Ensembles for Time Series Forecasting
Proceedings of the Sixth Asian Conference on Machine Learning, PMLR 39:360-370, 2015.
This paper describes a new type of ensembles that aims at improving the predictive performance of these approaches in time series forecasting. Ensembles are recognised as one of the most successful approaches to prediction tasks. Previous theoretical studies of ensembles have shown that one of the key reasons for this performance is diversity among ensemble members. Several methods exist to generate diversity. The key idea of the work we are presenting here is to propose a new form of diversity generation that explores some specific properties of time series prediction tasks. Our hypothesis is that the resulting ensemble members will be better at addressing different dynamic regimes of time series data. Our large set of experiments confirms that the methods we have explored for generating diversity are able to improve the performance of the equivalent ensembles with standard diversity generation procedures.