The model went 11-2 at UFC Shanghai on Saturday, its strongest card of the season. Three of those 13 picks were produced from prices that could not have existed in a real market.
We found that ourselves, before the fights, and we are printing it because the alternative is letting an 84.6% card stand on numbers we already know were wrong.
The odds file we build from carried 15 rows for a 13-bout card. Two bouts appeared twice at prices that contradicted each other, pulled from a scrape target that had drifted onto the wrong event. The damage showed up in nine of the 13 bouts: priced as probabilities, the two sides of each fight added up to less than a whole. A real market always adds up to more than a whole, because that gap is how the book makes its money. Ours added up to less, which is not a tight market or a generous one. It is not a market.
The model corrects for the book's margin by dividing by that total. When the total is below a whole, the correction runs backwards and inflates the error instead of removing it.
On three bouts it inflated far enough to move the market from backing one fighter to backing the other.
Hector Santiago vs. Lawrence Lui. We picked Lui. Santiago won by TKO at 0:53 of the second round.
Yan Xiaonan vs. Denise Gomes. We picked Gomes. Gomes won by TKO at 4:49 of the first.
Cam Nelson vs. Ding Meng. We picked Nelson. Nelson won a unanimous decision.
Two of the three won, and that is the part worth sitting with. A corrupt input that produces a correct pick is worth less than a clean input that produces a wrong one, because only one of the two will do it again. Those two hits are not evidence the model works. They are evidence it got lucky in a direction nobody can repeat.
The Nelson bout is the clearest case. A valid quote timestamped 7:12 a.m. GMT had Ding Meng as the favourite at -138. We served Nelson as the favourite instead, off a pair that summed to roughly half a market. Nelson won. So we were right about the fight, we disagreed with the real price, and the thing that put us on the right side was a number no book ever offered.
None of this touches the main event, where the prices were sound and the miss was clean. Song Yadong stopped Umar Nurmagomedov by TKO at 1:48 of the second round. We had Nurmagomedov at 76%, at our highest tier, and we called a decision. Wrong fighter, wrong method, nothing to blame but the read.
That was the only conviction pick to lose. The tier finished 3-1: Andre Lima by submission in the third, Bilal Hasan by TKO in the second, and Rei Tsuruya by submission in the first. All four conviction bouts sat on prices that held together, so the tier record for this card is unaffected by the fault.
Method was the weak half again, correct in four of 13. Ten of the 13 fights finished, eight of them by knockout or TKO, and the model called for a decision 10 times. It read the winners well and the endings badly, which is close to the opposite of last month.
For the season, and this is history rather than a forecast: all picks including toss-ups now stand at 247-104, or 70.37% on a sample of 351, across 29 graded cards as of Aug. 29. We are not publishing a conviction-tier season figure alongside it, because that cohort is under an internal hold and cannot currently be reproduced.
The odds fault is fixed at source. Pairs that add up to less than a whole are now refused rather than corrected, and where a valid earlier quote exists it is used with its timestamp attached. Nothing on this card was regraded. The card stands at 11-2, with three of the wins and losses inside it produced by an input we cannot stand behind.