The Second Round of the World Cup Group Stages Starts Today!
Our model predicts all four games, and we ask Claude to do the same. Can we be AI?
The second round of group games starts today. Every team has played once, the tables are forming, and we get four matches across Groups A and B. Here are the plays.
Today’s key picks
Czech Republic to win, against South Africa. Our biggest winner-market edge in a while.
Switzerland vs Bosnia: Under 2.5 goals.
Canada vs Qatar: Under 2.5 goals.
Mexico vs South Korea: no play, our model lands right on the market.
The four games
Czech Republic vs South Africa, 12 ET
This is the bold one. The market makes Czech Republic about a 53 percent favorite. Our model has them at 79. That is a 26-point gap, and it is unusual for us, most of our edges live in the goals markets, not the winner. Czech lost their opener to South Korea, but the model still rates them well clear of a South Africa side that was beaten by Mexico. Czech to win is the play.
Switzerland vs Bosnia, 3 ET
Switzerland were held by Qatar and will want to put it right. The model sees a controlled, low-scoring win, a projected 1-0, so the Under 2.5 is the play. Nothing on the winner, where we agree with the market that Switzerland are favored.
Canada vs Qatar, 6 ET
Canada are heavy favorites and we do not argue with that. The edge is in the goals again. The model expects Canada to manage this without a shootout, so Under 2.5 is the play.
Mexico vs South Korea, 9 ET
A pass. The model has Mexico narrowly ahead, almost exactly where the market sits, so there is no ten-point gap to bet. Some nights doing nothing is the right call.
New this round: we are racing Claude
Here is something we are starting today, and I want to explain how it works because it is genuinely interesting.
Alongside our statistical model, we are now logging a second forecaster: Claude, the AI, predicting the same games completely blind. It does not see our model, it does not see the betting market. Each matchday a fresh instance gets only the fixtures and the results so far, and it reasons from its own football knowledge. Then we put its picks next to ours, and next to the market, and we grade all three.
The scorecard is not just who picks the winner. We are tracking two things. The first is plain accuracy, how often each forecaster’s most likely call actually happens. The second, and the one that really matters, is calibration, measured with something called a Brier score. Brier rewards being honest about your confidence. If you say a team is 70 percent and it wins about 70 percent of the time, you score well. If you say 90 percent and it only comes in 60, you get punished for the overconfidence even if the favorite wins. Lower Brier is better, and it is the cleanest way to separate a forecaster who is genuinely sharp from one who just keeps backing the obvious favorite.
So three contestants: our model, a blind Claude, and the market as the benchmark to beat. We pre-register every pick before kickoff and grade them as the results land, so there is no hindsight.
The early read is already a good story. On today’s four games, Claude and our model agree on every winner, but they split on the goals, and they split hard on confidence. The clearest example is Czech Republic: our model says 79 percent, Claude says 54, right in line with the market. So today is a real test of a simple question. Is our model’s willingness to disagree with the crowd an edge, or just noise dressed up as conviction? The Brier score will tell us, and we will show it either way.
Across the tournament so far our published picks sit at 14-7 with a return around 51 percent, and on accuracy we are hitting about 70 percent on totals. Now we get to see how that holds up against an AI that has never seen a spreadsheet. Receipts out tomorrow.