CMStatistics 2021: Start Registration
View Submission - CFE
Title: Interpretable deep learning in asset pricing Authors:  Andreas Neuhierl - Washington University in St. Louis (United States) [presenting]
Jianqing Fan - Princeton University (United States)
Tracy Ke - Harvard University (United States)
Yuan Liao - Rutgers University (United States)
Abstract: A new nonparametric methodology is developed for estimating conditional asset pricing models using deep neural networks. We employ time-varying conditional information on betas carried by firm-specific characteristics. The method first applies cross-sectional deep learning, period-by-period to estimate spontaneous conditional expected returns, defined as the conditional expectation of asset returns given characteristics and factor realizations. We also estimate the long-term expected return as the predicted mispricing component and the product of the estimated risk exposures times the price of risk. We apply local kernel smoothing to capture the return dynamics that arise from time-varying alphas and betas. We formally establish the asymptotic theory of the deep-learning estimators for conditional expected returns, alphas and risk premia, which apply to both in-sample fit and out-of-sample predictions. Contrary to many applications of neural networks in economics, we can open the black box and provide an economic interpretation of the successful predictions obtained from neural networks. We decompose predicted returns into a risk-based and mispricing component. Empirically, we find a large, time-varying mispricing component. We find that the mispricing component is slowly decaying over time, but not monotonically. Mispricing tends to be high during times of high market volatility which is linked to periods of economic turmoil.