Overview
COV-(AI)D forecasts COVID-19 cases in U.S. states using historical case data plus human mobility trends from Apple Maps. The idea was that people moving around more meant more transmission risk.
How It Works
I noticed that case spikes seemed to follow increases in human movement. So instead of just training on past COVID numbers like most models, I fed the LSTM both case counts (from Johns Hopkins) and daily mobility data (from Apple Maps routing requests).
Used a 3-layer LSTM with 30-day sequences. The model hit an MAE of around 600 cases on 30-day forecasts, which beat traditional time-series methods like ARIMA.
Technical Details
- Python + PyTorch
- 3-layer LSTM for multivariate time-series
- Trained on COVID cases + Apple Maps mobility data