Methodology

1. The data

Match results, draws, rankings and pre-match odds come from a commercial tennis data feed covering the ATP and WTA tours and the Challenger level. Closing prices for closing line valueare taken only from sharp, high-limit markets, because a soft bookmaker's closing price says little about whether a pick beat the market. A price only counts as a closing price if it was recorded within two hours of the start; anything earlier, or any pick without a recorded sharp price, is shown as "not measured" rather than given a made-up number.

2. Ratings: surface Elo

Every player has an overall Elo rating and one per surface, rebuilt every week from real results (ITF events excluded as too noisy). New ratings move quickly at first and settle after about 20 matches. A surface rating only starts to count after five matches on that surface and reaches its full weight after thirty, so a player with two grass matches is not judged on grass alone. Indoor and outdoor hard courts are treated as close relatives. Players with fewer than 15 tracked matches fall back to an estimate from their ranking, and Grand Slam matches are converted to best of five. Details: why tennis ratings need to be surface-specific.

3. From rating to pick

1. A short list of matches is chosen each run, then enriched with form, head-to-head on the surface, serve and return statistics, rest days and recent workload.
2. A fair win probability is estimated blind, without seeing any odds, then pulled back toward 50 % to correct the usual overconfidence and blended with the Elo estimate.
3. A player who retired from their last match (within 60 days) is marked down by 4 percentage points. We measured this on almost 28,000 matches: such players win less often than their ranking suggests, and the market already prices it in.
4. The estimate is compared with the bookmaker price after removing the margin (no-vig). A pick needs an edge of at least 3 percentage points and an expected value of at least +2 %. Markets with a margin above 8 % are skipped.
5. Stakes follow fixed confidence tiers; the Kelly fraction is used only as a reference and is capped.

4. Tracking and settlement

Picks are posted in the Discord before the match starts; Discord's own message timestamp proves when. Each pick is settled from the official result, including retirements and walkovers under fixed rules, and every result goes into the public track record, win or lose. When a rule changes, all past picks are recalculated under the new rule and any differences are corrected.

5. Model board vs. picks

The model board shows the raw Elo view for every match next to the market. That is not a pick: a gap between model and market is only a starting point, and most gaps disappear once the full analysis and filters above are applied. Since late October 2026 the board's market price is refreshed every hour, so for most matches it is now taken within about two hours of the start; before that it was often several hours to a day old, so older and newer board entries are not measured against quite the same market. How accurate the board is gets published there once 200 matches are settled; below that the numbers are mostly noise.

Known limitations

• Bookmaker odds are efficient. Beating them consistently is genuinely hard, and we don't claim otherwise.
• With a single odds source, "value" is an estimate against one book's line, not a guaranteed cross-market edge.
• Elo knows results, not injuries, schedules or motivation. The checks in step 3 catch some of that, not all.
• Small samples lie. A few weeks of results, good or bad, say little about whether the model has a real edge.
• This is a research aid, not financial advice, and past results don't guarantee future ones.