An adaptive forecasting strategy with hybrid ensemble models

Abstract

We propose hybrid ensemble models for time series forecasting. A hybrid ensemble combines the output of several different models by a weighted mean that forms the final forecast. The final hybrid ensemble model consists of several individual models: A nearest neighbor/trajectory ensemble model, a feed-forward neural network ensemble, a trend cycle model, an autoregressive model and an ensemble model based on the returns of the time series. The best performing models with respect to a left-out part of the time series are selected by cross-validation and are combined by taking the inverse SMAPE prediction error as a weight for the combination of the respective single forecasts. We show the application of this approach in the WCCI 2016 CIF challenge.