Predicting chaotic coal prices using a multi-layer perceptron network model

Coal prices in China has risen steadily and been unusually volatile because of the state's contradictory policies in coal sector. This paper sets up a multi-layer perceptron network model to make short terms prediction after identifying the chaotic characteristics of coal price. Coal prices of Qinhuangdao port are selected as the experiment data. Firstly, coal price time series was studied from the chaotic point of view. Three classic indicators: the maximum Lyapunov exponent, the correlation dimension and the Kolmogorov entropy are adopted to verify the chaotic characteristic. Then a multi-layer perceptron model is proposed to predict the trend of the chaotic coal price. Topology of the MLP 3−11−3 is described in detail. Four measurements in level and directional prediction, namely, mean absolute percentage error, root mean square error, direction statistic and THEIL index, are used to evaluate the performance of the model. The selected model better recognizes the pattern and nonlinear characteristic of the coal price time series compared to the autoregressive integrated moving average model and MLP m−nh−1 model.