We built four different models to beat Vegas. They made the same mistakes

Trying to out-predict the betting market taught us something we did not enjoy learning: it almost doesn't matter which model you build. We built four that share no code and no philosophy, starting with gradient-boosted trees on efficiency stats, then a possession-by-possession Monte Carlo simulator, then a classic Elo rating, then a Kalman filter that treats team strength as a signal drifting through the season. The math is unrelated and the assumptions clash. They come out of different eras of statistics. Then we lined up the games each one got wrong.

They get the same games wrong. The correlation between any two of these models' errors runs between 0.94 and 0.97. Same estimator, four costumes. Blending all four buys a rounding error, about a twentieth of a point.

Why four different models converge

Because they all ask one question, how good is this team, and they ask it of one body of evidence: what teams have done on the field. On-field results contain a fixed amount of information about team strength, and every competent method extracts essentially all of it. Once you've wrung that towel dry, a fancier method can't squeeze out water that isn't there. The ceiling is the information.

What this means for beating the number

The market's closing line is the aggregate of everyone doing exactly what we did, plus information we don't have: injuries that break late, weather, and the money of people who know things. Across every full slate from 2019 through 2025 our best honest model picks sides at 50.7% against that close. A coin flip. That is what an efficient market looks like from the inside.

We didn't take that as cause for despair. If a better model can't win, stop building models and go looking for the places the market isn't efficient. The opening line before it sharpens, for one, and the price differences between books. That's where the next three things we found actually live.