Stock selection combining rule generation and risk/reward portfolio optimization

Abstract

Datamining Technologies such as rule generation applied to historical database of equities data can be combined with optimization-based portfolio selection methodology. We report on issues concerning the combination of these technologies and a numerical exercise covering ten years of equity data. Rule generation is performed on the database to classify equity performance into subset from subsets from which next-period performance of equity returns can be successfully predicted. Such rules are based on relationships between up to 40 attributes (P/E ratios, running averages, etc). The relationships are used to identify a subset of historical data points which are then used as an input to a risk/reward optimal stock selection program. The selection program can accommodate constraints customary in applications of the Markowitz mean/variance portfolio management methodology. It can be tuned to model downside risk formulations such as semivariance.