Show HN: DeepShot – NBA game predictor with 70% accuracy using ML and stats (github.com)

DeepShot is an open-source NBA game predictor that combines historical box-score data from Basketball Reference with rolling, momentum-aware statistics to forecast matchups — the author reports roughly 70% accuracy. The project wraps a reproducible ML pipeline (trainable via a provided Jupyter notebook) into a sleek, cross-platform NiceGUI web app so users can run the model locally, inspect predictions in real time, and visually compare the key stat differences driving each forecast. Technically, DeepShot emphasizes recent form by computing Exponentially Weighted Moving Averages (EWMA) for team stats, feeding these weighted features into a machine-learning model (model type saved as deepshot.pkl). The interface highlights which metrics most distinguish teams, making the system useful for analysts, hobbyists, and anyone who wants a transparent, data-driven lens on matchups. The code is MIT-licensed and built from free public data; the repo includes installation steps and instructions to train and launch the app. Note the “70% accuracy” claim depends on dataset, splits, and evaluation choices, so users should validate performance for their use case — the project encourages feedback (emailware) and contributions.