Building multi-factor stock selection models using balanced split regression trees with sorting normalisation and hybrid variables

Abstract

This research employed regression trees to build the predictive models of the rate of return of the portfolio and conducted an empirical study in the Taiwan stock market. Our study employed the sorting normalisation approach to normalise independent and dependent variables and used balanced split regression trees to improve the defects of the traditional regression trees. The results show (a) using the sorting normalised independent and dependent variables can build a predictive model with a better capability in predicting the rate of return of the portfolio, (b) the balanced split regression trees perform well except in the training period from 1999 to 2000. One possible reason is that the dot-com bubble achieved its peak in 2000 which changes investors' behaviour, (c) during the training period, the predictive ability of the model using data from the bull market outperforms the model using data from the bear market.