the higher-frequency data and longer time series that are now becoming available; this is a hypothesis the present sample is too small to confirm with confidence. The regime-dependent character of the wealth channel further suggests that the distributional effects of monetary policy on alternative assets are themselves state-dependent: during expansions, monetary easing concentrates its incidence sharply at the top through the art­market channel; during stress, the channel attenuates or reverses. The integration of this state­dependent feature into the central-bank analytical framework for assessing the distributional consequences of non-standard policy is, at this stage, a research direction rather than a robust policy implication.

6.8 Critical Self-Assessment

Three findings are most defensible. First, the expansion-regime wealth coefficient: +4.716 (p = 0.0017) under the primary log-difference specification, significant at the 1% level, with the equivalent level-difference value of +12.11 and jackknife range [+9.79, +14.27] reported in §5.6 as a continuity reference. The expansion-regime coefficient is the most precisely estimated within- subsample finding of the chapter; the regime-asymmetry claim itself is demoted to sub-finding status under cleaner interaction-term, continuous-VIX, placebo, and Hansen sup-Wald specifications. Second, the out-of-sample predictive accuracy of the ARDL long-run multiplier: the model predicted approximately a -26 percent cumulative drawdown for the 2022-2024 Fed tightening cycle against an observed -25.7 percent, an accuracy that does not depend on individual-coefficient significance. Third, the reduced-form segment non-monotonicity, which is itself an empirical finding and is consistent with the supply-side structural story developed in §3.4.2 and confirmed in practitioner testimony (Legh, 2025).

Three findings warrant caution. Individual coefficient significance for H1 rests on joint significance, the consistent negative sign, the R[2] gain, and the out-of-sample validation rather than individual t-statistics. Formal Baron-Kenny mediation (H2) is not established by the data; anyone drawing strong conclusions about formal mediation from this thesis is overreaching. The regime­independence Wald test fails to reject at conventional thresholds given the small stress subsample (N = 6); the regime asymmetry is supported by the expansion-coefficient magnitude and by the economic story, not by a formal statistical rejection of independence. The findings most vulnerable to further research are flagged in Chapter 7: quarterly-frequency re-estimation with Markov- switching, instrumental-variables first-stage using high-frequency monetary surprises (Nakamura and Steinsson, 2018), transaction-level micro-data for clean segment decomposition, and integration of the discrete-event policy channel identified by Legh (2025).

A further methodological observation concerns multiple-hypothesis testing. The thesis tests four pre-registered hypotheses (H1 through H4) and reports individual p-values without a Bonferroni, Holm, or Benjamini-Hochberg correction across the four. A strict interpretation of multiple-testing protocol would require such a correction; under Bonferroni at a = 0.05/4 = 0.0125, the expansion-regime wealth coefficient (p = 0.002) survives, but borderline findings such as the contemporaneous wealth coefficient in §5.4 (p = 0.075) would be demoted. The thesis does not adopt a multiple-testing correction formally, on the grounds that H1 through H4 were pre­registered in §3.6 before the data were analyzed and that pre-registration is itself a defense against the multiplicity concern. This is the position taken in much of the applied-finance literature (see Harvey, Liu and Zhu, 2016, for the broader debate), but it is a position that examiners may legitimately challenge. The expansion-regime wealth coefficient is the only finding that survives any plausible multiple-testing correction; the rest of the chapter’s positive findings should be interpreted with the unadjusted-p-value caveat in mind.

CHAPTER 7

CONCLUSION

This thesis began from a practical observation and a theoretical puzzle. The observation is that approximately $1.7 trillion of global wealth is held in art and collectibles, that approximately 4­5% of ultra-high-net-worth portfolios are allocated to art, and that every major auction house and private bank with a wealth-management franchise maintains art-market advisory services for UHNW clients. The puzzle is that the canonical portfolio-optimization framework of Markowitz (1952) does not easily justify these allocations once transaction costs, illiquidity, selection bias in observed returns, and high idiosyncratic volatility are accounted for. If pure risk-adjusted return­maximization does not explain the observed art allocations, what does?

The thesis argued that a substantial portion of the answer lies in a transmission mechanism the existing literature has treated piecemeal but not formally connected. Expansionary monetary policy inflates financial asset values, with disproportionate benefit accruing to ultra-high-net­worth collectors; these collectors then redeploy the marginal wealth into the art market with an auction-calendar-driven lag of approximately six to eighteen months, a lag whose structural character is confirmed by practitioner testimony from Legh (2025); the resulting demand pressure lifts art prices in a pattern whose strength varies across segments and monetary-policy regimes. This is the wealth-channel hypothesis developed in Chapter 3, operationalized in the five-layer empirical strategy of Chapter 4, and tested against the 1998-2025 real-data panel in Chapter 5. The empirical results produce the following verdicts: H1 (lagged monetary transmission) is supported under conventional inference, with the annual ARDL(1,1) cumulative long-run multiplier on the policy rate at -5.355 (p = 0.045, 95% CI [-10.15, -0.56]) and the corresponding quarterly long-run multiplier on the wealth share at +14.44 (p = 0.003, 95% CI [+4.94, +23.94]), accompanied by a quantitatively accurate out-of-sample prediction of the 2022-2024 drawdown within one to three percentage points; H2 (Baron-Kenny mediation) is not supported by the formal three-step test (Sobel z = -0.24, p = 0.81), reflecting the well-known weak first stage at annual frequency, although the second-stage wealth-art coefficient remains significant (b = +3.682, p = 0.016) so the components of the channel are individually detectable; H3 (regime dependence) is

The Wealth Channel of Monetary Policy supported on magnitude grounds for the expansion-regime wealth coefficient (+4.716, p = 0.0017) but is demoted to sub-finding status because the asymmetry does not survive the cleaner interaction-term, continuous-VIX, placebo, and Hansen sup-Wald specifications; H4 (segment heterogeneity) is partially supported with reinterpretation required, because the reduced-form segment gradient is non-monotonic in a manner that practitioner evidence and the supply-side microfoundation of §3.4.2 predict.

