Forecasting with Almost No Data

Panama's official census data gave me exactly 40 usable data points for unemployment. My Extended Essay asked which machine learning models, interpretable versus non-interpretable, forecast better when data is that scarce.

I built and compared two models, reaching about 80 percent accuracy on test data. The honest finding: everything overfits at that scale, no matter the regularization. A forecaster the government could rely on was out of reach with the data available.

The failure opened a better question that still keeps me curious: how do you build accurate models without massive data? Nobody has a great answer yet.