Daily Sugar Price Forecasting Using The Mixture Of Local Expert Models
Free (open access)
B. de Melo, C. L. Nascimento Júnior & A. Z. Milioni
This article concerns the application of the Mixture of Local Expert Models (MLEM) technique to predict the daily prices of the Sugar No. 14 contract in the New York Board of Trade (NYBOT). The MLEM technique can be seen as a Data Mining tool that performs data exploratory analysis and mathematical modeling simultaneously. Given a set of input-output data points, the basic idea is as follows: 1) firstly, considering only the input side of the training data, a Kohonen Neural Network is used to divide the data into non-overlapping clusters of points, 2) several modeling techniques are then used to construct competing models for each cluster, considering only the data points in that cluster, 3) the best model for each cluster is then selected and called the Local Expert Model. Finally, the output of all Local Expert Models are combined by a so-called Gating Network (a Radial Basis Function) that considers: a) the distance of the input data point to the center of the cluster of data points used to generate each Local Expert Model, and b) the size of the region of the input space taken by each cluster of training data points. The following modeling techniques were used to develop the Local Expert Models: Artificial Neural Networks, Multiple Regression and Carbon Copy. For comparison purposes, the same modeling techniques are also evaluated when acting as Global Experts, i.e., when the technique uses the entire data set without any clustering. Keywords: mixture of local experts, clustering, artificial neural networks, multiple regression analysis, time series forecasting. Division of Mechanical & Aeronautical Engineering, Brazil Instituto Tecnológico de Aeronáutica,
mixture of local experts, clustering, artificial neural networks, multiple regression analysis, time series forecasting.