EddieTGH / kalshi-weather-predictor
What it does
Custom temporal transformer ensemble with XGBoost. XGBoost achieves 1.55°F MAE; the custom temporal transformer achieves 1.0°F MAE. Optuna Bayesian hyperparameter search. Converts continuous forecasts to integer probability mass functions across bracket boundaries. Fractional Kelly with a 2% bankroll cap. Full frontend dashboard with auth.
What we learned from it
the accuracy ceiling on day-of weather forecasts is around 1°F MAE for the best ML models. That's tight enough to often pick the right 2°F bracket. If you're going to compete with this kind of bot, you need either a better model (hard) or different markets (easier).
Find the author
- Repository: github.com/EddieTGH/kalshi-weather-predictor
- GitHub profile: @EddieTGH
- Website / blog: edmond-niu.dev/
- Author bio: I am a senior at Duke studying CS. Passionate in utilizing AI/ML and software engineering to facilitate human productivity and automate inefficient processes.
Don't want to maintain your own bot?
Bot for Kalshi is the no-code platform for automated prediction-market trading — visual builder, signals, encrypted credentials, kill switch. The infrastructure these open-source projects build, ready to use.