Abstract: We study the critical issue of handling missing entries in hedge fund data. We introduce a deep learning approach, the BRITS, for recovering data for fund returns and 23 fund predictors. We compare its performance with popular imputation methods, such as the cross-sectional mean and singular value thresholding. BRITS' ability to capture information from past and future values in time series and the whole cross-section of observations yields the highest imputation fidelity in our simulations. The recovered information improves predictions of nonlinear and linear methods. At the same time, it helps to select top-performing funds that earn significant out-of-sample annual alphas of 13.4% net of all costs. 

Presentations: Paris December Finance Meeting 2025 (Paris, scheduled), FMA 2025 (Vancouver, scheduled), SFA 2025 (Orlando, scheduled), Alpine Finance Summit 2025 (Grenoble), 32 Finance Forum (Spanish Finance Association, AEFIN) (Pamplona), University of Edinburgh 2025.

Abstract: We propose a novel asset pricing factor model employing a new estimation method based on sparse second-order stochastic dominance (SSD), implemented via a greedy algorithm combined with linear programming. Initially drawing from a prominent set of 24 candidate factors featured extensively in the asset pricing literature, we find statistically and economically that no incremental benefit emerges from including more than ten factors. To further ensure robustness and validate our results, we extend our factor universe to a comprehensive set of 177 factors—153 factors from Jensen et al. (2023) plus the original 24. Remarkably, our empirical results again demonstrate that a ten-factor SSD-based sparse model remains optimal and consistently outperforms all leading benchmarks. This SSD-based sparse factor model significantly generalizes the mean-variance Arbitrage Pricing Theory (APT) framework by explicitly accommodating broader classes of investor preferences, thus providing enhanced theoretical rigor and clear economic interpretation. Comprehensive parametric and non-parametric tests, conducted in-sample and out-of-sample on cross-sectional and time-series equity returns, reveal that our proposed model consistently dominates prominent benchmark models, including LASSO-based machine learning approaches. These robust findings emphasize the considerable advantages of explicitly integrating investor welfare considerations into factor selection, thereby delivering rigorous theoretical insights and practical asset pricing relevance.

Abstract:  This paper introduces a new procedure to control for family error rate (FWER) in picking outperformers. The method utilizes multiple side information to estimate the FWER and gains much higher power in detecting out-performers compared to existing ones. In assessing hedge fund performance context, the new method allows investors picking out-performing funds with high confidence, that is, with low FWER. The yearly rebalancing portfolios of hedge funds constructed by the new method using of available covariates beat passive benchmarks in various settings. Our further experiments show that the new method detects truly out-performing hedge fund managers who can repeat their past performance over a long horizon.

Presentations: 31 Finance Forum (Spanish Finance Association, AEFIN) (Tenerife), EFMA 2024 (Lisbon)