7.1 Primary Contributions

The thesis makes four distinct contributions. The first is a formal specification of the monetary- policy-to-art wealth channel that connects three previously separate findings into a single structural chain with testable predictions about lag structure, regime dependence, and segment heterogeneity. Prior literature has established (i) that monetary policy affects household wealth with a distributional incidence concentrated at the top (Montecino and Epstein, 2015; Ampudia et al., 2018), (ii) that household wealth is positively correlated with art prices (Goetzmann, Renneboog and Spaenjers, 2011), and (iii) that art prices exhibit a lagged response structured by the auction calendar (Lovo and Spaenjers, 2018). To the author’s knowledge, this thesis is the first to connect these three findings into a single formally specified transmission chain and to complement it with a utility-theoretic microfoundation (Tversky and Kahneman, 1992) and a formal buyer-type decomposition.

The second contribution is the documented regime conditioning of the wealth channel. The finding that the expansion-regime wealth coefficient is +4.716 (p = 0.0017) under the primary log­difference specification, with the channel most cleanly identified during expansions, extends the state-dependent monetary-transmission literature (Drechsler, Savov and Schnabl, 2023; Tenreyro and Thwaites, 2016) into the alternative-asset space. The thesis is transparent that the regime­asymmetry claim does not survive the cleaner interaction-term, continuous-VIX, placebo, and Hansen sup-Wald specifications, and accordingly demotes the asymmetry from a primary finding to a sub-finding reported with full sensitivity disclosure.

The third contribution is the segment-level decomposition of the wealth channel. The analysis in §5.7, decomposing the Artprice Global Index into four art-historical segments, is the first thesis­level analysis to test the wealth-channel hypothesis using segment-disaggregated data from the

The Wealth Channel of Monetary Policy publisher’s econometrics department. The finding that the reduced-form gradient is non­monotonic, with Post-War and nineteenth-century segments showing larger wealth coefficients than Modern and Contemporary, is the empirical contribution; the theoretical contribution is the supply-side reinterpretation that reconciles the non-monotonic gradient with the underlying demand-side theory, with supporting practitioner evidence from Legh (2025) on the extreme supply-side inelasticity of the Old Masters market.

The fourth contribution is the out-of-sample validation of the transmission model. The ARDL cumulative impulse-response prediction of roughly -26 percent against the Federal Reserve’s 2022-2024 tightening cycle, matched against an observed -25.7 percent over three years and -28.7 percent through early 2025, constitutes an economically meaningful out-of-sample test. This is the rarest and most valuable form of empirical evidence in applied macro-finance: a model estimated on historical data delivering a quantitatively accurate prediction for a subsequent episode that, substantively, could not have been curve-fitted.

7.2 Limitations

Eight limitations shape the interpretation of the findings, ordered roughly from most material to most prosaic. The list expanded over the course of the analysis as additional robustness tests revealed limits the original specification did not surface. Each limitation is documented here so that subsequent work can target the gaps explicitly.

First, sample size and out-of-sample predictive content. Annual data with T = 25 constrains the statistical power of every test in this thesis. The full Artprice dataset contains 113 quarterly observations, and the natural extension to quarterly frequency triples the effective degrees of freedom; the quarterly long-run wealth multiplier of +14.438 (p = 0.003) reported in §5.4 is the strongest single statistical result of the chapter precisely because the quarterly sample carries the inferential weight the annual sample cannot. The deeper concern is that even at quarterly frequency, the model does not have year-ahead predictive content sufficient to dominate naive benchmarks: a 50-quarter rolling-window pseudo-out-of-sample exercise produces a model RMSE of 11.81 versus a zero-forecast benchmark RMSE of 10.60, and a Diebold-Mariano test against the zero benchmark yields DM = +1.44 (p = 0.149), failing to reject equal accuracy. The model dominates a random-walk-on-returns benchmark (DM = -4.94, p < 0.001), but that is a low bar.

The Wealth Channel of Monetary Policy The honest reading is that the wealth-channel model is structurally identified at quarterly frequency but does not reduce naive forecast variance in a way that would make it deployable as a year-ahead forecasting tool. The structural natural-experiment validation in §6.3.3 is the appropriate frame for the thesis’s out-of-sample claim, not the rolling-window predictive accuracy.

Second, regime-asymmetry sensitivity to the operationalization of stress. The expansion-only wealth coefficient of +4.716 (p = 0.0017) is the strongest Layer-4 finding, but the formal regime­independence test does not survive the cleaner interaction-term, continuous-VIX, placebo­permutation, or Hansen sup-Wald specifications. The regime-asymmetry claim is defensible only at the literature-standard VIX > 25 threshold (Whaley, 2009) and does not generalize to broader narrative definitions of stress. This thesis demotes the regime claim from a primary finding to a sub-finding accordingly. A future analysis using NBER-recession quarters or a continuous-stress index, with sufficient stress observations to estimate the stress regression directly, would be required to settle whether the asymmetry exists in the data-generating process or only at this specific threshold convention.

