A0628
Title: Test-based trimming for combined forecast
Authors: Viktor Eriksson - Uppsala University (Sweden) [presenting]
Abstract: The common practice of combining all single forecasts falls short on not trimming off poor forecasts. We propose a new method for trimming in the combination of forecasts that utilizes a test statistic for differences in the AIC. Application on M4-data shows that test-based trimmed combination forecast typically has better forecasting performance than forecast produced by the best model. The improvement in forecasting performance by test-based trimming is greater for uniform combination weights than it is for Akaike combination weights.
