Sparse convex combinations of forecasting models by meta learning

Abstract

Combining forecasting methods is a well known successful strategy that improves the accuracy of prediction by decreasing the variance of forecast. One drawback of this approach is that the computation of the final forecast could be extensive requiring the fitting of all the models of the combination. This could be problematic, especially in business applications, when accurate forecasts are required in a short time. On this matter, two different meta-learning systems, based on a deep neural network, are proposed in order to automatically provide sparse convex combinations of forecasting methods. Depending on time series features, both the meta-learners recommend to combine only a subset of the available forecasting methods. There is no need to compute the forecasts of the left out methods and therefore it is evident the computational savings of the strategy and its reliability in real time applications. Both the meta-learners have been trained and evaluated on the M4 competition dataset by building sparse combinations of nine forecasting methods. Results of the experiments show that the proposed meta-learners achieve predictive performances comparable with the top-ten ranked benchmarks of the competition by combining on average five to six of the nine methods available.