UFC Sacramento predictions: the model went 8 for 13
UFC Sacramento predictions scored: our model went 8 for 13, the closing market 9 for 13, and four underdogs won after their prices drifted all week.
Two days before UFC Sacramento I published our prediction model's report card, and it was not a good one. Across five weeks and 60 UFC fights, the model and the closing betting line had picked opposite corners exactly ten times, and the model had lost all ten. I wrote it up under the headline that our model was 0 for 10 against the market, because burying that would have been worse than printing it.
Saturday's card gave it one more chance to disagree with the market. One fight out of 13.
Kennedy Nzechukwu against Shamil Gaziev. Our model gave Nzechukwu 58 percent. The books had Gaziev at -123, an implied 53 percent the other way. Gaziev knocked Nzechukwu out in 80 seconds.
The model is 0 for 11.
That is the headline number and I am not going to dress it up. Here is the rest of the night, which is more interesting than the streak.
The model went 8 for 13 on Saturday. The closing market went 9 for 13. On the 12 fights where they agreed, they were both right eight times and both wrong four times, which is the honest description of what a model that mirrors the market buys you: nothing, at a slightly worse price.
Four underdogs won, and both of us missed all four
Photo: ReinierdeRidder / CC BY-SA 4.0
The interesting part is not the one-fight gap in the win column. It is that four fighters won as betting underdogs on Saturday, and the model had all four of them as underdogs too.
| Fight | Underdog winner | Closing price | Model's number | Market's number |
|---|---|---|---|---|
| Main event | Gregory Rodrigues | +182 | 35% | 34% |
| Main card | MarQuel Mederos | +305 | 32% | 24% |
| Prelims | Jamall Emmers | +305 | 42% | 24% |
| Prelims | Stan Dorsainvil | +155 | 23% | 38% |
On two of the four the model was meaningfully closer to the truth than the market was. It gave Emmers a 42 percent chance where the books gave him 24, and Mederos 32 where the books gave him 24. On Dorsainvil it was worse than the market by a wide margin, having liked Gauge Young at 77 percent in a fight Young lost on all three cards to a debutant.
Scored properly rather than by win-loss, the two are close enough to be indistinguishable on a sample this size. Brier score, which punishes confident wrong answers hardest, came out at 0.239 for the model and 0.223 for the closing line. The market wins the night on the number, which is what usually happens and what should happen: the closing line has absorbed injury news, camp reports, and the opinions of everyone with money at stake, and our model has a frozen feature set and no idea who limped out of the gym on Wednesday.
Where the money went is its own story
The other thing worth recording from Saturday is what the lines did between the day they opened and the moment they closed, because our odds history keeps the open and the close for every book we track.
All four of those winning underdogs got LONGER over fight week. Rodrigues opened at +132 and closed at +182. Mederos opened +210 and closed +305. Emmers opened +279 and closed +305. Dorsainvil opened +120 and closed +155. Every dollar of that movement was money walking away from the four men who ended up winning.
And where the money moved hardest toward a favorite, it was right. Reinier de Ridder opened at -350 and closed -390, then finished Roman Dolidze in the first round. Carli Judice opened -480 and closed -614, then finished Jeisla Chaves in 99 seconds. Shanelle Dyer opened -710 and closed -841, then stopped Elise Reed in the third.
The market was not confused on Saturday. It was confident, and it was confident in the wrong direction on four fights and the right direction on nine, which is roughly what a 69 percent hit rate looks like when you are pricing an entire Fight Night card.
What I actually take from it
A model that agrees with the closing line on 12 of 13 fights is not adding information. I said that before the card rather than after it, which is the only reason I get to say it again now. Agreement is not accuracy: it went 8 for 13 while agreeing, and the one time it broke ranks it lost inside 80 seconds.
What it did do, in two places, was refuse to price a big favorite as big as the market did. Emmers at 42 percent when the world said 24, and he knocked his man out in the first round. That is the shape of a useful disagreement, and nobody could have acted on it, because the model still had him losing.
Our predictions for every upcoming card sit on the event pages, corner-stamped so a late opponent change voids the number rather than silently reassigning it to a different fighter. The next batch goes up inside 48 hours of the next UFC card. The disagreement counter reads 0 for 11. I will keep printing it either way.