Genetic Programming Application for Combining Directional Changes Indicators in the International Stock Market

The majority of algorithmic trading studies often rely on datasets with fixed physical time intervals, like hourly or daily data, resulting in a discontinuous representation of time. Directional Changes (DC) is an alternative method that transforms these datasets from physical time intervals into event-driven sequences, allowing for unique price analysis. While previous work has primarily focused on introducing novel DC-based indicators, there has been limited exploration of training these indicators using machine learning algorithms. At the same time, Genetic Programming (GP) has shown promise in algorithmic trading, but its performance within the DC framework has not attracted much attention.

Our study delves into the potential of DC as a complementary tool to traditional time-based approaches and explores the incorporation of DC-based indicators in algorithmic trading. Thus, we first propose a novel approach, combining 28 DC indicators through a genetic programming (GP) algorithm, referred to as GP-DC. Furthermore, within the same GP framework, an alternative trading strategy is presented, fusing 28 DC indicators with 28 traditional physical time-based technical analysis indicators, creating the GP-DC-PT algorithm. We extensively evaluate the return and the risk of DC-based trading strategies across a dataset containing 220 sets from 10 international financial markets. We compare the performance of these GP-based strategies against the GP-PT, a non-DC-based GP using strategies based on physical time intervals.

Our results demonstrate that both GP-based algorithms significantly outperform GP-PT. GP-DC-PT also reduces risk compared to GP-DC. In the next step, to investigate whether a multi-objective optimisation approach could improve the performance of GP-based strategies in the market, we propose a novel algorithm that follows the same process as NSGA-II but uses GP training by a combination of the DC indicators and technical indicators. The proposed algorithm demonstrated robust performance by benchmarking with the GP-based single objective optimisation and statistically outperformed the benchmark.

Based on joint work with Michael Kampouridis and Panagiotis Kanellopoulos that appeared in PPSN’22