Open data: every settled bet our AI has placed

The complete settled record behind our published figures — wins and losses, with the entry price we actually took. Downloadable, and computed from the database on every page load rather than written down once.

It is published so it can be checked. Every number on this page is the sum of rows in the file below; if your arithmetic disagrees with ours, ours is wrong and we want to hear about it.

The record

Settled bets
800
Matches
600
Competitions
20
Return on stake
+14.60%
Won
257
Lost
215
Half / push
328
Total staked
132,904
Net profit/loss
+19,423

800 settled bets, 2026-01-28 to 2026-07-22. A seven-day delay applies, so the most recent week is not included here or in the download.

Why win rate is the wrong question

Across this record the models won 257 bets and lost 215, and returned +14.60% on stake. That is not a contradiction. A bet has two numbers — how often it lands, and the price paid for it. A service winning 70% of its bets at odds of 1.20 loses money; one winning 40% at 3.00 makes it. Win rate is the easiest metric to inflate, which is why it is the one most often published alone.

By competition

Every competition is listed, including the losing ones. 14 of 20 fall below 50 settled bets and are flagged: at that size a single quarter-line settling the other way moves the return by whole percentage points, so those rows describe variance, not skill. They are shown because omitting them would turn this table into a selection of whichever competition got lucky.

CompetitionBetsMatchesW–LReturn
Brazilian Serie B88581824-0.90%
La Liga83603120+27.90%
Premier League82563621+31.40%
World Cup74691225-13.00%
Serie A64482319+11.00%
Bundesliga52332115+29.00%
J1 Leaguesample too small47351410-16.70%
Pro Leaguesample too small45321711+35.90%
Eredivisiesample too small42271213+4.00%
Brazilian Serie Asample too small38341211+23.20%
Ligue 1sample too small37301413+20.00%
Super Leaguesample too small3331108+5.20%
Primeira Ligasample too small271883+32.60%
K League 1sample too small242487+3.20%
UEFA Champions Leaguesample too small231476-5.40%
Indian Super Leaguesample too small191251+57.70%
Süper Ligsample too small161476+7.80%
FA Cupsample too small3220+79.40%
UEFA Europa Conference Leaguesample too small2201-31.90%
UEFA Europa Leaguesample too small1101-100.00%

Take the data

Row-level, one line per settled bet, with the minute and scoreline at entry, the odds recorded before the outcome was known, and the settled profit or loss.

How the numbers are produced

1. Attack and defence strength, weighted toward recent matches

A Dixon-Coles model estimates each side's scoring and conceding rates, adjusted for the opponent, with older matches carrying less weight. It also corrects the low-score dependency that a plain Poisson model gets wrong — 0-0, 1-0 and 1-1 are not independent outcomes in practice.

2. Expected goals rather than the scoreline alone

A side can win 1-0 having been outplayed. Training on results alone teaches the model the wrong lesson, so shot quality enters the estimate: xG measures what a team actually created rather than what fell for it that day.

3. Monte Carlo simulation

The fixture is simulated thousands of times from those rates. The output is a distribution rather than a single scoreline: the probability of each exact score, each 1X2 outcome, each goals line and each Asian handicap.

4. Comparison against the market, and the filter

The model probability is compared with the probability implied by the offered price. A bet is only recorded when the gap between them is wide enough. That is why most fixtures produce no bet at all — in most matches the market is already right, and betting into it just pays the margin.

Limitations — read before citing

Stated here so that anyone quoting this record is not caught out by them later.

  • The sample is small. Several hundred bets over roughly six months does not separate genuine edge from a good run with any confidence, and no single competition here comes close to a sample that would.
  • This is a trading record, not a backtest. These are bets that were recorded at prices that existed at the time. It is not a simulated result over historical data and cannot be reproduced by re-running the model on old fixtures.
  • Figures are in model stake units, not currency. Real-money return is lower, because the price actually executed is rarely the price displayed.
  • A seven-day publication delay applies, so the most recent week never appears.
  • The models changed over the period. Earlier rows were produced by earlier versions, so the record is not a single system measured end to end.

How to cite

If you use this in a story or a paper, this is the reference we would ask for.

OddsFlow AI, "Settled Predictions Record" (2026-01-28–2026-07-22). Retrieved from oddsflow.ai/open-data.

The live record

This page is the archive. The running record, updated after each match, is here.

This is data analysis published for information and research. It is not betting advice, and no bet is placed on this platform. Past results do not guarantee future results. 18+.