A framework for optimization of pattern sets for financial time series prediction

Abstract

In this paper, a framework is introduced for generating human-interpretable structures, here called pattern sets, for short-term prediction of financial time series. The optimization is carried out using an evolutionary algorithm, which is able to modify both the structure and the parameters of the evolving pattern sets. The framework has been applied in two different modes: A tuning mode, in which the user provides a starting point in the form of loosely defined pattern set, and a discovery mode, in which the starting points consist of random pattern sets. The best results were obtained in the tuning mode, for which the top-performing pattern sets gave strongly statistically significant results in excess of one-day market returns (p-values below 0.0007 and, in many cases, even below 0.0001) over validation data (not used during optimization) for two data sets, involving stocks with large and small market capitalization, respectively. The average one-day returns ranged from 0.518 to 1.147%, with one-day Sharpe ratios ranging from 0.138 to 0.258.