"A Postmortem Analysis Of Covid-19 Ensemble Forecasts: Forward Or Stick" by Jinman Pang

Date of Award

8-1-2023

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College/School/Department

Department of Economics

Dissertation/Thesis Chair

Kajal Lahiri

Committee Members

Pinka Chatterji, Ulrich Hounyo

Subject Categories

Economics

Abstract

This study conducts a postmortem analysis to evaluate the overall forecasting performance of a large number of pandemic transmission models for COVID-19 mortality. For a comprehensive accuracy analysis, we selected thirty-nine prolific forecasting models, which reported continuously national-level weekly incident death forecasts at least fifty times at four weekly horizons between June 8, 2020, and May 28, 2022. We found that the forecasting performance of individual models varies over target weeks and horizons, even after adjusting for time-varying forecast difficulty. Most individual models predicted the turning points with a lag, and very few models excelled at predicting the troughs and peaks of COVID deaths ahead of time. Generally, the models were found to be biased and informationally inefficient, except for COVIDhub-ensemble and OliverWyman-Navigator, which performed well in terms of forecast accuracy, forecast rationality, unbiasedness, and efficiency. We also found that the information updating processes of individual models reflect both sticky and noisy information theories at different horizons and incorporate forward information when making projections. Furthermore, our study confirms that the dynamics of disagreement in forecasts generated by the forecasting models are determined by heterogeneity in the initial prior beliefs implicit in their model formulations and the heterogeneity in how they interpret widely available new public information.

Download

Included in

Economics Commons

Share

COinS