Futures Trend Forecasting Based On Dilated Causal Convolution And Multi Head Self Attention

Futures trend forecast is an important research field in quantitative trading.With the rapid growth of artificial intelligence technology and computer computing power,futures trend forecast extends from the traditional mathematical statistical analysis method to the machine learning for modeling and analysis.Deep learning,as a machine learning method based on deep artificial neural network,has stronger ability of feature extraction and learning.Deep learning is a data-driven algorithm,the massive and real datasets have laid the data foundation for it.By collecting historical financial data in futures trading,a trend forecasting model established by using deep learning technology can forecast the rise and fall categories and change ranges of futures price trend,so as to provide guidance for investors to formulate automatic quantitative trading strategies and also some support for the financial transaction supervision department.In the existing research on futures trend forecast,people often only study and predict a single futures time series data,ignoring the factors that the futures trend is affected by its related varieties.The prediction model usually adopts cyclic sequence networks such as Long Short-Term Memory,which has limitations in network parallel computing.Aiming at the research problem of multiple correlated futures data and trend prediction,a multivariate time series prediction network using multi head extended causal convolution and multi head self attention mechanism is proposed to learn the multiple futures trading data and predict the price trend of the target futures.In the study,the futures varieties of rebar are studied,and the Pearson correlation coefficient is used to determine three related varieties,namely hot rolled,coke and iron ore.The dataset obtained and constructed is the five minute futures trading data from January 5,2015 to December 31,2021.In order to improve the quality of the dataset,the complete ensemble empirical mode decomposition with adaptive noise method is used to denoise the financial dataset,and the improved three boundary method is selected to mark the rise and fall categories and change amplitude values of the price trend.The data features are expanded by selecting a variety of financial technical indicators.In order to evaluate the performance of the multi-variety futures trend prediction model,the multivariate time series data composed of four varieties of rebar,hot rolled,coke and iron ore were selected for experiments.The results show that the accuracy of the proposed model is 9.3% higher than LSTM(Long Short-Term Memory),6.4% higher than TCN(Temporal Convolutional Network),and 3.4% higher than the multivariate time series prediction network TST(Time Series Transformer)proposed in 2021,which verifies the effectiveness of the proposed model.