Research And Implementation Of Commodity Futures Trend Prediction Based On Deep Learning

In recent years,deep learning technology has been continuously applied in all walks of life.This paper applies deep learning technology to the financial field to predict future price trends of commodity futures.Numerical data includes transaction data and index data.Numerical data can help investors to judge the trading point and price trend.Text data includes news,institutional analysis articles,investors' comments and other forms.The future price trend can be judged by analyzing the emotions transmitted by text data.Therefore,the analysis of these two types of data is of great significance to investors,which can help investors make decisions.In the trend prediction model of numerical data,there is no model that directly performs self attention operation on numerical data.Besides,the model only trains for a single commodity,and does not play the role of related commodities.In the trend prediction model of text data,the model feature extraction ability is poor,and the key features cannot be highlighted.In view of the problems of the above two models,this paper has done the following research:1)Constructing a futures numerical data trend prediction model that adds a self-attention layer before the two-layer LSTM model.Existing trend prediction models for futures numerical data do not have self-attention calculations on the data from the beginning.The model verifies that direct self-attention calculations on numerical futures data can improve the prediction effect of the model.2)Based on the self-attention mechanism,a futures numerical data trend prediction model fused with related commodity data is constructed.The current known models do not use the data of related commodities,and they are all single-input single-output models.This model is a two-input two-output model.The input is the data of two related commodities,and the output is the predicted trend of the two.There is a certain correlation between related commodities.This model integrates the information of the two through the self-attention mechanism,which brings great gains to the effect of the model.It also provides a way to improve the effect of the model by fusing the relevant information of the two in a disguised form.3)A text sentiment analysis model based on a hybrid neural network is constructed.Most of the existing text data trend prediction models have poor performance in capturing long-distance dependence and semantic word order feature extraction and do not have the ability to highlight key features,so the model effect is not ideal.This paper constructs a hybrid neural network for text sentiment analysis to judge the futures price trend.It mainly uses Bi LSTM to extract long-distance dependence,LSTM-CNN structure to extract semantic word order features and self attention mechanism to learn key features.4)Model comprehensive effect analysis.In actual transactions,traders often analyze from multiple perspectives.Therefore,this paper first conducts experimental analysis on the above two models,and then combines the two to analyze the overall effect.