Combinations in predictive analytics by using machine learning

This paper investigates the predictive accuracy of various forecast combinations by using machine learning techniques and proposes a model selection and hyperparameter optimization process for achieving better accuracy in given set of examples from the energy market. Various econometric models are estimated prior to the combination process by utilizing autoregressive integrated moving average (ARIMA), Holt-Winter’s exponential smoothing and other leading univariate forecasting models. Neural networks and support vector machine are employed to find the most accurate combinations of those models. Validation sets are used for finding the most accurate (post- sample) combinations with certain hyperparameter configurations. Models and combinations are compared in the test set based three accuracy metrics. Neural network combinations with inputs generated from Autoregressive- Distributed Lag model (ARDL) empirical mode decomposition (instinct mode functions) performed significantly better than single models and other combinations in majority of given data sample.