Improving Estimation of Factor Risk Premia via Robust Covariance Techniques

Abstract

Accurate estimation of factor risk premia is a key to success in factor-based investing and risk management. Current risk premia estimation methodologies are subject to instability and other issues caused by financial data's tendency to follow heavy-tail distributions and contain outliers. We propose a toolkit of three robust covariance estimation techniques, namely winsorization, adaptive Huber covariance estimation, and U-type covariance estimation. They are incorporated in a three-pass regression methodology that constructs a latent factor model using principal component analysis of portfolio returns. We then investigate their effects on common factors well explored in asset-pricing literature and a large set of macroeconomic factors available from the Federal Reserve Economic Data database. The results show that the three covariance estimators often have similar effects on the risk premia estimates. Since the true risk premia of factors are unknown, we perform a simulation study which demonstrates the effectiveness of the covariance techniques in reducing the estimation error. The overall results suggest that robust covariance estimators are effective tools for superior risk premia estimation.