Futures Data Analysis With Mutual Information Continual Graph Representation Learning

In recent years,the continuous development of big data,artificial intelligence and data science technology has made breakthroughs in quantitative investment in strategy richness,model complexity and transaction speed,but China’s futures market related research and trading practices are not mature,containing huge arbitrage space.On the other hand,the existence of futures leverage also brings huge risks,and the prediction of futures prices is always the focus of government management departments and investment subjects.By analyzing the high-frequency price characteristics of the domestic futures market,this paper predicts the intraday prices of various futures varieties,so as to help investors avoid risks and increase returns.The existing research methods for futures price prediction mainly include econometric methods and machine learning algorithms,and the non-stationary and nonlinear characteristics of high-frequency financial data make it difficult for econometric models and machine learning algorithms to solve long-term dependencies in the sequence,fortunately,the development of deep learning has innovated the original feature extraction system,which can more accurately mine the feature patterns of financial data,and is widely used in futures price prediction and prediction.However,most of the existing models ignore the spatial correlation and spatiotemporal heterogeneity characteristics of futures contracts,and only learn the information at the time domain level.Based on the differences in the long-term and short-term signals of spatiotemporal graph convolutional networks STGNNs and high-frequency futures data,this paper proposes a heterogeneous continuous graph neural network HCGNN based on mutual information loss to predict the price of multi-variety futures.The main work is as follows:(1)Aiming at the temporal autocorrelation of futures prices and the spatial correlation between various variable series,the spatiotemporal characteristics of futures prices are dynamically obtained through the self-attention mechanism and the spatiotemporal graph neural network.Considering the long-term and short-term volatility and trend differences in futures prices,four heterogeneous tasks of price prediction,mean reversion,range regression and variable point classification are designed to learn the long-term and short-term trends of prices respectively.Using the importance of the original features and model features,the mutual information loss between the predicted value of the model and the real value,the parameter optimization iteration is carried out by maximizing the mutual information loss,which solves the catastrophic forgetting problem in multi-task continuous learning,and verifies the effectiveness of the model in the main contract data of 49 active varieties in the domestic futures market from February to March 2022.(2)This paper takes STGNN,SSTGNN,MTGNN and Stem GNN as the baseline model,predicts the futures price of 1 minute and 15 minutes,in the short-term forecast,the HCGNN model is the best prediction performance for various tasks,in the long-term forecast,the ability of HCGNN to capture the rise and fall signals is significantly higher than the baseline model,and the model’s ability to predict the long term and short-term is more balanced.Based on the effectiveness of the overall model,a series of ablation experiments are designed,and it is found that under the traditional continuous learning mechanism,the increase of tasks will reduce the performance of the model,while the addition of learning tasks in the mutual information loss can improve the performance of the model,indicating that the catastrophic forgetting of parameters in traditional multi-task learning can be effectively alleviated by updating the parameters of mutual information loss control.