An Empirical Study of Multi-Objective Algorithms for Stock Ranking

Abstract

Quantitative models for stock selection and portfoliomanagement face the challenge of determining the mostefficacious factors, and how they interact, from largeamounts of financial data. Genetic programming usingsimple objective fitness functions has been shown to bean effective technique for selecting factors andconstructing multi-factor models for ranking stocks,but the resulting models can be somewhat unbalanced insatisfying the multiple objectives that portfoliomanagers seek: large excess returns that are consistentacross time and the cross-sectional dimensions of theinvestment universe. In this study, we implement andevaluate three multi-objective algorithms tosimultaneously optimise the information ratio,information coefficient, and intra-fractile hit rate ofa portfolio. These algorithms the constrained fitnessfunction, sequential algorithm, and parallel algorithmtake widely different approaches to combine thesedifferent portfolio metrics. The results show that themulti-objective algorithms do produce well-balancedportfolio performance, with the constrained fitnessfunction performing much better than the sequential andparallel multi-objective algorithms. Moreover, thisalgorithm generalises to the held-out test data setmuch better than any of the single fitnessalgorithms.