Fast Factor Extraction for Mixed Type Financial Data

Abstract

Empirical research has access to ever larger datasets of mixed data types. The sheer amount of available data may often lead to the use of techniques such as dimensionality reduction or shrinkage. To reduce dimensionality, it is not uncommon to assume that the data are driven by a small number of common factors. For metric, or continuous, data, factor extraction may be conducted by means of standard principal component analysis [PCA]. PCA is, however, not directly applicable to count or binary data. Mixed data types may be dealt with via generalized linear models driven analogously by latent factors, and we discuss fast implementations of maximum likelihood estimators based on specific alternating least squares regressions. We demonstrate their practical applicability in simulations and in an application to factor-based forecasts of excess returns.