Here's a question most cappers can't answer: what's your model's calibration error?

If someone selling you picks doesn't know what that means, they're not using a model. They're using vibes.

I spent 6 months building a 4-model consensus system for NHL betting. Here's the stack:

CMP-Dixon-Coles — adjusts for home ice, recent form, and scoring rates using a Poisson framework. It's the baseline.

Elo+MOV — classic Elo ratings but weighted by margin of victory. Catches teams that are better (or worse) than their record suggests.

Glicko-2 — adds rating volatility. A team on a 5-game win streak against weak opponents doesn't get the same confidence as one beating top-10 teams.

Market Ensemble Devigging — strips the vig from betting lines to extract implied probabilities, then compares against our other 3 models.

All 4 models have to agree before a pick gets released. Then it passes through a Bayesian Kelly gate — if P(EV>0) isn't at least 60%, it doesn't go out.

263 NHL games of training data. Every pick independently logged on CapperTek.

Most people in this space sell confidence. We sell math.