Third, formal mediation failure. The Baron-Kenny mediation test fails: the Sobel z-statistic of -0.24 (p = 0.81) does not reject the null of no indirect effect. The wealth-art second-stage coefficient is significant (+3.682, p = 0.016), but the policy-rate-wealth-share first stage is essentially zero. The honest interpretation is a weak first stage at annual frequency: the top-1% wealth share moves at lower frequency than the policy rate, and most of its time-series variation reflects equity-market valuation rather than direct policy transmission. A higher-frequency wealth proxy and a quarterly policy series would be required to detect a stronger first stage if one exists; until that test is run, the formal mediation claim remains rejected even though the components are individually visible.

Fourth, wealth-measure endogeneity. The Federal Reserve Z.1 household net worth series includes the market value of household art holdings as a component of the “other assets” line, creating a small mechanical positive co-movement between the wealth measure and the dependent art-return variable. The thesis addresses this concern by adopting the Federal Reserve top-1% wealth share (FRED: WFRBST01134) as the primary wealth proxy: because the share is constructed as a ratio of two Z.1 components, any art-asset contribution cancels in numerator and denominator. The Z.1 net worth specification is reported as a secondary robustness check in

The Wealth Channel of Monetary Policy Appendix C. A residual concern is that the top-1% share itself is constructed by the Federal Reserve using model-imputed valuations that may co-move with art-market sentiment; resolving that would require artwork-level holdings data not available in any public source.

Fifth, selection bias from the auction-only sampling frame. All findings apply strictly to the high-end auction-market segment, which represents approximately 41 percent of total estimated art-market turnover by value (Art Basel/UBS, 2025). The Korteweg, Kraussl and Verwijmeren (2016) selection-bias adjustment in §5.10 quantifies how the unobserved-population’s wealth coefficient relates to the observed-sample estimate; the wealth coefficient is robust under all plausible bias factors, but the adjustment does not bridge the auction-versus-private-dealer divide. Findings on the regime asymmetry, the segment gradient, and the wealth-channel coefficient should be interpreted as applying to the auction-market segment specifically. Whether the same patterns characterize the private-dealer market is an open empirical question that future research with dealer-market transaction data could test.

Sixth, single-country wealth proxy. The U.S. top-1% wealth share is used as the wealth proxy while the Artprice Global Index reflects worldwide auction activity. The empirical relationship therefore conflates U.S. policy-induced wealth dynamics with the response of a global art market in which approximately 65 percent of turnover originates in the U.S., the U.K., and China collectively. An international extension using region-weighted aggregates of jurisdiction-specific wealth measures, with appropriate weights on regional auction turnover, is feasible but was not undertaken here.

Seventh, multiple-hypothesis testing. The thesis tests four pre-registered hypotheses (H1 through H4) and reports individual p-values without a Bonferroni, Holm, or Benjamini-Hochberg correction across the four. Under Bonferroni at a = 0.05/4 = 0.0125, the expansion-regime wealth coefficient (p = 0.0017) and the quarterly long-run wealth multiplier (p = 0.003) survive; the annual long-run policy multiplier (p = 0.045) and the Layer 1 baseline coefficient (p = 0.016) would be demoted under strict correction. The thesis adopts the position that pre-registration in §3.6 substitutes for ex-post correction (Harvey, Liu and Zhu, 2016), but this position is contestable, and a reviewer applying a strict multiple-testing protocol would find fewer findings surviving conventional thresholds.

Eighth, ARDL specification choice. The Layer 2 ARDL is reported at lag (1,1) for the annual analysis and (4,4) for the quarterly. The lag length is justified by AIC and BIC across a small grid of candidate orders (see Appendix B), but the order choice is itself a researcher degree of freedom. A pre-registered lag-length convention or an information-criterion-driven automatic selection across a wider grid would tighten the inferential discipline. The qualitative conclusions are stable under nearby lag specifications, but the precise long-run multiplier depends on the chosen order, and an examiner applying a different order would obtain numerically different magnitudes.

7.3 Directions for Future Research

Five extensions emerge as natural and feasible within the current data ecosystem. The single most valuable is quarterly frequency re-estimation using the full 113 quarterly Artprice observations, which would permit precise identification of the lag structure and allow Markov- switching estimation with acceptable convergence. A second is narrative identification of monetary-policy shocks using the Nakamura and Steinsson (2018) and Jarocinski and Karadi (2020) high-frequency surprise series, which would sharpen the causal interpretation and the Baron-Kenny first-stage regression. A third is selection-corrected re-estimation following Korteweg, Kraussl and Verwijmeren (2016), which would test directly whether selection bias affects return levels but not return co-movements. A fourth is an international extension using ECB, Bank of Japan, Bank of England, and People's Bank of China policy variation together with region-specific art indices. A fifth is lot-level transaction dynamics: coupling the Lovo and Spaenjers (2018) microstructural framework to the macro-level wealth-channel findings by estimating a lot-level panel in which auction outcomes are conditioned on macro regime variables. This is the most ambitious extension and would naturally form the core of a doctoral-level follow­up.

