Forecast Combination with Truncation of the Negative Weights
This thesis examines the predictive power of forecast combinations when the negative weights are not ignored. Instead of removing the negative weights, we use truncation to limit the negative weights to a select range of levels. Using the Survey of Professional Forecasters from the European Central Bank, a dataset with high correlations in forecast errors, this thesis reports a positive performance compared to both equal weights and non-negative weights in five out of six cases. We also show that the truncation is also able to improve the forecast when the selection of the threshold is out-of-sample. Additionally, illustration using conditional-biasadjusted weights, and a simulation study has been carried out to demonstrate the robustness of the method.