Working Papers

Risk-On, Risk-Off: A Decision-Focused Learning Framework for Market Regimes 

Market regime models are typically estimated to fit return distributions, yet investors require regimes optimized for decisions. We introduce a utility-optimized unsupervised learning framework in which regime classifications are endogenously determined by their contribution to investor welfare rather than statistical fit. The method jointly learns interpretable regime classifications and regime-specific optimal portfolios in a unified optimization procedure. In an application to U.S. equities, gold, and treasuries, the framework endogenously recovers a robust Risk-On/Risk-Off structure, implying 100% allocations to equities in Risk-On states and to gold in Risk-Off states. Utility-optimized regime learning produces economically large welfare gains because it clearly identifies this safe-haven switching structure. The framework also provides an economically interpretable ranking of the value of uncertainty and risk-aversion signals for regime identification. Relative to standard regime classification benchmarks, the approach delivers economically significant certainty equivalent gains and improved out-of-sample market timing performance.

 

Updating Density Estimates using Conditional Information Projection: Stock Index Returns and Stochastic Dominance Relations (with Stelios Arvanitis and Thierry Post) 

We propose, analyze, and apply a Conditional Information Projection Density Estimator (CIPDE) that estimates latent conditional density functions by projecting a prior time-series estimator onto distributions that satisfy a set of conditional moment conditions with functional nuisance parameters. The derivation of limit theory coupled with information-theoretic results characterizes the estimator and its improvements over the prior estimator. Theoretically, CIPDE is shown to achieve a lower limiting relative entropy to the latent distribution, provided that the prior is inconsistent and the moment conditions are well specified. An application to stock index options is presented using conditional moment restrictions based on market prices and pricing restrictions for index options. CIPDE is shown to enhance index return density forecasts out-of-sample and improve the out-of-sample investment performance of index option strategies by better timing protective put purchases and covered call writing.

 

Empirical Likelihood Tests for Pairwise Stochastic Dominance (with Stelios Arvanitis and Thierry Post) 

We propose and study an empirical likelihood–based test for stochastic dominance between two random variables in time-series settings. The test statistic is pointwise and self-normalizing, allowing it to exploit local violations and achieve higher intrinsic power than global measures based on integrated distribution functions. Critical values are consistently estimated by first localizing candidate binding constraints as minima of a functional linking the dominance inequalities, and then computing the empirical likelihood statistic on these candidates using recentered subsamples. We establish first-order asymptotic validity and show that the test converges to self-normalized quadratic limits with Gaussian-cone representations, revealing that its local power systematically exceeds that of procedures aggregating violations across the support. Monte Carlo experiments demonstrate that the proposed test maintains strong size control across various specifications and sample sizes, exhibits higher size-adjusted power compared to competing distance-based tests, and shows robustness to different choices of the localization threshold. 

Option-implied Physical Probabilities (with Thierry Post & Valerio Poti)

Results of Empirical Likelihood Ratio tests support the joint specification of the density estimates and the pricing kernel for stock index options by Constantinides, G. M., J. C. Jackwerth, and S. Perrakis, 2009, Mispricing of S&P 500 Index Options, Review of Financial Studies 22, 12471277. The test results include implied probabilities which are shown to be superior density forecasts for stock index returns compared with the original density estimator and the estimated risk-neutral density, using both statistical and economic goodness criteria. Further improvements of predictive ability are obtained by refining the initial density estimates and the pricing kernel system.

Photo taken at Alcázar, Seville, Spain.