Model selection in reconciling hierarchical time series

Model selection has been proven an effective strategy for improving accuracy in time series forecasting applications. However, when dealing with hierarchical series, apart from selecting the most appropriate model, forecasters have also to select a suitable method reconciling base forecasts produced each make sure they are coherent. Although some methods like minimum trace strongly supported both theoretically and empirically forecasts, there still circumstances under which might not produce accurate results, being outperformed by other methods. In this paper we propose approach dynamically reconciliation leading more coherent forecasts. The approach, call conditional forecasting, is based on machine learning classification that use features hierarchy. Moreover, it allows be tailored according measure of preference level(s) interest. Our results suggest can lead significantly than standard approaches, especially at lower levels.