In the preview we published for UFC 330 we wrote that "across a validated cohort of 80 flagged LOCK picks, the LOCK label has hit 91.2%."
Six days later our four LOCK picks on that card went 2 of 4.
We are not walking that sentence back, because it was true. The 91.2% was a real measurement on a real cohort of 80 picks, and we labelled it in the preview as a past record and not a forecast. That labelling was correct. It also does not get us off the hook for how the line reads this week.
The statistical context, offered as context and nothing else: four picks is a tiny sample. Going 2 of 4 does not disprove 91.2%, in the same way that a single clean card would not have proved the model was flawless. A four-fight result carries almost no information about a rate measured over 80 picks.
That is true, and it is also the first argument any model owner reaches for after a bad weekend. So the uncomfortable part goes next to it rather than after it.
Our two biggest LOCK misses were both submissions.
Mansur Abdul-Malik was our highest-confidence pick on the entire card at 80.4%, and the model projected him to win by KO/TKO. He was submitted by Dustin Stoltzfus in the second round.
Myktybek Orolbai, at 76.6%, was submitted by Jeremiah Wells in the third.
We do not have an explanation for that. Two submissions is two fights, and we are not going to build a theory about submission blindness out of a sample that size. If it is a real weakness it will show up again, and we will say so. What we can honestly report today is that it happened.
The LOCKs that held ended inside the distance the way the model expected. Donte Johnson, at 77.2%, won by KO/TKO in the first round. Esteban Ribovics, at 73.5%, won by KO/TKO in the second. Our HIGH-flagged loss was Joel Alvarez at 76.8%, beaten by Chidi Njokuani on the cards.
Flagged picks finished 3 of 6: LOCKs 2 of 4, HIGH 1 of 2.
The second point is really the same point.
The card before this one, UFC Vegas 120 on Aug. 8, we picked all 12 winners. In the autopsy for that card we wrote that it was "one card and 12 fights," and we said nothing at all about how good the model is.
Six days later we went 6 of 11.
That is why we refused to headline the perfect card. A 12-of-12 card and a 6-of-11 card came out of the same model six days apart, and neither number describes it. Had we run a victory lap on Aug. 8, the Aug. 15 card would have made a liar of us inside a week.
We are not publishing a season-to-date flagged figure next to this. Our internal number for that cannot currently be reproduced, and it stays under review until it can be. Card-level results are what we can stand behind, so card-level results are what you get.
The 91.2% stays on the record where we published it. We are not recomputing it here and we are not presenting it as current. It measured 80 picks that were settled before this card started. Last Saturday added four more.