How to Read Our Model-Health Page
The methodology page isn't decoration. Here's how to read tier hit rates, sample size, and edge decay so you know when to lean in and when to sit out.
Most picks services show you a record and ask you to trust it. We show you the machinery. The methodology page pulls every number live from the same ledger that grades our picks, so you can audit us instead of believing us. Here's how to actually read it.
Tier hit rates are the headline, sample size is the fine print
Every pick carries a confidence tier: S, A+, A, B, C. The hit rate next to each tier is the percentage of completed picks at that tier that won against the spread. That number is meaningless without the count beside it.
A tier sitting at 100% on 4 picks tells you nothing. A tier at 58% on 240 picks is a real edge. When you read the page, read the two numbers together:
- n < 30 at a tier: treat the rate as noise. Don't size up on it.
- n between 30 and 100: a signal forming, not yet a fact.
- n > 100: this is the rate you can plan around.
The breakeven for a standard -110 spread bet is 52.4%. Any tier clearing that on a large sample is paying its rent. Anything under it, regardless of how confident the label sounds, is not.
Calibration: do the probabilities mean what they say
A model that says "65%" should win 65% of the time on those picks. When it does, it's calibrated. When it says 65% and wins 50%, it's overconfident, and you're the one paying for that confidence.
We surface this as a calibration curve and an ECE (expected calibration error) figure. Lower ECE is better. A well-calibrated model lets you trust the gap between tiers: an S pick really should be safer than a B pick, not just labeled that way.
Edge decay: the part nobody wants to show you
Markets adapt. A strategy that printed in 2021 can be dead by 2024 because the books closed the inefficiency. We run change-point detection on rolling accuracy to catch that drift, and we publish it.
If you see a tier's recent accuracy diverging hard from its lifetime accuracy, the edge may be decaying. That's not a reason to panic. It's a reason to weight recent performance over the all-time number when you decide how much to put behind a pick.
The disagree flag is where the real money is
When our pick is on the opposite side of the consensus moneyline, we mark it
disagree and surface it with a ▲. Those are the picks where the model sees
something the market doesn't. They're rarer, they're higher-variance, and
disagree_roi is the one ROI figure we compute in real units (1u = $100).
If you only watch one number on the health page, watch how the disagree bucket is performing. Agreeing with the market is easy. Beating it is the whole point.
What to do with all this
Read it in this order: tier hit rate, sample size next to it, calibration, then recent-vs-lifetime accuracy. If a tier clears 52.4% on a large, calibrated sample and isn't visibly decaying, that's a pick you can stand behind. Everything else is a maybe, and we'd rather you sit out a maybe than chase a label.
The page updates every hour off live results. Bookmark it and check it before you trust any pick we drop.