Evolving Financial Trading Strategies with Vectorial Genetic Programming
Rui Menoita, Sara Silva
TL;DR
The paper tackles the problem of automatically evolving profitable trading strategies for financial markets, a challenging task under the efficient market hypothesis. It employs Vectorial Genetic Programming (VGP) and introduces two novel variants—Complex Vectorial GP (CVGP) and Strongly-Typed Vectorial GP (STVGP)—to learn trading rules from vectorial inputs that encode historical data and technical indicators. Experiments on three instruments over a seven-year span demonstrate that standard GP often underperforms, while STVGP consistently yields among the best results, with CVGP also showing strong potential; ROI-based fitness is augmented by a win-rate factor when profitable and includes an inactivity penalty to encourage active trading. The findings underscore the viability of automated, interpretable trading-rule discovery using vectorial and typed GP frameworks and point to fruitful avenues such as multi-agent collaboration, richer feature sets, and larger search spaces for further improvements.
Abstract
Establishing profitable trading strategies in financial markets is a challenging task. While traditional methods like technical analysis have long served as foundational tools for traders to recognize and act upon market patterns, the evolving landscape has called for more advanced techniques. We explore the use of Vectorial Genetic Programming (VGP) for this task, introducing two new variants of VGP, one that allows operations with complex numbers and another that implements a strongly-typed version of VGP. We evaluate the different variants on three financial instruments, with datasets spanning more than seven years. Despite the inherent difficulty of this task, it was possible to evolve profitable trading strategies. A comparative analysis of the three VGP variants and standard GP revealed that standard GP is always among the worst whereas strongly-typed VGP is always among the best.