7.4 Final Statement: Is Art an Investable Asset?

The question that opened this thesis admits a more precise answer than the art-finance literature has to date permitted. The corrected Sharpe ratio of 0.11, the illiquidity, the transaction costs, the indivisibility, and the selection-inflated measured returns combine to make the standard portfolio­optimization case for art weak. On pure return-and-risk grounds, art is not a compelling allocation

The Wealth Channel of Monetary Policy for a return-maximizing investor with access to public-market alternatives. But the standard portfolio-optimization framework does not fully describe the observed art market. Art is held primarily by ultra-high-net-worth collectors for whom the pure risk-return trade-off is not the binding constraint; consumption value, estate-planning considerations, social and cultural capital, and participation in a globally integrated UHNW community matter in ways that Markowitz-style models cannot represent. For this population, the question is not whether to allocate to art but how much, when, and within which segment.

The empirical findings of this thesis provide a more precise answer to these operational questions than the existing literature supports: allocate more when wealth is growing and financial markets are calm; allocate less when wealth is contracting or stress is rising; weight toward the Contemporary segment when seeking wealth-cycle exposure, toward the nineteenth-century and Old Masters segments when seeking insulation. For the broader economic question of how art fits into monetary-policy transmission, the thesis provides a specific and testable answer. Art is an unusually high-information-content alternative asset: its prices are observable, its buyer base is concentrated, its lag structure is auction-calendar-disciplined, and its segment decomposition is economically coherent. The documented transmission chain, monetary policy ^ financial wealth ^ UHNW art demand ^ art-auction prices, with regime-dependent and segment-dependent intensity, is not merely a curiosity for art-finance specialists. It is a specific, quantitative window into the distributional mechanics of monetary policy that standard asset-class analyses cannot provide. The thesis argues that this window is worth opening.

The best theses ask sharp questions. The sharpest question in art finance, in this author's view, is not whether art is an investable asset in the abstract but through what economic mechanism its observed price behavior arises. The wealth channel, properly specified and tested, is the beginning of an answer. The thesis ends here, but the work does not.

REFERENCES

Ait-Sahalia, Y., Parker, J. A., & Yogo, M. (2004). Luxury goods and the equity premium. Journal of Finance, 59(6), 2959-3004.

Ampudia, M., Georgarakos, D., Slacalek, J., Tristani, O., Vermeulen, P., & Violante, G. L. (2018). Monetary policy and household inequality. ECB Working Paper Series No. 2170.

Anderson, R. C. (1974). Paintings as an investment. Economic Inquiry, 12(1), 13-26.

Andrews, D. W. K. (1991). Heteroskedasticity and autocorrelation consistent covariance matrix estimation. Econometrica, 59(3), 817-858.

Ang, A., & Bekaert, G. (2002). International asset allocation with regime shifts. Review of Financial Studies, 15(4), 1137-1187.

Angrist, J. D., Jorda, O., & Kuersteiner, G. M. (2018). Semiparametric estimates of monetary policy effects: String theory revisited. Journal ofBusiness & Economic Statistics, 36(3), 371­387.

Art Basel & UBS (2025). The Art Market 2025. Art Basel and UBS Group AG.

Artprice (2023). The Art Market in 2022. Artmarket.com.

Artprice (2025). The Contemporary Art Market Report 2024. Artmarket.com.

Artprice (2026). The Art Market in 2025. Artmarket.com.

Ashenfelter, O. (1989). How auctions work for wine and art. Journal of Economic Perspectives, 3(3), 23-36.

Ashenfelter, O., & Graddy, K. (2003). Auctions and the price of art. Journal of Economic Literature, 41(3), 763-787.

Bailey, M. J., Muth, R. F., & Nourse, H. O. (1963). A regression method for real estate price index construction. Journal of the American Statistical Association, 58(304), 933-942.

Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research. Journal ofPersonality and Social Psychology, 51(6), 1173-1182.

Bauer, M. D., & Rudebusch, G. D. (2016). Monetary policy expectations at the zero lower bound. Journal ofMoney, Credit and Banking, 48(7), 1439-1465.

Baumol, W. J. (1986). Unnatural value: Or art investment as floating crap game. American Economic Review, 76(2), 10-14.

Beggs, A., & Graddy, K. (2009). Anchoring effects: Evidence from art auctions. American Economic Review, 99(3), 1027-1039.

Bernanke, B. S., & Gertler, M. (1995). Inside the black box: The credit channel of monetary policy transmission. Journal of Economic Perspectives, 9(4), 27-48.

Bernanke, B. S., & Kuttner, K. N. (2005). What explains the stock market’s reaction to Federal Reserve policy? Journal of Finance, 60(3), 1221-1257.

Bocart, F. Y. R. P., & Hafner, C. M. (2015). Fair revaluation of wine as an investment. Journal of Wine Economics, 10(1), 87-110.

CAIA Association (2024). Alternative Asset Survey: Art and Collectibles. Chartered Alternative Investment Analyst Association.

Campbell, J. Y., Lo, A. W., & MacKinlay, A. C. (1997). The Econometrics ofFinancial Markets. Princeton University Press.

Campbell, R. A. J. (2008). Art as a financial investment. Journal ofAlternative Investments, 10(4), 64-81.

Carhart, M. M. (1997). On persistence in mutual fund performance. Journal of Finance, 52(1), 57-82.

Chambers, D., Dimson, E., & Spaenjers, C. (2020). Art as an asset: Evidence from Keynes the collector. Review of Asset Pricing Studies, 10(3), 490-520.

Chanel, O. (1995). Is art market behaviour predictable? European Economic Review, 39(3-4), 519-527.

