Research On Multi-feature Futures Price Trend Prediction Based On Transformer Model

The futures market,as a physical commodity-based market organization,plays a crucial role in stabilizing the spot market,hedging against price risks,and enhancing market liquidity.It is of great social and economic value to accurately analyze and forecast the futures market.It is a great challenge to extract valid information from the complex futures market to predict futures price trends.In this thesis,in order to improve the accuracy of futures price trend prediction,deep learning models are applied to futures price prediction from the perspectives of model and feature,and compares different models to obtain a better prediction model.Additionally,news and commentary texts are transformed into structured data and integrated as new features to enhance prediction accuracy.Its key contributions are as follows:Based on the target of futures price trend prediction,this thesis selects the daily frequency trading data of five futures varieties,namely,CSI 300 stock index futures,steel rebar futures,iron ore futures,fuel oil futures,and petroleum asphalt futures,as the research target,and also obtains the commentary text and industry news text data from "Futures Bar" of the East Money Information.In this thesis,we construct the Transformer model to predict the price trends of futures varieties and explore the prediction effects of the Transformer model and the comparison model under different time windows.The study yielded the following experimental results: Firstly,the model’s effectiveness varied across different time windows.The experimental results show that when the time window is 7 days,the accuracy of Transformer,CNN,and gradient enhancement models is better than other time windows.When the time window exceeds 14 days,the prediction accuracy of the model decreases,indicating that the model cannot effectively capture relevant historical information when the time window is too long.Secondly,when predicting futures price trends based on technical indicator characteristics,the Transformer model outperformed the benchmark model.Each type of futures showed high prediction accuracy on the Transformer model,with a prediction accuracy of 61.7% for rebar futures.These results demonstrate the effectiveness of using the Transformer model for futures price trend forecasting.Thirdly,the inclusion of text features improved the forecasting accuracy of all three neural network models.Combining news and commentary text indicators produced the most significant improvements.From a model perspective,the Transformer model showed the greatest improvement,with an average increase of 1.12% for oil asphalt futures.This thesis improves the accuracy of futures price trend prediction from two perspectives: model and features.The Transformer model is shown to outperform other comparable models,and the addition of commentary and news text indicators to the model contributes to its effectiveness,enriching research content and having practical significance for futures market practice.