Searching for the Best Conditional Equity Premium Model

Abstract

Three most prominent theoretical state variables, inflation, scaled market prices, and market variance, are identified as best market return predictor variables via best subset variable selection. They work better together than individually, consistent with multifactor equity premia stressed by recent asset pricing models. Bootstrap simulation shows that best subset has high accuracy rates. The variables are consistently chosen by alternative variable selection techniques. The multifactor model is strikingly stable across subsamples and has an out-of-sample R2 of 9.6% over the 1965Q1 to 2020Q4 period, outperforming machine learning models and the combination forecast method. We find similar results using international data.

Keywords: Market equity risk premium, Machine Learning, Random Forecasts, Boosted Regression Trees, Neural Networks, Stock market variance, Scaled market prices, Inflation, Multifactor model, Consumption-based asset pricing model, The Great Depression

JEL Classification: G10, G12, G17