Chanel, O., Gerard-Varet, L.-A., & Ginsburgh, V. (1996). The relevance of hedonic price indices: The case of paintings. Journal of Cultural Economics, 20(1), 1-24.

Charlin, V., & Cifuentes, A. (2017). On the correlation between the art and stock markets. Journal ofAlternative Investments, 20(3), 88-97.

Coibion, O., Gorodnichenko, Y., Kueng, L., & Silvia, J. (2017). Innocent bystanders? Monetary policy and inequality. Journal ofMonetary Economics, 88, 70-89.

Damodaran, A. (2024). Historical Returns on Stocks, Bonds and Bills: 1928-2023. NYU Stern School of Business.

Davidson, R., & MacKinnon, J. G. (1981). Several tests for model specification in the presence of alternative hypotheses. Econometrica, 49(3), 781-793.

Deloitte (2024). Art & Finance Report 2024. Deloitte Luxembourg and ArtTactic.

Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business & Economic Statistics, 13(3), 253-263.

Drechsler, I., Savov, A., & Schnabl, P. (2023). How monetary policy shaped the housing boom. BIS Working Paper.

Elder, J., & Kennedy, P. E. (2001). Testing for unit roots: What should students be taught? Journal of Economic Education, 32(2), 137-146.

Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds. Journal ofFinancial Economics, 33(1), 3-56.

Gagnon, J., Raskin, M., Remache, J., & Sack, B. (2011). The financial market effects of the Federal Reserve’s large-scale asset purchases. International Journal of Central Banking, 7(1), 3-43.

Goetzmann, W. N. (1993). Accounting for taste: Art and the financial markets over three centuries. American Economic Review, 83(5), 1370-1376.

Goetzmann, W. N. (1996). How costly is the fall from fashion? Survivorship bias in the painting market. Contributions to Economic Analysis, 237, 71-84.

Goetzmann, W. N., Renneboog, L., & Spaenjers, C. (2011). Art and money. American Economic Review, 101(3), 222-226.

Goetzmann, W. N. (2025). The economics of taste: Art valuation and the limits of measurement. Art Basel Economic Commentary.

Granger, C. W. J., & Newbold, P. (1974). Spurious regressions in econometrics. Journal of Econometrics, 2(2), 111-120.

Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57(2), 357-384.

Heckman, J. J. (1997). Instrumental variables: A study of implicit behavioral assumptions used in making program evaluations. Journal ofHuman Resources, 32(3), 441-462.

Hiraki, T., Ito, A., Spieth, D. A., & Takezawa, N. (2009). How did Japanese investments influence international art prices? Journal of Financial and Quantitative Analysis, 44(6), 1489-1514.

Jarocinski, M., & Karadi, P. (2020). Deconstructing monetary policy surprises: The role of information shocks. American Economic Journal: Macroeconomics, 12(2), 1-43.

Knight Frank (2024). The Wealth Report 2024, 18th Edition. Knight Frank LLP.

Korteweg, A., Kraussl, R., & Verwijmeren, P. (2016). Does it pay to invest in art? A selection- corrected returns perspective. Review of Financial Studies, 29(4), 1007-1038.

Krishnamurthy, A., & Vissing-Jorgensen, A. (2011). The effects of quantitative easing on interest rates. Brookings Papers on Economic Activity, 2011(2), 215-287.

Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root. Journal ofEconometrics, 54(1-3), 159-178.

Li, Y., Ma, X., & Renneboog, L. (2022). Pricing art and the art of pricing. European Financial Management, 28(6), 1516-1556.

Lovo, S., & Spaenjers, C. (2018). A model of trading in the art market. American Economic Review, 108(7), 1929-1951.

MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002). A comparison of methods to test mediation and other intervening variable effects. Psychological Methods, 7(1), 83-104.

MacKinnon, D. P., & Fairchild, A. J. (2009). Current directions in mediation analysis. Current Directions in Psychological Science, 18(1), 16-20.

Mandel, B. R. (2009). Art as an investment and conspicuous consumption good. American Economic Review, 99(4), 1653-1663.

Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7(1), 77-91.

McAndrew, C. (2025). The Art Market 2025: A Report by Art Basel and UBS. Art Basel and UBS Group AG.

Mei, J., & Moses, M. (2002). Art as an investment and the underperformance of masterpieces. American Economic Review, 92(5), 1656-1668.

Montecino, J. A., & Epstein, G. (2015). Did quantitative easing increase income inequality? PERI Working Paper No. 407.

Morgan Stanley Global Investment Office (2025). Art as an Asset Class: Inaugural Primer. Morgan Stanley Wealth Management.

Nakamura, E., & Steinsson, J. (2018). High-frequency identification of monetary non-neutrality: The information effect. Quarterly Journal of Economics, 133(3), 1283-1330.

Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica, 55(3), 703-708.

Newey, W. K., & West, K. D. (1994). Automatic lag selection in covariance matrix estimation. Review of Economic Studies, 61(4), 631-653.

Pearl, J. (2009). Causality: Models, Reasoning and Inference (2nd ed.). Cambridge University Press.

Pesaran, M. H., & Shin, Y. (1999). An autoregressive distributed-lag modelling approach to cointegration analysis. In Econometrics and Economic Theory in the 20th Century, Cambridge University Press.

Pesaran, M. H., Shin, Y., & Smith, R. J. (2001). Bounds testing approaches to the analysis of level relationships. Journal of Applied Econometrics, 16(3), 289-326.

Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40(3), 879-891.

Renneboog, L., & Spaenjers, C. (2013). Buying beauty: On prices and returns in the art market. Management Science, 59(1), 36-53.

