How TennisPicksHub predictions work

ATP men's singles only, for now. The homepage lists scheduled matches for roughly the next seven days, including qualifying when it appears on the tour calendar.

Because we do not yet have the rating history to do it properly. The ATP model rests on about a decade of completed match results per player and per surface; that history has not been built for the WTA tour here. Without it, the model would output something close to 50/50 on nearly every match, which is worse than showing nothing — it would look like a prediction while carrying no information. WTA coverage will appear only once the rating history behind it is real.

Two Elo ratings per player — one overall, one for the match surface — are blended, with thin surface samples pulled back toward the overall rating. That produces a first probability, which is blended with one from current ATP ranking, then adjusted slightly for rest and recent workload. A calibration step fitted on past matches converts that into the published figure.

Court speed and bounce change which skills pay off. Clay slows the ball and rewards defence, spin and stamina; grass keeps the ball low and fast and rewards serving and short points; hard courts sit between them. The same two players can have genuinely different chances against each other depending on where they meet, and it is common for a player to be a hundred or more rating points stronger on one surface than another. A single all-surface rating would quietly misprice a large part of the tour.

That player has fewer than five completed matches in our history on that surface, so there is no meaningful surface-specific read on them. The prediction falls back on their overall rating and ranking. Treat it as a rough estimate, and note that market comparisons are withheld entirely when either player is unrated.

Between five and nineteen matches on that surface — enough to be informative, not enough to be trusted on its own. The surface rating is used but pulled substantially toward the player's overall rating, and the marker stays on the card so the thin sample is visible.

We could not confirm a live ATP ranking for that player in the latest published list — common for qualifiers, players returning from injury, and those outside the main computer ranking. The ranking part of the model is either skipped or uses a stale figure, so the prediction leans harder on Elo.

Projected games is the model's expected total games in the match; straight sets is the chance it ends without either player dropping a set. Both follow from the win probability and match format, so they are averages across many similar matches, not a scoreline forecast.

When bookmaker prices exist for a match, we strip the bookmaker's built-in margin from both sides so the two implied probabilities sum to 100%, then show how far the model's number sits from that adjusted market number, in percentage points. It is context — a note that the model and the market read the match differently. It is not a claim that the model is right, and it is not a claim that acting on the difference is profitable. Disagreement most often means one side has information the other lacks, and that side is frequently the market.

Odds coverage from our data source is per tournament, not per match, and only the larger events are carried. Smaller tour stops, most qualifying draws, and events far ahead of their start date have no prices available at all, so those matches are prediction-only. We also withhold the comparison when either player is unrated on the surface, because a market gap computed from a rating we do not trust is noise.

Usually because an opponent is not yet decided. Later-round fixtures appear on the schedule as placeholders while the earlier match is still being played, and we will not attach a probability to a player who does not exist yet. A prediction is generated once both names are confirmed. A match can also stay unpredicted if we could not confidently match a player name to our database, or if the tournament's surface is unknown.

No. Nothing is sold, no picks are supplied, no wagers are taken, and no claim is made about profit. It is a free statistics site publishing a model's win probabilities together with the quality of the data behind them.

These percentages represent the plain gap between the model's prediction and the margin-free market price. For example, if the model says 58% and the market says 49%, that is a 9% gap (previously referred to as percentage points or 'pp'). It means the model and the bookmaker price the match differently; it is not a tip and not a claim that either side is right.

Because ranking and Elo answer different questions, and the app uses both. ATP ranking is points earned over the last 52 weeks, so it collapses when a player misses months through injury. Elo remembers how they actually played, and against whom, so a returning player can sit far higher on surface Elo than on ranking — Jack Draper, ranked outside the top 100 while injured, still rates inside the top ten on hard. Blending the two lets the model respect current form without pretending a strong player became weak while sitting out.

The list shows players with a completed match in the last twelve months. Ratings are frozen at a player's final match, so Roger Federer or Juan Martín del Potro would otherwise sit among active players with figures that no longer describe anyone playing this week. Tick 'include inactive players' to see them; their ratings are unchanged, just marked inactive.