We built four different models to beat Vegas. They made the same mistakes
Boosted trees, a possession simulator, an Elo system and a Kalman filter — four unrelated ways to rate teams — and their errors agree 95% of the time. That agreement is the whole reason the closing line is unbeatable on prediction alone.
Here is the uncomfortable thing we learned trying to out-predict the betting market: it almost doesn't matter which model you build. We built four that share no code and no philosophy — gradient-boosted trees on efficiency stats, a possession-by-possession Monte Carlo simulator, a classic Elo rating, and a Kalman filter that treats team strength as a signal drifting through the season. Different math, different assumptions, different eras of statistics. Then we lined up the games each one got wrong.
