关键词: 多元网络信息, 新闻信息, 网络关注度, 深度学习模型, 玉米期价预测

Abstract: In China’s financial market, the futures market has the important functions of risk aversion and price discovery. The futures price is the embodiment of the market expectations of agricultural products, and can reflect the actual supply and demand of agricultural products. Corn futures is a representative variety of bulk agricultural products, and it is also one of the most active futures in China’s futures market. It is always at the forefront of commodity futures in terms of trading volume and investor participation. With the improvement of the marketization level of corn, the uncertainty of the market increases sharply, the price of corn futures fluctuates continuously, the risk of investing in grain storage intensifies, and the demand for industrial hedging increases. Effective and accurate price prediction can provide farmers with guidance for planting and trading, give a reference basis for the production and trade of spot enterprises, and transmit efficient information to regulators to enhance the predictability and pertinence of national policy regulation.
Benefiting from the rapid development of the Internet, a large number of unstructured data related to the market has brought an unimaginable amount of information, which not only affects the investment decisions of market participants, but also affects the specific performance of the market. In the performance, news is passively received by the public, especially reports on major policies, emergencies and weather in the place of origin, which can have a great impact on market sentiment, thus promoting the fluctuation of futures prices. On the other hand, the search engine is a way for people to access information actively. Each trader sends a request signal to the network in the form of unstructured keywords, and the engine returns the corresponding content to assist traders in making judgments on market conditions and developing trading strategies. At the same time, the engine also records the search frequency with structured data. In the current futures market, these two forms are also the easiest channels for traders and supervisors to obtain multiple information. Among them, news topics and emotional tendencies are the main expressions of the influence of news information on market fluctuations, and search data is the main reflection of the trend of public concern. The key issues of the research are the measurement of the number of topics, the quantification of sentiment index and the selection of search keywords.
In addition, many studies in the field of price forecasting show that neural network model can efficiently fit and model the observed data due to its structural characteristics and high intelligence, which greatly improves the accuracy of financial time series forecasting.
Based on the above ideas, in addition to the closing price of corn futures, the basic data set carried out in this paper also introduces two source data of relevant news and keyword search volume, and proposes a corn futures price prediction method integrating multiple network information. Since all kinds of major events are mostly published in the form of news text, and there is a lack of effective prior knowledge about the influence of corn futures prices, this method first uses Kullback-Leibler divergence (KL) to determine the key parameters of the topic model Latent Dirichlet Allocation (LDA), and then analyzes the news texts and extracts the topic index. Secondly, the SnowNLP method is used to judge the emotional tendency of news, and the text sentiment index is further optimized from the perspective of time cumulative effect. In addition, the related keyword map of corn futures is constructed with the help of Baidu. According to the keyword map, the Baidu search volume corresponding to each keyword is captured, and the network attention index is synthesized after it is filtered by the Spearman correlation test. In order to integrate the most effective predictor variables and reduce the redundancy of information, the Recursive Feature Elimination (RFE) method is used to construct a combination of predictor variables. Considering that the extracted multivariate indexes are all time series data, the Long Short-Term Memory (LSTM) model is used to complete the final multi-step prediction of corn futures price. The empirical results show that the proposed method is better than the benchmark models SVR, RF and BPNN, and has better prediction performance in the medium and long term. In addition, compared with the LSTM model based on univariate, the proposed method reduces the MAE, RMSE and MAPE indexes by 45%, 41% and 43% respectively, and shows significant performance advantages in Diebold-Mariano (DM) statistical tests. It shows that the proposed method can effectively measure the value of multivariate network information for the prediction of corn futures price and improve the prediction accuracy of the model.
Since this paper focuses on the enabling effect of multiple network information on the prediction of futures price, it only discusses the effectiveness of predictor variables, but does not make a further analysis of the importance of different complex variables. Meanwhile, we only use corn futures price as an financial variable in the research. In the subsequent research work, the relevant content will continue to be improved and expanded.

Key words: multivariate network information, news information, Internet attention, deep learning model, corn futures price prediction