Zookeeping: A Comparison of Factor Pruning Methods

Sammanfattning

The proliferation of proposed risk factors in empirical asset pricing, often referred to as the factor zoo, complicates the identification of a true stochastic discount factor (SDF). To address this, we assess in- and out-of-sample cross-sectional performance for Fama-French models, Principal Component Analysis (PCA), Risk-Premium PCA (RP-PCA), Bayesian Model Averaging (BMA), and a nonlinear Deep Latent Factor Model (DLFM) across varying regularization regimes. We find that the performance of dense and latent factor models depends strongly on the choice of regularization method and strength, and that the optimal choice of least-squares method itself depends on how factors are constructed, while a regularization-agnostic BMA model provides robust pricing of the cross-section. Finally, we find a positive relationship between model dimensionality and cross-sectional pricing ability; however, performance improvements diminish rapidly for latent factor models, suggesting that the underlying SDF is likely to be dense in observable factors, yet adequately represented by a relatively small set of latent factors.