Allen vs Duncan scored: the coin flip the market priced at -209

On Wednesday I wrote that the sportsbook consensus had Malcolm Wellmaker at -225 against Otari Tanzilovi while our model had the fight at a coin flip, and I promised that if the market turned out right, I would say so on Sunday.

It is Sunday. Tanzilovi knocked Wellmaker out in the second round.

That is the headline, and it is also the easiest thing in the world to write about, because it makes the model look clever. So this piece does the other half too. Every bout at UFC Fight Night: Allen vs Duncan, the model's last number before the fight against the last consensus line before the first bell, misses included. The model missed the main event. It missed by a lot.

A note on method, once. The market number is the closing sportsbook consensus with the bookmakers' margin stripped out. The model never sees a line. None of this is betting advice; it is a record of who believed what, and how wrong they were.

The coin flip

At the bell our model gave Tanzilovi 51.4 percent. The market closed at -209 on Wellmaker and +173 on Tanzilovi, which works out to 35.1 percent for the man who won. That was the widest gap on the card, 16.3 points, and the line had barely moved toward Tanzilovi all week: -223 on October 4, when ten books were first pricing it, -209 at the close.

Here is where I owe some honesty about the reasoning, not just the result. Wednesday's case for the model was wrestling: Tanzilovi's 4.5 takedowns per 15 minutes against Wellmaker's 47 percent takedown defense. Tanzilovi did land 2 of 3 takedowns in the first round. But across the whole fight he held 51 seconds of control. This was not a man getting smothered.

It was decided standing. Wellmaker scored a knockdown in the first round, which is exactly the fight the market had priced. Tanzilovi still out-landed him 24 to 9 in significant strikes in that round, dropped him in the second, and the fight was stopped at 2:39. Over the two rounds the significant strikes went 38 to 17, and 36 of Tanzilovi's 38 landed on the feet, at distance or in the clinch. It was Tanzilovi's first win in a UFC-promoted fight, and it came with a Performance of the Night bonus.

So the model got the winner and half of the why. I picked Tanzilovi by decision. I got the winner and none of the how.

The one the model lost

Brendan Allen closed at -115 against -105 for Christian Leroy Duncan, which is 51.1 percent for Allen once the margin comes out. Our model gave Allen 37.1 percent. On Wednesday I laid out its case: every striking number favored Duncan.

Every striking number still favored Duncan on Saturday night. He out-landed Allen 106 to 51 in significant strikes over four and a half rounds, and won the fourth round 46 to 17. If you only read the striking columns, the model nailed it.

Allen won anyway. He landed 4 of 11 takedowns, banked 8 minutes and 39 seconds of control time, and submitted Duncan at 2:46 of the fifth round. That is precisely the part of the fight I wrote the model could not see: Duncan had never fought past 15 minutes in the UFC, and Allen had headlined four cards without losing. Make it five.

That piece closed by saying somebody was going to be wrong in public by Sunday morning, and that it might be us. It was us. On the main event, the market's 51.1 beat the model's 37.1 by 14 points.

Less wrong is not right

The other big gap went the model's way, sort of. The market closed with RJ Harris at -366 and Allen Frye Jr. at +287, or 24.8 percent for Frye. The model gave Frye 39.1 percent.

The line had walked one way all week. When ten books were first pricing it on October 4, Frye was +194. By the close he was +287. The market got more sure of Harris almost every day. Frye then landed 40 of 55 significant strikes, knocked Harris down, and stopped him at 2:59 of the first round. Harris arrived 6-0. He leaves 6-1.

I am not going to dress this up. The model still picked Harris, at 60.9 percent. Giving the winner 39 percent instead of 25 is a smaller miss, not a hit, and the record below counts it as a loss.

The full card, scored

Every number here is the probability each side gave the eventual winner.

Ernesta Kareckaite and Melissa Gatto fought to a majority draw (28-28, 28-28, 27-29), so that bout is out of both records.

Straight up, both sides picked six winners. The market went 6 for 11. The model went 6 for 10, because it called Prado vs Bonfim at exactly 50.0 and refused to pick, while the market flipped that fight across the week, from Bonfim at -130 on October 4 to Prado at -160 at the close, and was right to.

The better test is how much probability each side put on what actually happened. The Brier score does that: the average squared miss, where saying 50 percent on every fight scores 0.250 and lower is better. The model scored 0.283. The market scored 0.296.

Read that twice. Both lost to a coin.

The model's edge, such as it is, comes almost entirely from two fights. On Wellmaker vs Tanzilovi and Frye vs Harris alone, the market piled up 0.38 more squared error than the model. Take those two out and the market wins the other nine comfortably, and it was closer to the winner on six fights to the model's five.

Where they disagreed most

Six bouts had the model and the market at least 8 points apart. The split is as even as it gets.

Three each. If you came here hoping a model had solved fighting, it has not. What it did on Saturday was stay put on two fights where the market spent a week getting more confident in the wrong man.

A bad night for confidence

The card punished certainty from every direction. Market favorites priced at 60 percent or better went 2 for 6: Godinez and Walker won; Matheus Camilo, Wellmaker, Felipe Franco and Harris lost. At those prices the market expected roughly four of them to win. The model's own calls at 60 percent or better went 1 for 4, Harris among them. It had Camilo at 65.7 percent, mostly on age, 25 against 38, and Herbert out-landed him 54 to 43 and took all three cards 29-28. It had Franco at 72.4 percent, and Ribeiro, on a four-fight losing streak, submitted him in 1:49.

Five closing underdogs won, at +119, +159, +173, +287 and +306.

Then there is the line movement, which is supposed to be the market learning. From the first ten-book consensus on October 4 to the close, the line moved toward the eventual winner in six fights and away in five. The two biggest wrong-way moves came on the two fights where the model and the market were furthest apart after Wellmaker vs Tanzilovi: the market drifted 7.3 points toward Duncan in the main event, which was the model's side, and 7.8 points away from Frye, which was against it. It was wrong both times.

For the record, my own picks in Friday's preview went 8 for 11, including Allen by submission. All three I got wrong were the model's favorites too. One card of eleven decided fights proves nothing about anyone, me included, and I would say the same if the numbers had gone the other way.

Next card, same rules: the model's number, the closing line, and every miss on the table. On Saturday the coin beat both of them. The coin did not publish a preview.