Assuming a number of forecast models are available to a user: How can these be usefully combined to improve forecast accuracy on unseen data compared to choosing individual models? Which models should be selected and should the learned weights be corrected?
If this topic sounds interesting to you, we would like to refer you to the just published research article "Linear Hybrid Shrinkage of Weights for Forecast Selection and Combination", co-authored by Felix Schulz with Thomas Setzer and Nathalie Balla.
In this paper, a new methodological procedure for simultaneous selection, weighting and shrinkage of forecast weights is presented. Using information criteria from statistical learning theory, forecast models are first selected and, depending on the selection state, shrunk linearly from their optimal weights toward the mean or toward zero. Simulation results show conditions (scenarios) under which the presented method achieves higher accuracy than linear and nonlinear shrinkage methods from optimal weights to the mean, estimated optimal weights, and the simple averaging of forecasts.