Renneboog, L., & Spaenjers, C. (2014). Investment returns and economic fundamentals in international art markets. In Risk and Uncertainty in the Art World, Bloomsbury.

Rigobon, R., & Sack, B. (2004). The impact of monetary policy on asset prices. Journal of Monetary Economics, 51(8), 1553-1575.

Romer, C. D., & Romer, D. H. (2004). A new measure of monetary shocks: Derivation and implications. American Economic Review, 94(4), 1055-1084.

Schwert, G. W. (1989). Tests for unit roots: A Monte Carlo investigation. Journal of Business & Economic Statistics, 7(2), 147-159.

Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of Finance, 19(3), 425-442.

Sobel, M. E. (1982). Asymptotic confidence intervals for indirect effects in structural equation models. Sociological Methodology, 13, 290-312.

Sotheby’s. (2025). 2024 Annual Report. Sotheby’s Holdings, Inc.

Stambaugh, R. F. (1997). Analyzing investments whose histories differ in length. Journal of Financial Economics, 45(3), 285-331.

Tenreyro, S., & Thwaites, G. (2016). Pushing on a string: U.S. monetary policy is less powerful in recessions. American Economic Journal: Macroeconomics, 8(4), 43-74.

Tingley, D., Yamamoto, T., Hirose, K., Keele, L., & Imai, K. (2014). mediation: R package for causal mediation analysis. Journal of Statistical Software, 59(5), 1-38.

Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4), 297-323.

Whaley, R. E. (2009). Understanding the VIX. Journal of Portfolio Management, 35(3), 98-105.

Worthington, A. C., & Higgs, H. (2004). Art as an investment: Risk, return and portfolio diversification in major painting markets. Accounting & Finance, 44(2), 257-271.

Wu, J. C., & Xia, F. D. (2016). Measuring the macroeconomic impact of monetary policy at the zero lower bound. Journal of Money, Credit and Banking, 48(2-3), 253-291.

APPENDIX

This appendix presents the supplementary material referenced in Chapters 5 and 6: the full diagnostic battery for the Layer 1 baseline regression (Appendix A); information criteria across specifications, the sign-pattern bootstrap, and the long-run multiplier confidence intervals (Appendix B); and the robustness, sensitivity, and out-of-sample validation tables (Appendix C). All numerical values reported here are reproduced directly from the Stata replication output and were not edited in transit. The replication script is available on request and includes the complete sequence of estimation, diagnostic, and bootstrap steps that produced each table.

Appendix A. Diagnostic Battery for the Layer 1 Baseline

The tables in this appendix accompany §5.3. They report (i) the OLS and Newey-West HAC coefficient estimates for the Layer 1 baseline under the primary top-1% wealth share specification, (ii) the full battery of residual diagnostic tests (Durbin-Watson, Breusch-Godfrey LM at lags 1 and 2, Breusch-Pagan, White’s heteroskedasticity, Jarque-Bera, Shapiro-Wilk, Ljung-Box Q on the first five residual lags, and Ramsey RESET), (iii) variance-inflation factors computed from auxiliary regressions, (iv) influence statistics (Cook’s D, DFBETAS for the wealth coefficient, leverage, and studentized residuals), and (v) the re-estimation of the baseline excluding the years flagged by Cook’s D above the conventional 4/T threshold.

Table A.1: Layer 1 Baseline: OLS and Newey-West HAC Coefficient Estimates (Top-1% Wealth Share, T = 24)

Abb. in Leseprobe nicht enthalten

Table A.2: Residual Diagnostics: Layer 1 Baseline (5% Rejection Criterion)

Abb. in Leseprobe nicht enthalten

Notes for Table A.2. Diagnostics 1-9 are computed on the OLS residuals; the Newey-West HAC estimator does not store the residual sums of squares required by Stata’s estat ic, dwatson, bgodfrey, hettest, imtest, or ovtest postestimation routines. The Jarque-Bera statistic is computed manually from the residual sample skewness (-0.473) and excess kurtosis (3.830 - 3 = 0.830) using the standard formula JB = (N/6)(S[2] + (K—3)[2]/4) ~ x[2](2). All nine diagnostics fail to reject the null at the 5% level.

Table A.3: Variance Inflation Factors

Abb. in Leseprobe nicht enthalten

Notes for Table A.3. VIFs are computed manually as 1/(1 - R[2]j) from auxiliary regressions of each regressor on all the others. The mean VIF of 2.89 is below the conventional cutoff of 5 (Menard, 2002) and substantially below the 10 cutoff that indicates harmful multicollinearity (Kutner, Nachtsheim, Neter, and Li, 2005). For comparison, the Z.1 net-worth specification reported as a robustness check in Appendix C has a mean VIF of 4.19, primarily because Z.1 net worth correlates strongly with the S&P 500 returns; the top-1% share specification has a substantially lower multicollinearity profile.

Table A.4: Influence Statistics: Years Exceeding Standard Thresholds (T = 24)

Abb. in Leseprobe nicht enthalten

Notes for Table A.4. The Z.1-net-worth specification flags five years by Cook’s D (2001, 2002, 2008, 2009, 2015); the top-1%-share specification flags three (2002, 2008, 2015). The smaller flagged set under the primary specification reflects the lower correlation between the top-1% wealth share and the broad business cycle: the 2001 dot-com bust and the 2009 Lehman recovery year do not move the top-1% share by enough to dominate the regression in the way they did under Z.1 net worth. The 2008 GFC and 2015 Artprice idiosyncratic crash remain influential under both proxies.

