UFC Shanghai predictions scored: our model went 8 for 13
We said our model faded the UFC Shanghai favorites. Scored against the results it went 8-for-13, level with the market, and badly wrong on Umar Nurmagomedov.
Two days before UFC Shanghai I published a piece here headlined "our model fades the favorites." It flagged two fights where our AI model picked a different winner than the betting market, and noted that on eight of the thirteen bouts the model rated the favorite lower than the odds board did.
The card is over. Here is the scorecard on that, including the part where the model was badly, expensively wrong about the main event.
The headline number
Across the thirteen bouts on the card where our model had a corner-matched prediction, it picked eight winners out of thirteen, or 61.5 percent. The closing betting favorite went 8-for-13 as well. On raw hit rate, the model and the market finished the night level.
That is the boring version. The interesting version is where they actually disagreed, which was on exactly two fights out of thirteen.
| Fight | Our model | Market (closing) | Winner | Model |
|---|---|---|---|---|
| Umar Nurmagomedov vs Song Yadong | Umar 74.2% | Umar -621 | Song Yadong, KO/TKO R2 | Wrong |
| Yan Xiaonan vs Denise Gomes | Gomes 62.8% | Yan -147 | Denise Gomes, KO/TKO R1 | Right |
| Aoriqileng vs Kai Asakura | Asakura 74.0% | Asakura -463 | Kai Asakura, KO/TKO R2 | Right |
| Sumudaerji vs Alex Perez | Sumudaerji 60.2% | Sumudaerji -313 | Sumudaerji, decision | Right |
| Ce Liu vs Levi Rodrigues Jr. | Rodrigues 51.1% | Ce Liu -180 | Ce Liu, KO/TKO R1 | Wrong |
| Bilal Hasan vs Nilson Rojas | Hasan 63.4% | Hasan -837 | Bilal Hasan, KO/TKO R2 | Right |
| Namsrai Batbayar vs André Lima | Lima 72.3% | Lima -333 | André Lima, submission R3 | Right |
| Rei Tsuruya vs Kevin Borjas | Tsuruya 78.2% | Tsuruya -796 | Rei Tsuruya, submission R1 | Right |
| Jack Jenkins vs Sean Woodson | Woodson 58.2% | Woodson -155 | Sean Woodson, split decision | Right |
| Xiao Long vs Francesco Nuzzi | Xiao Long 64.8% | Xiao Long -176 | Francesco Nuzzi, KO/TKO R1 | Wrong |
| Lawrence Lui vs Hector Santiago | Lui 72.4% | Lui -319 | Hector Santiago, KO/TKO R2 | Wrong |
| Xiong Jing Nan vs Julia Polastri | Polastri 58.6% | Polastri -227 | Julia Polastri, KO/TKO R1 | Right |
| Ding Meng vs Cameron Nelson | Ding Meng 57.2% | Ding Meng -121 | Cameron Nelson, decision | Wrong |
The model probabilities above are the final stored numbers, which is what it was saying by fight night rather than on Wednesday. It rescores as cards change, and on this one the Ding Meng line moved a long way in the last two days.
The two fights where they actually disagreed
On eleven of the thirteen bouts the model and the closing odds pointed at the same fighter. They split on two, and they went one apiece.
The model won the bigger one. It had Denise Gomes at 62.8 percent against Yan Xiaonan while the books had Yan as a -147 favorite. That is a genuine disagreement, not a rounding difference: the market implied Yan should win about 59 percent of the time, the model said Gomes should win about 63 percent, and the two numbers cannot both be close to right. Gomes knocked out the fourth-ranked strawweight in 4:49.
The market won the other. The model had Levi Rodrigues Jr. at 51.1 percent, a genuine coin flip, against a Ce Liu the books priced at -180. Ce Liu knocked him out in the first round. In fairness to both, Rodrigues took that fight on short notice after Junior Tafa withdrew, which is the kind of thing neither a model nor a betting line handles well.
One other row is worth flagging even though it was not a disagreement. Sumudaerji came in at 60.2 percent from the model and -313 from the market, and won a unanimous decision by a single round on all three cards, 29-28 from Ben Cartlidge, Vito Paolillo and Clemens Werner. A -313 favorite is not supposed to need the last round. The model priced that fight much closer to what it turned out to be.
What it got wrong, and the worst of it
Umar Nurmagomedov at 74.2 percent. He got knocked out in the second round.
There is no way to soften that. The model was more confident in Umar than a coin flip by a wide margin, the market was even more confident at -621, and both of them were looking at a fighter who had beaten Cory Sandhagen, Mario Bautista and Deiveson Figueiredo by decision and had never been finished in his career. Everything about the profile said "wins rounds, does not get stopped." Song Yadong stopped him at 1:48.
That is a real limitation and it is worth naming rather than burying. A model built substantially on what a fighter's record and per-fight numbers have done in the past is going to be slow to price the possibility that a durable decision fighter meets someone who ends fights. Song has eight post-fight bonuses and seven of them came for a finish. That is not a hidden fact. It is a fact the numbers underweighted.
The three prelim misses tell a smaller version of the same story: Lawrence Lui at 72.4 percent, Xiao Long at 64.8 percent, Ding Meng at 57.2 percent, all beaten, two of them by knockout. On a card where five underdogs won and ten of the thirteen completed bouts ended inside the distance, a model that leans on established form was always going to have a hard evening.
The honest read
Level with the market on hit rate, one-for-one on the two fights where they actually disagreed, and confidently wrong together on the one everybody was watching. If you had used the model as a fade signal on Yan Xiaonan you were paid. If you had used it on the main event you were not, though the market would have hurt you more.
I would rather publish this than a piece about the eight it called. Our predictions live on every upcoming event page and get scored against results whether or not I write about them, and the accuracy tracker is public for exactly this reason. A prediction system that only gets written up when it wins is a marketing asset, not a measurement.
Next test is UFC Fight Night: Hooker vs Parnasse on September 5.