Stock Price Prediction Based on Non-Stationary Transformers Model

In recent years, a growing number of investors and researchers have been applying deep learning techniques to stock price prediction in the stock market, yielding promising predictive outcomes. This study employs the newly introduced deep learning model, Non-Stationary Transformers, for stock price forecasting. This model extends the Transformer architecture by incorporating a Series Stationarization module and a De-stationary Attention module. This augmentation not only enhances the predictability of stock price time series data but also effectively harnesses the temporal modeling capabilities inherent in deep models. To demonstrate the effectiveness of the Non-Stationary Transformers model in stock price prediction, we conducted prediction experiments and backtesting experiments, comparing it with four other deep learning models. The prediction experiments revealed that the Non-Stationary Transformers model outperformed others, yielding the lowest Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) values. Furthermore, the backtesting experiments also demonstrated that the Non-Stationary Transformers model achieved the highest Return Ratio, Sharpe Ratio, and an appropriate Max Drawdown. Both sets of experiments indicate that the Non-Stationary Transformers model provides reliable guidance for investors in making stock market investment decisions.