The equity premium puzzle: an artificial neural network approach

Purpose - In recent years, equity premiums have been unusually large and efforts to forecast them have been largely unsuccessful. This paper presents evidence suggesting that artificial neural networks (ANNs) outperform traditional statistical methods and can forecast equity premiums reasonably well. Design/methodology/approach - This study replicates out‐of‐sample estimates of regression using ANN with economic fundamentals as inputs. The theory states that recent large equity premium values cannot be explained (the equity premium puzzle). Findings - The dividend yield variable was found to produce the best out‐of‐sample forecasts for equity premium. Research limitations/implications - Although the equity premium puzzle can be partly explained by fundamentals, they do not imply immediate policy prescriptions since all forecasting techniques including ANN are susceptible to joint assumptions of the techniques and the models used. Practical implications - This result is useful in capital asset pricing model and in asset allocation decisions. Originality/value - Unlike the findings from previous research that are unable to explain equity premium behavior, this paper suggests that equity premium can be reasonably forecasted.