Draw detection
Active tournaments from the ATP and Challenger calendar are loaded and the day's matches are detected with their surface, round and tournament context.
Technical foundation
How it works
Deep Tennis AI isn't "an AI that has opinions about tennis." It's a pipeline that every day gathers player and market data, feeds it to an AI model anchored to the real price, and then puts every pick through a deterministic control layer that decides its quality, its stake and whether it deserves to be published. The AI proposes; the system disposes.
Daily pipeline
From the ATP calendar to your pick, in 5 steps.
The process runs automatically every day before the day's matches. No pick is generated by hand or adjusted after publication.
Performance data
For each player: ranking and points, surface-specific level ratings, dominance ratio (relationship between points won on serve and return), recent match-by-match form, H2H history and stats from their previous meetings.
Real multi-bookmaker odds
8 markets are extracted per match (winner, sets, game and set handicaps, totals, correct score...) comparing several bookmakers and fixing a reference price. Without enough odds, the match isn't analyzed.
Market-anchored model
An AI model with temperature 0 (same input, same output) receives the whole package. Its estimated probability starts from the market price and can only deviate within a limited margin: the AI can't "make up" a giant edge.
Score, stake and publication
Every proposed pick goes through the deterministic layer: quality score Q, market filters, stake assignment and publication decision. This is where the system can say NO even if the AI wants to bet.
Internal score
Q: the quality grade the AI doesn't decide.
The Q score is calculated in the backend with a deterministic formula — the AI model can't inflate it. It combines the value detected, the edge against the price, the quality of the available data and market stability. The exact threshold and its weights are internal calibration.
A premium pick needs two things: to clear the current Q threshold and survive the hard gates — markets that historically and statistically don't pay off are banned from premium even if their Q is high. New rules apply with an effective date: they never rewrite history that's already been published.
Probability vs price
Difference between the estimated probability and the one implied by the odds. Without enough margin against the price, there's no value no matter how good the read looks.
How much we really know
Players with a small sample, no history on the surface, or incomplete data score lower. The system prefers not knowing over pretending it knows.
Flags that subtract
Doubtful injuries, stale odds, contradictory signals between sources, or forced picks trigger penalties that lower the grade or cancel the stake outright.
Stake
The system sets the units, not enthusiasm.
The stake (1 to 5 units) is assigned with a ladder based on pick quality, with per-market caps and risk cutbacks. And it has an exit almost nobody shows you: stake 0.
Ladder by quality
The higher the Q, the more units — but only within markets that have proven they deserve it. The ladder is capped by market and format.
Caps by market
The most volatile markets (game totals, demanding handicaps) have their own stake ceilings even if the pick looks good. No single read concentrates too much risk.
Risk cutbacks
If there are active flags — unconfirmed injury, small sample, sensitive market — the stake drops even if quality is there. The top of the ladder requires a clean record.
Stake 0: the system says no
When the AI's proposal doesn't pass the control layer, the pick stays at 0 units and doesn't count as a bet. This filter systematically discards the statistically weakest picks before they reach anyone.
Formats
BO3 and BO5 aren't analyzed the same way.
A 5-set match changes the math: a strong favorite comes back more often, a -1.5 sets handicap means 2-0 in BO3 but 3-0/3-1 in BO5, and game totals are far more volatile. That's why Grand Slam draws are routed to a format-specific analysis with their own filters, banned markets and stake limits.
Active version
The current generation of the system: probability anchored to the market price, temperature 0, deterministic Q score in the backend, and market gates. Validated against hundreds of matches before deployment.
Validation against historical data
Every change to a threshold, gate or stake is first tested against the real historical record of picks. If an adjustment only wins in theory, it doesn't reach production.
Parallel candidates
New analysis versions run silently alongside the active one and only replace it if they show consistent improvement, not just one good week.
Verification
Every pick audits itself — in public.
Publishing a win is easy; the hard part is leaving the miss in plain sight at the same font size. This is how the loop closes:
Automatic closing
When the match ends, the real result marks each pick as won, lost or void. No manual editing, no picks that disappear.
The market as judge
The movement of the odds is logged up until shortly before the match starts. Systematically beating the closing price is the most honest signal that there's real edge, whether or not the individual pick wins.
Changes with effective date and backup
When a rule changes (threshold, gate, signal), it's applied with an effective date and a backup of the prior state. The public track record includes the good and the bad.
Public, verifiable rules
The winner signals (Winner Elite, Tapado IA, Combi Elite) keep a stable public identity and show their full track record. Internal filters can evolve with an effective date and traceability, without breaking the public read of each signal.
Data
What the model looks at before assessing a match.
The decision combines player, match and market context. We don't publish specific data providers or the full prompts.
Recent form
Recent matches with real scorelines, set trends, accumulated load and competitive continuity.
Surface
Surface-specific level rating, historical adaptation to the tournament and the court's expected speed.
Dominance
Dominance ratio: how much a player dominates their points on serve and return. It tells apart the player who wins narrowly from the one who wins comfortably.
Matchup
Full H2H with a breakdown by surface, style clash and how each player arrived at their previous meetings.
Context
Round, format (BO3/BO5), tournament level, rest between matches and both players' competitive situation.
Market
Multi-bookmaker reference price, the line available in each market, and consistency between the price and the rest of the signals.
Public data to verify it
The Premium record keeps date, market, odds, stake, result and strategy version. It can be checked on the site or downloaded in reusable formats.