NPR’s Morning Edition featured a small segment on the development of an ensemble model for COVID-19 by a biostatistician at UMASS Amherst, Nicholas Reich. Basically, combining multiple models intelligently produces a a better model than the individual components.
This finding is helpful if not surprising. Ensemble techniques have been used in multiple disciplines, especially weather forecasting. In that application there are two types of uncertainty to be “integrated out”: Uncertainty in the actual initial state and uncertainty from the remaining difference between the model output and the real world. In some applications with which I’ve worked in the area of geopolitical forecasting the ensembles feature weighted combinations of alternative models, similar to the ensemble described in the NPR piece. (According to that piece the COVID-19 ensemble from Reich is an unweighted average of the several model forecasts with accuracy weighting to come in a future version of the ensemble.) The uncertainty surrounding the initial state doesn’t really come into play in this domain in the same way that it does with weather forecasting. In my experience the ensemble isn’t always uniformly the best model but it’s never a bad model. Since we don’t know a priori which of the models will be the best using an ensemble is a nice hedging strategy.