A Hybrid Approach to Combining CART and Logistic Regression for Stock Ranking

Abstract

The performance of a stock relative to a suitable peer group is often influenced by a multitude of factors and their interactions. Traditional parametric models, albeit very useful, are often inadequate in capturing complicated relationships. In contrast, the nonparametric decision tree technique, such as classification and regression trees (CART), is more capable of capturing any nonlinearities and high-order interactions among stock characteristics, with the additional convenience of graphically representing the model, but discontinuous and coarse-grained responses produced by CART are potentially undesirable. In contrast, traditional regression models such as logistic regression produce a smooth response surface, which is more tractable in practice. The authors use a hybrid approach that takes advantage of the strengths in both parametric (logistic regression) and nonparametric models (CART). An application of this sophisticated technique to North American defensive stocks demonstrates its usefulness.