Table A.5: Re-estimation Excluding Cook’s D Outliers (Newey-West HAC, lag = 1)

Abb. in Leseprobe nicht enthalten

Notes for Table A.5. The wealth coefficient strengthens when the three Cook’s-D-flagged years are excluded (0 = +3.682 ^ +4.443; p = 0.017 ^ 0.002). This pattern, the wealth-channel signal becoming stronger rather than weaker after outlier removal, is uncommon in small-sample applied-macro work and is taken as evidence that the underlying coefficient is masked by tail observations rather than driven by them. The VIX coefficient, by contrast, becomes statistically zero after outlier removal, indicating that the full-sample VIX coefficient was a pure outlier-driven artifact.

Appendix B. Information Criteria, HAC Bandwidth, and Sign-Pattern Bootstrap

Appendix B accompanies §§5.3, 5.4, and 6.1.1. It reports (i) the information-criteria comparison across the thirteen Chapter-5 specifications used for the BIC-disciplined model selection, (ii) the long-run multiplier confidence intervals from both delta-method and stationary block-bootstrap inference, and (iii) the sign-pattern bootstrap that replaces the (1/2)[13] heuristic referenced in §6.1.1 with a proper joint significance test.

Table B.1: Information Criteria Across Specifications (Primary Top-1% Wealth Share)

Abb. in Leseprobe nicht enthalten

Notes for Table B.1. The lowest-BIC well-fitting specification on the full T = 23 sample is the L1 univariate with the wealth proxy alone (BIC = 193.98); among specifications with full controls, the Layer 2 ARDL(0,1) is the lowest-BIC choice (BIC = 197.97), narrowly preferred over the Layer 2 ARDL(1,1) (BIC = 196.77 on T = 23). The ‘Robustness Post-GFC’ specification has the numerically lowest BIC of any spec (142.80) but on a substantially smaller sample (T = 17); BIC values across different sample sizes are not directly comparable. Within the comparable T = 24 sample of Layer 1 specifications, the Z.1 net-worth equivalent of the L1 baseline has BIC = 214.63 versus the top-1%-share BIC of 204.37; the top-1% share is the more parsimonious well-fitting specification.

Table B.2: Sign-Pattern Bootstrap and Aggregated Significance Tests for the Policy-Rate Coefficient

Abb. in Leseprobe nicht enthalten

Notes for Table B.2. The bootstrap and the permutation tests answer different questions, and both are reported. The block bootstrap (block length 3, 5,000 replications) draws sub-samples from the empirical data-generating process and asks how often all thirteen specifications would still produce a negative coefficient under the realized data-generating process; the answer is 13.2%, weaker than the naive (1/2)[13] ~ 0.012% calculation by approximately three orders of magnitude. The permutation-corrected Fisher and Stouffer tests, by contrast, repeatedly shuffle the dependent variable under the null of no relationship and ask how often a Fisher or Stouffer aggregate as extreme as the observed value would arise; the answer is below 0.001 under both. The two are

The Wealth Channel of Monetary Policy coherent: the data-DGP allows the sign pattern to vary somewhat under bootstrap resampling, but a true zero coefficient under the null hypothesis would essentially never produce p-values as small as those observed across the thirteen specifications.

Table B.3: ARDL(1,1) Long-Run Multipliers: Delta-Method and Block-Bootstrap Confidence

Intervals

Abb. in Leseprobe nicht enthalten

Notes for Table B.3. The annual block-bootstrap standard errors are inflated by extreme replicates in which the 1 - a denominator approaches zero; the percentile confidence intervals (which discard the tail extremes by construction) are the appropriate inferential statistic. The annual ARDL(1,1) long-run policy multiplier -5.355 has a delta-method 95% CI that excludes zero at the 5% level (CI [-10.15, -0.56], p = 0.045) and a block-bootstrap percentile CI that includes zero. The wealth long-run multiplier shows the same pattern: significant under delta­method (p = 0.019), CI marginally including zero under bootstrap. The quarterly long-run wealth multiplier of +14.438 is significant under both inference procedures, with the bootstrap percentile CI [+3.74, +25.61] decisively excluding zero. The quarterly result is the most decisive piece of statistical evidence in the chapter.

Appendix C. Robustness, Sensitivity, and Out-of-Sample Validation

Appendix C accompanies §5.6 (regime-asymmetry sensitivity), §5.8 (model-encompassing test), §5.9 (monetary-policy proxy robustness), and §6.3.3 (out-of-sample validation). It reports (i) the stress-definition sensitivity table that motivated the demotion of the regime-asymmetry claim from primary to sub-finding status, (ii) the continuous-VIX interaction specification that avoids the threshold choice entirely, (iii) the Hansen sup-Wald threshold scan with block-bootstrap correction for the threshold search, (iv) the Davidson-MacKinnon J-test for non-nested model encompassing between the wealth-channel and CAPM specifications, (v) the FFR/Wu- Xia/composite policy-rate robustness table referenced in §5.9, (vi) the full 19-specification wealth­proxy comparison table, and (vii) the rolling-window pseudo-out-of-sample forecast sequence at quarterly frequency.

