Futures Price Forecasting Method Based On The Feature Fusion LSTM Model

With the continuous development of the economy,the financial assets that can be invested by Chinese residents are gradually increasing,and the futures is also chosen by more and more investors.Futures trading refers to the way that buyers and sellers sign a standard agreement to trade a certain commodity on a decided date in the future.It has low cost and allows investors to sell first and then buy,which makes the investment risk increases.In recent years,along with the continuous development of computer technology and the maturation of interdisciplinary subjects,futures market participants and researchers expect more scientific and objective means to conduct market analysis and predict the future trend of prices,thereby reducing the one-sidedness of subjective judgments and avoiding investment risks.The existing analysis of futures prices mainly includes fundamental analysis,traditional measurement models,and machine learning algorithms.Fundamental analysis relies on analysts’ empirical knowledge,which is subjective and one-sided.Traditional econometric models have no obvious advantages in nonlinear data such as finance,and machine learning algorithms such as SVM have not solved the long-term dependence problem in price series.Later,the emergence of a memory neural network solves the weakness in the above models and has a strong prediction effect on long-term,nonlinear financial data.However,in these neural network prediction models,most researchers only use basic price indicators and common technical indicators as input features and do not consider the impact of other complex features.To integrate more features into the futures price prediction model,this thesis proposes a feature fusion futures prediction model based on data mining and long short-term memory neural network(LSTM).Considering the impact of price patterns,this thesis designs a price pattern feature extraction algorithm.The algorithm slides the continuous price data for a period to form a price pattern vector and then uses the K-means clustering algorithm to analyze the price pattern to extract the similarity between the price patterns.At the same time,considering the impact of investor sentiment on price changes,the sentiment analysis is carried out on texts posted in the futures bar,and weight optimization rules are designed to quantify investors’ sentiment values more accurately.After the extraction of new feature indicators is completed,it is integrated with basic price indicators and key technical indicators,and a joint LSTM neural network is chosen for model training and verification.Finally,the model proposed in this thesis is experimented on several futures products of Shanghai Futures Exchange and Dalian Commodity Exchange and compared with similar priceprediction models.The results show that the model proposed in this thesis has excellent performance on different futures products,and also has better results than other similar algorithms.This fully verifies the advantages of LSTM and the effectiveness of feature fusion.At the same time,investors can guide the formulation of strategies according to the prediction results of the model.Therefore,the model proposed in this thesis has great significance and also provides new views for other such studies.