Forecast the price of chemical products with multivariate data

Sales price of staple commodities plays an important role in human life and reflects production and sales of enterprises, so predicting the price accurately is of great significance. The price of chemical products has the characteristics of time series, nonlinear, unstable, etc, and has relationship with multiple variables which are affected by seasons, national policy and macro-economy. Therefore, predicting their price has become a challenging task. In this paper we propose a new prediction algorithm that exploits multivariate data with analysis including crawled web data related to chemical products and expert experience data. History data is first disposed and analyzed to build statistic and machine learning forecasting models. Then sentiment analysis is performed based on related data crawled from the internet measured by text analyzing techniques. Finally expert experience on forecasting the price is used to optimize the prediction results. We use methanol as an example to evaluate the accuracy of prediction results tracked for eight months, the MAPE (average absolute percent error) of our method is 2.91% better than other models. Compared with traditional prediction models, our model based on multivariate data has higher accuracy.