Research On Agricultural Futures Price Prediction Based On Text Feature And LSTM

China is one of the world’s leading agricultural countries,meanwhile,its agricultural futures market is gradually developing and maturing.Recently,more and more scholars begin to conduct research on agricultural futures price prediction.Changes in agricultural futures price can be influenced not only by climate,supply and demand,but also by political issue,economic issue,domestic and international policy,investor sentiment and so on——information that can usually be expressed through online information texts.Most of the existing studies have combined influencing factors to forecast agricultural futures prices,while there is still a gap in the study of extracting features from relevant texts to apply to price prediction.Therefore,using the information in the text to more accurately forecast agricultural futures prices is important for agricultural producers,futures investors,industry operators and other participants involved in all aspects of trading.In order to better predict agricultural futures prices,this study proposes a price prediction method that incorporates text features,using wheat futures prices and related online information text as the object of the research.In this study,price and text data from 1 October 2021 to 30 September 2022 were selected.Firstly,Top2 Vec and Snow NLP were used to analyze the text and output sentiment feature values classified by topic,which were combined with price data to build a multivariate timeseries dataset;secondly,the multivariate dataset was used as input to an LSTM model to forecast multiple sets of agricultural futures prices.Other models were also selected for prediction of single column price data;finally,the results of the experiments were compared and analyzed.The study used four indicators to evaluate the predictive effect of the model:adjusted R2,RMSE,MAE and MAPE.The results indicate that the prediction of agricultural futures prices with the addition of textual sentiment features has improved,comparing with the prediction effect of univariate input,and the price prediction effect of fused textual sentiment features under different topic categories also differs.The research method proposed in this study can effectively optimize the prediction model of agricultural futures price and has certain guiding significance for the research and application of agricultural futures price prediction.