Empirical Asset Pricing via Neural Networks and Macroeconomic Data
Abstract
This study showcases the benefits of expanding the dimensions of the variable input vector
with macroeconomic predictors when predicting monthly out-of-sample stock-level risk
premiums. Using 610 predictor variables, we achieve a prediction performance of 3.19%
R2
oos for our best model and for stocks with a large market value of equity, over a fourfold
increase in performance compared to previous research on this area. Furthermore, by using
Shapley values, we show the pricing importance of each group of variables, challenging
the view of neural networks as black boxes. The resulting Shapley values indicate that the
neural networks weight higher variables such as labor market, interest and exchange rates,
and prices, during recession, and stock characteristics during expansion periods.