Integration of Genetic Algorithms and Trading Simulation for Enhanced Stock Price Prediction and Optimization of Quantitative Trading Strategies

We propose a novel methodology integrating genetic algorithms (GAs) with quantitative back-testing to optimize a multilayer perceptron (MLP) model for stock price prediction and trading strategies. Addressing three major challenges in financial time series forecasting—the low signal-to-noise ratio (SNR) inherent in financial data, neglect of critical trading constraints during model training, and homogenization of predictive signals due to lack of diversity—we employ the ZigZag Indicator as a target transformation function to filter out minor price fluctuations and capture significant market trends, thereby improving the SNR. Specifically, our approach enhances the SNR from − 10.19 dB to − 7.28 dB, representing a 28.56% improvement. Unlike conventional models focused solely on minimizing statistical errors, our method incorporates back-testing criteria and trading constraints directly into the MLP’s training and optimization process, ensuring both high predictive accuracy and robust trading performance. Furthermore, we introduce an enhanced parent selection mechanism within the genetic algorithm by implementing a convex optimization framework that facilitates multi-objective optimization by maximizing performance metrics—including the Sharpe ratio, Sortino ratio, and trend accuracy—while minimizing correlations among selected models to maintain diversity. Extensive experimental results demonstrate that our approach significantly surpasses existing methods in stock market prediction. These findings highlight the immense potential and versatility of GAs in navigating complex financial datasets and offer substantial improvements in both prediction accuracy and trading strategy optimization.