misc project

tennis.computer

Turning historical tennis data into calibrated win probabilities, then betting them against the market.

Stack: Elo → gradient boosting → stacked ensemble
Validation: Walk-forward vs closing line
Status: Live

What it does

An Elo rating anchors the model, a gradient-boosted layer learns what Elo misses, and the two stack into a single calibrated probability for any ATP match on any surface.

Those match probabilities then feed a Monte Carlo simulation of the tournament draw, which turns them into outright win percentages for a whole event.

Why the closing line matters

A prediction model is easy to fool yourself with. The betting market’s closing line is the hardest public benchmark available, because it has absorbed every piece of information right up to the first serve.

So the model is validated walk-forward against it, and paper-traded before any real money. Beating the closing line is the bar for going live.