A regularized instrumented principal components analysis

This paper examines instrumented principal components analysis (IPCA), one of the newly developed machine learning methods to model asset returns. It proposes the implementation of a support vector regression (SVR) into IPCA to better capture the relationship between firm characteristics and returns. Using a set of 9689 equities and 46 characteristics we show that IPCA(SVR) has better out-of-sample performance compared to the classic IPCA and multi-factor models in the majority of cases. The results imply that the least squares approach of IPCA overfits the characteristic sensitivity matrix, resulting in better in-sample results but worse out-of-sample fit than regularized IPCA. Although the regularized specifications exhibit superior predictive performance, the null hypothesis of no pricing errors is rejected in several testing environments, highlighting the distinction between predictive accuracy and asset-pricing validity.