Built for a hackathon during the early months of COVID-19 to identify forecasting approaches that could inform real production use cases.
The project tested multiple methods—including SARIMAX, VAR, linear regression, and nonlinear curve fitting—and explored highly debated public-policy questions, such as whether lockdown stringency could help predict case trends.
The project finished 2nd place against a field that included PhD-level researchers. I walked away with some swag, bragging rights, and—slightly more importantly—the opportunity to contribute to production forecasting strategy.
Used a multi-model forecasting approach rather than assuming one method would work equally well everywhere. Each model was backtested by region, then selected and weighted based on how accurately it performed on that region’s historical data.
Methods included:
-
ARIMA with automated hyperparameter testing Tested multiple ARIMA configurations to find the best fit for each region’s historical case patterns.
-
Logistic curve fitting Fit an S-shaped growth curve to estimate how quickly cumulative cases could rise and where growth might begin to level off.
-
ARIMA with lockdown stringency as an external factor Added lagged government-restriction data to test whether changes in lockdown policy could improve future case predictions.
-
Vector Autoregression with stringency data Modeled case counts and lockdown stringency together, allowing the historical behavior of each to inform the forecast.
The final forecast combined the strongest-performing models for each region, with greater weight given to the methods that produced the best backtested results.
