Online Program

Aggregating Statistical Models and Human Judgement (310008)

Keywords: COVID-19, risk perception, public health, forecasting

Forecasts of COVID-19 based on computational models are restricted to train on reportable, objective data. However, human judgment forecasting uses a combination of objective and subjective data to produce probabilistic predictions. Our work aims to build a metaforecast, a combination of computational forecasts with probabilistic predictions from subject matter experts and trained forecasters. We asked forecasters for predictions of the number of incident infections, hospitalizations, deaths, the spread of the B.1.1.7 variant, and the cumulative number of first dose/fully vaccinated individuals. The average rel. difference in accuracy between the consensus and an ensemble of computational models (Fhub) was 2.78 (p=0.18) and between the metaforecast and the Fhub was 1.46 (p=0.22). When ranked from least (rank of 0) to most (rank of 1) accurate, the mean rank for the consensus was 0.55, for the metaforecast was 0.42, and for the Fhub was 0.30. Adding human judgement to computational models can generate fast and potentially more accurate probabilistic predictions to support public health risk communications that realign perceptions of COVID-19 risk with objective data.