Multi-scale Combined Model And Its Applied Research On Commodity Price Forecasting

In recent years, with the rising of China’s external dependence degree on commodities, the international commodity price fluctuates violently and becomes higher overall, which caused a great impact on China’s economy. Predicting the price fluctuation and trend effectively has a very important significance on the national’s economic policy formulation and enterprises’decision making.Based on the idea of decomposition-reconstruction-integration, this paper built a new multi-scale combined forecasting model by using empirical mode decomposition (EMD), run-length-judgment method, artificial neural network (ANN), support vector machine (SVM) and time series methods. The model building process is as follow:firstly used empirical mode decomposition (EMD) to decompose the time series; secondly, proposed a new idea to use run-length-judgment method to reconstruct the component sequences and generally obtained high-frequency, medium-frequency, low-frequency and trend term four parts; thirdly, selected ANN, SVM and time series methods to predict the four parts respectively; finally, select SVM model to integrate the four parts’prediction results.Select crude oil, copper and wheat respectively as the representative of energy commodities, basic raw materials and bulk agriculture commodities. This multi-scale combined model we built was used to do empirical research. Based on the analysis of the characteristics of the these three commodity markets, firstly we used EMD to decompose these three price sequences, results showed that oil and wheat price can be decomposed into7IMF and1R, and copper price can be decomposed into6IMF and1R; then we used run-length-judgment method to reconstruct these sub-sequences, we found that oil price sub-sequences can be reconstructed into four items including high frequency, medium frequency, low frequency and trend item which respectively represent the irregular factors, seasonal factors, the impact of major events and long-term trends, and copper and wheat price sub-sequences can only be reconstructed into three items; after that, we selected ANN, SVM and time series methods to predict the four parts respectively; and finally web selected SVM integrate model to integrate the four parts and got the final predict result. The above empirical research showed that comparing with the GARCH, GM (1,1), Elman and all the other single models, and comparing with ARIMA-SVM combined model and EMD-SVM-SVM multi-scale combined model, the multi-scale combined model this paper built obtained the best forecast result, which is more suitable for commodity price forecast. On one hand, the multi-scale combined model this paper built enhanced the prediction accuracy, on the other hand, it gave the reconstructed sequences economic implications, which can generally be considered that the commodity price series are constituted by long-term trend, seasonal factors, major events and irregular factors influencing factors. So this multi-scale combined model is a "data-driven modeling" and "theory driven modeling" combination prediction method, which is suitable for commodity price forecast.