There are two general approaches to predictive modeling in economics: structural and reduced form. The structural approach estimates the parameters of a causal model in order to estimate the effectiveness of interventions and make long-term forecasts. The reduced form approach uses extrapolation and smoothing to develop short run "black box" forecasts, sometimes called "nowcasts".
Here we focus exclusively on the reduced form approach and use some standard machine learning models to nowcast covid-19 cases in the US. We pay particular attention to survey data and the role of "proxy questions," which are questions of the form "are there any people in your community with the following symptoms ...?" Proxy questions turn out to be highly useful for short-term predictions.
We also examine county level data using model selection techniques and find suggestive examples of covid-19 incidence and socio-economic covariates.
This keynote talk will be 30 minutes long, followed by 10 minutes of commentary by Michal Gal (University of Haifa) on "Access Barriers to Data", and 20 minutes of Q&A.