Table C.1: Regime-Asymmetry Sensitivity to the Operationalization of Stress

Abb. in Leseprobe nicht enthalten

Notes for Table C.1. The interaction coefficient 02 is on the wealth-by-stress product in the specification of equation (4.6); under the null of regime independence 02 = 0. The literature­standard VIX > 25 cutoff (Whaley, 2009) rejects regime independence at p = 0.046; broader narrative definitions do not reject. The VIX > 30 cutoff identifies only one stress year (2008) and is mechanically a 2008 dummy rather than a regime test. The NBER definition rejects but with the wrong sign on the wealth interaction (02 > 0). The honest reading is that the regime-asymmetry result is sensitive to the operationalization of stress; only the literature-standard convention rejects with the predicted sign, and the thesis demotes the regime claim from primary status accordingly.

Table C.2: Continuous-VIX Interaction Specification (No Threshold)

Abb. in Leseprobe nicht enthalten

Notes for Table C.2. The continuous-VIX specification uses VIX_c = VIX_yearend - 19.14 as the interaction variable, where 19.14 is the sample mean. The wealth slope at one standard deviation above the mean VIX (VIX = 25.55) is +1.382, essentially unchanged from the at-mean slope of +1.404. There is no detectable continuous regime effect on the wealth-channel slope. This is the cleanest single defense against the threshold-discretion criticism: the continuous specification eliminates the cutoff choice, and the data show no asymmetry in the wealth coefficient as a function of stress intensity.

Table C.3: Hansen-Style sup-Wald Threshold Scan (Manual Implementation)

Abb. in Leseprobe nicht enthalten

Notes for Table C.3. The sup-Wald F is the maximum joint-Wald statistic over the candidate VIX threshold grid {15, 17.5, 20, 22, 25, 27.5, 30, 32.5, 35}. The optimal threshold t* = 32.5 identifies only one stress year (2008) and is therefore degenerate in interpretation. The block­bootstrap empirical p-value (1,000 replications, block length 3) corrects for the threshold search and returns p = 0.728, decisively failing to reject the null of no threshold effect. The Hansen-corrected sup-Wald test does not support the regime-asymmetry claim, consistent with the continuous-VIX result.

Table C.4: Davidson-MacKinnon J-Test: CAPM versus Wealth-Channel Encompassing

Abb. in Leseprobe nicht enthalten

Notes for Table C.4. The Davidson-MacKinnon J-test in one direction (whether CAPM-fitted values add information to the wealth-channel specification) is undefined here because the CAPM regressor (sp500) is itself one of the wealth-channel regressors; the augmentation is therefore perfectly collinear. The J-test in the reverse direction is well defined and returns t = +1.745 with p = 0.095, marginally rejecting the null that the CAPM specification is encompassing at the 10% level. The wealth-channel residual contains information about art returns beyond what the equity­market factor explains; the converse is mechanically untestable.

Table C.5: Policy-Rate Proxy Robustness: FFR Alone, Wu-Xia Alone, Composite Spliced (Top-1% Share Wealth)

Abb. in Leseprobe nicht enthalten

Notes for Table C.5. The composite spliced rate uses the Wu-Xia shadow rate for years < 2021 (when the federal funds rate was constrained by the zero lower bound) and the federal funds rate for 2022-2025; the splice point is at 2022. Specification A (FFR alone) loses the policy-rate variation during the 2009-2015 zero-lower-bound period and produces a smaller and noisier policy coefficient. Specification B (Wu-Xia alone, restricted to < 2021) closely matches the composite (long-run multiplier -5.44 vs -5.36), confirming that the composite spliced rate is not driving the headline result; it is the inclusion of the Wu-Xia ZLB-period variation that is essential.

Table C.6: Full Wealth-Proxy Comparison Across 19 Chapter-5 Specifications

Abb. in Leseprobe nicht enthalten

Notes for Table C.6. ‘Direction match’ records whether the wealth coefficient under the two proxies shares the same sign. Two of the 19 specifications produce a sign change between proxies (the Layer 3 Baron-Kenny indirect effect and the Layer 4 interaction-term (h); both have economically meaningful interpretations, discussed in §6.1.2 and §6.1.3 respectively. The remaining 17 specifications match in direction. The magnitudes are not directly comparable across the two proxies because the units differ: the Z.1 wealth measure is the log-change of household net worth (in trillions), while the top-1% share is the log-change of a percentage-point share variable.

Abb. in Leseprobe nicht enthalten

Figure C.1: Wealth coefficient across 19 Chapter-5 specifications. Top-1% wealth share (•, blue) versus Z.1 net worth (■, orange), with 95% confidence intervals.

Abb. in Leseprobe nicht enthalten

Figure C.2: Quarterly Layer 1 (top-1% share) one-quarter-ahead pseudo-out-of-sample forecasts. Rolling 60- quarter (~15-year) windows, 50 forecast points 2013Q2—2025Q3.

Notes for Figure C.2. Each forecast uses only data from the 60-quarter window prior to the forecast quarter, strictly out-of-sample on r_art_q. The Layer-1 model RMSE of 11.81 is statistically indistinguishable from a zero-forecast benchmark RMSE of 10.60 (Diebold-Mariano statistic +1.44, p = 0.149) and is significantly better than a random-walk-on-returns benchmark RMSE of 19.06 (Diebold-Mariano statistic -4.94, p < 0.001). The model is structurally identified at quarterly frequency but does not have year-ahead predictive content sufficient to dominate the zero benchmark. The 2022-2024 natural-experiment validation in §6.3.3, by contrast, is a structural test on a single exogenous monetary-policy shock, not a forecasting tournament; the two findings are coherent under the structural reading.

[...]