The Aggregated Equity Risk Premium

Abstract

We propose a new approach for predicting the equity risk premium (ERP) that first estimates expected returns on individual stock before aggregating them to the market level. Our deep learning combination forecast aggregates firm-level return predictions from neural networks of varying complexity, trained on a comprehensive two-dimensional feature set of post-publication firm-level characteristics and aggregate macroeconomic variables. Using this aggregation method, we achieve an out-of-sample R² of 2.74% in a sample from 2000 to 2021. The forecasts demonstrate strong economic significance in trading strategies even with transaction costs. While the market generated a return of 376% over this period, a simple market-timing strategy based on our model's forecast signs yields a net cumulative return of approximately 768%. Our results show that aggregating firm-level predictions can lead to profitable market timing signals, challenging the conventional wisdom that the ERP is unpredictable out-of-sample and suggesting that valuable market-wide information can be extracted from the cross-section of individual stocks.

Keywords: Equity risk premium, Stock market anomalies, Machine learning models, Return prediction JEL classifications: G12, G14, G17, C45, C58

JEL Classification: G12, G14, G17, C45, C58