A real prediction, published before the match
Beijing · Thursday, 8 October 2026
Predicted score: 2-0
N. Bartunkova wins 2-0
MissedAlgorithm
Tennis prediction algorithm: how our AI predicts every match
Every morning, before the first point, Pronocast publishes a win probability for every ATP, WTA and Challenger match, computed by Pronocast Engine, our prediction algorithm. This page explains what it looks at, what it learned from, and how to check its accuracy yourself. No black box: every prediction is locked before the match and published alongside the result.
- training matches
- 100,000+
- unseen test matches
- 37,104
- accuracy in testing
- 70%
What the algorithm analyses
Six families of information, all known before the match. Nothing that happens during or after the match enters the calculation.
Swipe to see all six factors
Why 100,000+ training matches
The model learned from every ATP, WTA and Challenger singles match played since 2020, more than 100,000 matches, and player state (Elo, form, fatigue) is computed on close to 300,000 matches, ITF included. This volume is not a marketing figure: it is what lets players climbing up from small tournaments reach the main tour with a rating that is already reliable, and what ensures rare situations (return from injury, surface switch, unknown player) have been seen often enough.
- Chronological training: the model only learns from matches played before the ones it predicts, never the other way round.
- A single implementation of player state, shared between training and the day's predictions.
- Refitted every morning on the previous day's results, recipe re-validated every month.
Pronocast Engine: the model in detail
Pronocast Engine is the name of the algorithm behind every prediction on the site. Here is what it is, without the internal settings.
- Step 1
Model family
Gradient-boosted decision trees (scikit-learn's HistGradientBoosting), combined with an Elo rating computed by Pronocast. Probabilities are recalibrated with isotonic regression, then blended with the probability implied by bookmaker odds when it exists.
- Step 2
Training data
Every ATP, WTA and Challenger singles match played since 2020, more than 100,000 matches. Player state (Elo, form, fatigue, head-to-head) is computed on nearly 300,000 matches, ITF included. Doubles have their own model, built on an individual Elo.
- Step 3
Retraining
Pronocast Engine is refitted every morning on the history updated with the previous day's results. Every month its recipe (features and settings) is re-evaluated, and replaced only if the new one does strictly better on the same test set.
Accuracy measured on more than 37,000 matches
Tested on 37,104 matches from 2025 and 2026 that it never saw during training, the model picked the right winner in nearly 70% of cases, on par with the betting market. Misses count as much as hits, and every prediction published since has its own page, locked before the match.
All test matches70%
Matches announced at 70% confidence or more83.5%
Why not 90%? Because half of all tennis matches are so close that no model can call them reliably. On matches where the algorithm announces 70% confidence or more, its actual accuracy exceeds 80%. The dashboard's Recommendations page puts those first.
See every verified matchFree vs Pro: what changes
During the beta, the whole algorithm and dashboard are free, with no credit card. The Pro plan will follow with additional tools; members who signed up during the beta will be notified before any change.
Beta, free
- ATP, WTA and Challenger predictions
- Predicted set score
- Track record and measured accuracy
- Updates itself
Frequently asked questions
How does a tennis prediction algorithm work?
It turns what is known before the match (both players' Elo ratings, surface, form, fatigue, head-to-head, odds) into a win probability. The model learned this mapping from more than 100,000 past matches, then applies it to the day's matches.
Is Pronocast's algorithm free?
Yes, during the beta. A free account gives access to every ATP, WTA and Challenger prediction, the predicted set score, the track record and the recommendations. No credit card is required.
How accurate is the algorithm?
About 70% on more than 37,000 test matches from 2025 and 2026, which the model never saw during training. On matches announced at 70% confidence or more, it exceeds 80%.
Can predictions be checked after the fact?
Yes, that is the whole point. Each prediction is computed in the morning, locked, then published with the result on the match page: anyone can check it after the fact.
Does the algorithm beat the bookmakers?
No, and nobody does so consistently: consensus odds are the best public predictor known. The model matches them, also covers matches without odds (Challenger included) and provides a calibrated probability for each.
How is this different from the Methodology page?
This page explains how it works in plain terms. The methodology details the metrics (log-loss, Brier score), chronological validation and calibration, for anyone who wants to audit the approach in depth.
Going further
Try the algorithm on today's matches
Free during the beta, no credit card.