Non-Stationary Demand Forecasting Based on Empirical Mode Decomposition and Support Vector Machines

A company performance strongly depends on its ability of delivering the right quantity of of a given product at the time customers require. Even though some demand forecasting techniques have been developed, they have commonly used simplifying assumptions that limit their use like assuming that the relation between the inputs and the output is linear, for example. Therefore, machine-learning techniques, such as Support Vector Machines (SVM), arise as a promising alternative for accomplishing demand forecasting. SVM has the advantage of performing well in cases where the relationship between input and output data is unknown, and thus has brought good results when applied in different contexts. However, SVM presents some limitations in predicting non-stationary series. In this context, a method called Empirical Mode Decomposition (EMD) has been adopted for decomposing non-stationary and nonlinear time series into a group of Intrinsic Mode Functions (IMFs). Moreover, SVM performance strongly depends on the values of real-valued parameters, which need to be tuned to enhance the predictive ability of the model. This situation gives rise to the model selection problem, which may be solved by heuristics such as Particle Swarm Optimization (PSO). Therefore, this work proposes a non-stationary demand forecasting methodology based on EMD-PSOSVM. An example in the context of the food industry is presented and we compare the results obtained by the proposed methodology against the ones returned from a plain PSO-SVM. The results show that the proposed EMD-PSO-SVM presented superior performance.