Agricultural Futures Price Prediction Based On BERT-LSTM Model

Food security is an important foundation of our national security and Self-Reliance,and it is also an overall strategic issue concerning our economic development and social stability.Our country agriculture development faces the new challenge under the new era background.As planting work is affected by climate,market,policy and other factors,farmers’ income still cannot be effectively guaranteed.In this regard,the No.1 central government document has made specific arrangements for agricultural development.The 14th Five-Year plan also proposes to develop the mode of "insurance+futures" of agricultural products,which can effectively avoid risks through the price discovery of agricultural products futures,ensure food security and promote higher quality development of agriculture.Futures market since the establishment of Zhengzhou grain wholesale market,has more than 30 years of development history.Among them,soybean futures market of soybean No.1,soybean No.2,soybean meal,soybean oil and other related products has been gradually formed since soybean was listed on the stock exchange.Scholars from all over the world have made use of the function of price discovery and risk avoidance to create financial derivatives that can counter spot risks.In agricultural security.At present,the development of agricultural product output insurance,price insurance are relatively mature.The application of Copula model also provides technical support for the accounting of farmers’ income insurance.In terms of futures price prediction,the application of numerous statistical models and deep learning models has greatly improved the accuracy of the models.However,due to the impact of investor sentiment,it is difficult for a single model to accurately predict the futures prices of agricultural products.In order to improve the accuracy of futures price prediction,taking soybean as an example,a total of 1513 soybean futures closing prices and their influencing factors from August 1,2016 to December 2,2022 were selected,including spot price,futures trading volume,port purchase price,port inventory,futures closing price of upstream and downstream products,etc.,as well as news text classification training results.In order to be able to consider both quantifiable data and text data,the model selection is divided into two parts.In the text sentiment classification model,SVM model,random forest model and BERT model are selected.In recent years,BERT model is widely used in the process of natural language processing,is a pre-trained model after a large number of corpus training,which mainly uses the enconder part in Transform to learn the context relationship of the text,for the user,can greatly improve the training efficiency and training accuracy,the training results show that the BERT model has the highest accuracy in the comparison model,and the BERT model is selected as the main model for news text classification.In the process of time series forecasting,the XGBoost model,the GRU model and the LSTM model are compared.The results show that the MAPE value of XGBoost model is 6.5%,the MAPE value of GRU model is 1.78%,and the MAPE value of LSTM model is 1.54%.Finally,the BERT-LSTM combination model is formed as a closing price prediction model for soybean futures.The model evaluation results show that the model accuracy rate is high,the interpretation is strong,can provide a basis for insurance company pricing,but also the theoretical basis for futures companies to implement risk avoidance decisions,and provide a certain degree of protection for the farmers’ income.The accurate prediction of the futures price of agricultural products will also provide new ideas for the subsequent solution of the problems of agriculture,rural areas and farmers.