World Cup Day 9: We've Made Some Changes to the Model

We made some changes to the model to match reality. The four games today have some surprising modeled outcomes.

Round two rolls on. Groups C and D play their second games today, four matches, and the slate runs from the hosts at noon to Turkey and Paraguay at night. Here are the plays.

Today’s key picks

• Scotland vs Morocco: Under 2.5 goals. Our biggest edge of the day by a distance.

• Scotland vs Morocco: the draw, a second play in the same game.

• USA vs Australia: Australia to win. The market makes the US a clear favorite, our model has it a coin flip.

• Brazil vs Haiti and Turkey vs Paraguay: no play, the model lands where the market is.

The four games

United States vs Australia, 12 ET

The bold one today. The market makes the US about a 62 percent favorite. Our model will not separate these teams at all, it has the game a dead coin flip, 35 percent each way. So Australia, priced as a long shot, is the value. If you think this is closer to even than the market does, and we do, Australia to win is the play. The goals lean slightly under.

Brazil vs Haiti, 3 ET

About as lopsided as the board gets. The market has Brazil near 88 percent and we do not argue. In fact the model wanted Brazil even higher, and this is the first game where our new cap steps in and holds them at 65, which I will explain below. No winner play, and the goals are priced fairly. A pass on the card, a likely Brazil win on the pitch.

Scotland vs Morocco, 6 ET

The game of the day for us. Morocco are the better side and a fair favorite, but the model sees a tight, low-scoring grind, barely more than a goal in the whole match. The single most likely scoreline is 0-0. That makes Under 2.5 the standout play of the slate, a 28-point gap on the market, and the draw, at a 12-point edge, is the second play. Two bites at the same low-scoring read.

Turkey vs Paraguay, 9 ET

A pass. Turkey are a modest favorite and the model agrees, landing almost exactly on the market with no ten-point gap to bet. The goals lean under but not enough to publish. Some nights the market just has it right.

The nerdy bit: we capped our own confidence

Here is what changed under the hood overnight, and it comes straight out of yesterday.

Our model had the Czech Republic at 79 percent against South Africa. The market had them at 53. The game finished 1-1. When your single most confident call of the day misses, the scoring penalty is brutal, and it was not the first time the model has run hot on a big favorite.

So we tested a fix instead of guessing at one. On the last two World Cups, 2018 and 2022, that is 96 games the model never trained on, we capped the top win probability at a range of levels and measured what actually helped. Capping the favorite at 65 percent won cleanly. It lowered the error on games the model had never seen, and it never touched how often the model picked the right side. Then we checked it against this tournament too, and it helped here as well, pulling our winner score down further. The cap is live as of today, which is why Brazil reads 65 and not 80.

We also tried the opposite, propping the draw up, because this tournament has been draw-heavy and our draws looked a touch low. We threw it out. It flattered the recent games and hurt the backtest, which is the tell of a tweak that is just chasing what already happened. The rule we hold to is simple. If it only works on the games we have already seen, it does not ship.

The reason any of this matters is the only scoreboard that counts. Across every game this tournament where we have both our number and the market’s, we are now beating the market on game winners and on over-under both, the second one by a clear margin. A homemade model is not supposed to do that against a real exchange. We are, for now, and the cap is part of why.

Our published card sits at 14 and 10 on the tournament, a return around 32 percent, and we are a more disciplined version of ourselves today than we were yesterday. Receipts as the games land.