Selecting Neural Network Forecasting Models Using the Zoomed-Ranking Approach

In this work, we propose to use the Zoomed-Ranking approach to ranking and selecting artificial neural network (ANN) models for time series forecasting. Given a time series to forecast, the Zoomed-Ranking provides a ranking of the candidate models, by aggregating accuracy and execution time obtained by the models in similar series. The best ranked model is then returned as the selected one. In order to evaluate this proposal, we implemented a prototype to rank three ANN models for forecasting time series from different domains. In the experiments, the rankings of models recommended by Zoomed-Ranking were significantly correlated to the ideal rankings.