Tippmix
A self-correcting sports-betting analysis engine — rebuilds team form and Elo from match history, prices markets against the bookmaker's own odds, then grades its own predictions and folds the error back in.
Tippmix is an automated multi-sport analysis system for the Hungarian betting platform of the same name. It pulls upcoming fixtures, reconstructs each team's recent form and Elo rating from historical results, projects a scoreline, and scores every market it can price — producing confidence-rated recommendations rather than a flat yes or no.
The part I find most interesting is the feedback loop. After results settle, the system grades its own run and stores a signed calibration offset per market, so a market it has been persistently overconfident in gets pulled back on the next pass. The first version of this learned purely from wins and losses, which turned out to be dominated by variance — a correct bet that loses is still a correct bet. The current work replaces it with a decision-quality measure: how far the model's probability sat from the closing line, and whether the price moved in its favour after the bet was placed. That version runs in shadow mode until an A/B backtest over the archived run history proves it beats what is live.
The core is deliberately Python standard library only — no numpy, no ML frameworks. Every formula is a few lines you can read and argue with, which matters when the whole point is knowing why a number came out the way it did. A separate Deno and TypeScript layer provides the web front end.