Revisiting the Forecast Combination Puzzle: An Empirical Study

Honours Research Project at Monash University (2023)

Supervisor: David Frazier

This thesis demonstrates that the forecast combination puzzle is tightly related to the in-sample fit of the constituent models used in the analysis. The forecast combination puzzle refers to the common finding that an equally-weighted forecast combination often outperforms an optimally-weighted forecast combination calculated via a sophisticated scheme. We show that when constituent models have similar in-sample fit, the puzzle will be in evidence; it is ambiguous otherwise. We analytically show the relationship between the estimated optimal weight and the constituent models in terms of point combinations using mean squared error and empirically confirm these findings in density combinations using log score. As an additional contribution, the puzzle is shown to be evident in both time series and cross-sectional settings.

The thesis and relevant file are in the Research Thesis folder.

All code and data are in the R folder.

I cannot be held responsible for any use and mis-use of my code.

Sapphire Li

xlii0145@student.monash.edu

Monash University

October 2023