Developer Francesco Sacco has released DeepShot, an open-source tool that predicts NBA game outcomes with machine learning, on GitHub under the MIT license. It uses only public data, with a reported accuracy of about 66%.
Open Source · MIT License · Local ML App
DeepShot: Predicting NBA Winners With Machine Learning
An open-source app blends rolling averages, EWMA trends and chess-style Elo ratings to forecast game outcomes — reaching 66.45% accuracy on a three-week hobby build that runs entirely on your own machine.
66.45%
Reported model accuracy (validated on 2024 season)
~25
Games of rolling averages feeding recent-form features
149★
GitHub stars & 20 forks (as of July 2026)
How accurate is it, really?
Reported accuracy of leading approaches — where DeepShot sits on the scale.
~64%
Simple win-rate
66.45%
DeepShot
~65–72%
Professional
Only ~2 points above a naive baseline — and roughly in line with pro-grade Elo models.
From data to prediction — the local pipeline
1 · Collect
Historical team stats via BeautifulSoup / Requests
→
2 · Engineer
Rolling averages, EWMA trends & Elo ratings
→
3 · Train
scikit-learn + XGBoost → deepshot.pkl
→
4 · Visualize
Interactive NiceGUI web UI, runs locally
Praise
Clear, usable NiceGUI interface
Builds on free public data + mainstream libraries
Great educational material for ML learners
Criticism
66% is close to a ~64% baseline
Needs comparison against Vegas lines
No hosted version; Docker support still wanted
"A roughly three-week project that cannot yield a consistent betting edge — though it sometimes beats the odds."
— The author, noting it compares reasonably against professional bettors' ~68% accuracy. Feedback is welcomed through GitHub Issues.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…