Financial Time Series Forecasting Based on Adversarial Training and Dynamic Weight Design

To address the challenges of financial market instability, high data correlation , anomalies, and noise, traditional adversarial training methods enhance model robustness and nonlinear modeling capabilities by introducing adversarial samples. However, current research primarily relies on single models and singular adversarial methods, which limits the diversity and adaptability of these approaches. Additionally, perturbing only the output layer overlooks the dynamic characteristics present in various layers of the model. To overcome these shortcomings , this paper proposes a strategy that integrates multiple adversarial methods into different deep learning models, combined with a two-stage training and dynamic weight adjustment. In the first stage, the model identifies fundamental patterns by training on the original dataset. In the second stage, adversarial training is introduced to enhance the model’s ability to cope with complex market conditions. Furthermore, the introduction of dynamic weight parameters allows the model to flexibly adjust the weights of the two stages, thereby improving its adaptability to the characteristics of financial data. This paper also introduces two novel adversarial methods: the Hessian perturbation method and the return volatility perturbation method. These methods not only strengthen the model’s nonlinear modeling capabilities but also enhance its adaptability and generalization ability. Experimental results demonstrate that, on the CSI100 dataset, our proposed comprehensive strategy significantly outperforms baseline models in terms of annualized return, information ratio, and maximum drawdown.This achievement not only validates the effectiveness of adversarial methods in predicting financial time series but also illustrates the contribution of the two-stage training framework in enhancing adversarial training, thereby improving the model's adaptability under complex market conditions. Additionally, the introduction of dynamic weight parameters enables the model to flexibly adjust the focus of each stage, allowing for a more effective response to the dynamic characteristics of financial data.