The research behind the model

The canonical in-play forecasting paper — win probability updated point by point, not just before the match.

How to turn published serve statistics into the point-win probabilities a Markov model needs.

Compares two players through opponents they have both faced, rather than through raw averages.

Serve and return skill as a Gaussian random walk, varying by surface. Reports 68.8% accuracy against 66.3% for earlier point-based models.

A readable, self-contained walk through point-level prediction if you want the mechanics before the algebra.

Eleven published models tested head to head on 2,395 ATP matches. The single most useful benchmark paper in the field.

Weights each result by how emphatic it was, rather than treating every win as identical.

Four ways to fold margin of victory into Elo. Only the joint additive form stayed unbiased in simulation.

Glicko adds a reliability term to Elo — useful for players with thin recent histories, which is most of the ITF draw.

Treats the tour as a network of results and ranks players by their position in it.

Read this one before believing any ML claim: the models beat ranking-only baselines but could not beat odds-implied forecasts.

Recent, and specific about Grand Slams, where best-of-five changes the variance structure.

Feeds statistical model output into a learner rather than choosing between the two.

Short and practical, and unusually candid about how thin the betting margin turned out to be.

In-play prediction with neural networks — the closest academic analogue to the live re-pricing here.

Network features as inputs to a learner rather than as a ranking in their own right.

Tests whether momentum is measurable rather than assuming it. Relevant to any in-play signal.

In-play odds track the match closely — the benchmark any live model is competing against.

Finds longshot bias in tennis, but not enough of it to make backing favourites profitable.

Where sportsbook prices drift from fair value, using tennis as the setting.

A survey worth reading for the recurring methodological errors it catalogues.

Combining model output with human judgement, and when that helps rather than hurts.