Fundamental, Sentiment and Technical Analysis for Algorithmic Trading Using Novel Genetic Programming Algorithms

This research focuses on applications of genetic programming (GP) on the topic of algorithmic trading. In recent years, there have been significant advancements in algorithmic trading. Investors typically rely on either fundamental analysis (FA) or technical analysis (TA) indicators, while sentiment analysis (SA) has gained attention in more recent years. As a result, algorithms have become the primary method for developing pre-programmed trading strategies, leading to substantial financial benefits. While each analysis type has been studied individually, their combination has not been explored extensively. Considering the effectiveness of each analysis type, our motivation is to investigate whether combining the indicators from these analysis types can enhance the profitability of trading strategies. Thus, we first introduce a novel genetic programming algorithm, which combines the three analysis types into the same GP structure, wanting to understand the financial advantages of their combination. Following, we present a GP algorithm with a strongly-typed architecture. Each branch of the algorithm's tree represents one analysis type exclusively, allowing for improved exploration and exploitation of the indicator search space. Furthermore, we also showcase a novel fitness function that rewards not only a tree's trading performance, but also the trading performance of its FA, SA, and TA subtrees (branches). Venturing in understanding the performance of the GP algorithm better, we lastly propose a novel GP operator that encourages active trading by injecting trees into the GP population that can perform a high number of trades while achieving high profitability at low risk. To evaluate the performance of our GP algorithms, we conduct experiments on the stocks of 42 international companies. We compare the novel algorithms with the rest of the introduced GP variants, while also comparing the first two proposed GP algorithms with four machine learning benchmarks and a financial trading strategy. For the experiments we use three financial metrics: Sharpe ratio, rate of return, and risk. The results demonstrate that the proposed genetic programming algorithms improve the financial performance of the trading strategies and significantly outperform the respective benchmarks.

Based on joint work with Michael Kampouridis, Panagiotis Kanellopoulos, and Maria Kyropoulou.