A hybrid forecasting model based on deep learning feature extraction and statistical arbitrage methods for stock trading strategies

Abstract

The time series data of financial markets are nonlinear, owing to rapid data accumulation. Thus, research on stock price prediction has always been a challenge. This study proposes a quantitative trading strategy that combines basic quantitative trading rules and deep learning methods to help investors realize arbitrage. We combine basic quantitative trading arbitrage with deep learning frameworks to fully extract market characteristics and develop trading strategies for investors. The hybrid forecasting model is a new signal‐trading system that uses a genetic algorithm to obtain optimal parameters for the technical indicator timing method of the moving average price. The deep learning structure of the CNN‐Bi‐LSTM, with the attention mechanism and modified loss function, optimizes the trading signal to achieve local optimization. Its core concept is to determine the trading signal through the local trend of price fluctuations and then correct the trading signal through the prediction results. A‐shares in the Chinese market trading data are used as the statistical arbitrage analysis process to output actual trading signals and verify the effectiveness of the system. The results demonstrate that an arbitrage strategy based only on moving average trading rules is ineffective. With the optimization of the deep learning CNN‐Bi‐LSTM framework, the arbitrage ability improves significantly. The optimized strategy can increase the final profit by 1.6042 to the greatest extent. The annualized revenue increased by 35.16%, and the winning rate increased by 15.22%. In addition, we consider the transaction costs during the simulated transaction process. An optimized trading strategy can effectively seize arbitrage opportunities; hence, its profitability and stability are significantly improved.