For my UC Berkeley MIDS DATASCI 207 final project, my team and I studied next-day rainfall prediction in Australia using the WeatherAUS dataset. We framed the task as predicting RainTomorrow from daily weather variables, and built the pipeline with a strict chronological split and expanding-window cross-validation to avoid temporal leakage.
We compared Logistic Regression, Random Forest, RNN, and PatchTST baselines, then tested latent-space generative approaches (VAE and latent WGAN) to better capture rainy-day structure in a highly imbalanced setting. Our manifold-aware approach improved rain-day forecasting compared to purely discriminative models and gave better ROC-AUC with a more balanced precision-recall tradeoff for the rainy class. The project repo is here: MIDS DS207 Final Project (AARVE).