Model Complexity and Accuracy: A COVID-19 Case Study

When creating mathematical models for forecasting and decision making, there is a tendency to include more complexity than necessary, in the belief that higher-fidelity models are more accurate than simpler ones. In this paper, we analyze the performance of models that submitted COVID-19 forecasts to the U.S. Centers for Disease Control and Prevention and evaluate them against a simple two-equation model that is specified using simple linear regression. We find that our simple model was comparable in accuracy to highly publicized models and had among the best-calibrated forecasts. This result may be surprising given the complexity of many COVID-19 models and their support by large forecasting teams. However, our result is consistent with the body of research that suggests that simple models perform very well in a variety of settings.