Time Series Forecasting Through a Dynamic Weighted Ensemble Approach

Time series forecasting has crucial significance in almost every practical domain. From past few decades, there is an ever-increasing research interest on fruitfully combining forecasts from multiple models. The existing combination methods are mostly based on time-invariant combining weights. This paper proposes a dynamic ensemble approach that updates the weights after each new forecast. The weight of each component model is changed on the basis of its past and current forecasting performances. Empirical analysis with real time series shows that the proposed method has substantially improved the forecasting accuracy. In addition, it has also outperformed each component model as well as various existing static weighted ensemble schemes.