Risk Factor Centrality and the Cross-Section of Expected Returns

The Factor Zoo phenomenon calls for answers as to which risk factors are in fact capable of providing independent information on the cross-section of expected excess returns, while considering that asset-pricing literature has produced hundreds of candidates. In this paper, we propose a new methodology to reduce risk factor predictor dimensions by selecting the key component (most central element) of their precision matrix. Our approach yields a significant shrinkage in the original set of risk factors, enables investigations on different regions of the risk factor covariance matrix, and requires only a swift algorithm for implementation. Our findings lead to sparse models that pose higher average in samples R^2 and lower root mean square out of sample error than those attained with classic models, in addition to specific alternative methods documented by Factor Zoo-related research papers. We base our methodology on the CRSP monthly stock return dataset in the time frame ranging from January 1981 to December 2016, in addition to the 51 risk factors suggested by Kozak, Nagel, and Santosh (2020)