The confidence score

Every grade carries two numbers: the edge, and how much the engine trusts it. The second one is the more honest of the two.

What it actually measures

Confidence is a 0-to-1 score describing the strength of the evidence behind a probability. It says nothing about whether the bet is good.

Three inputs move it:

Sample depth. How many relevant historical observations the engine found. A player with forty logged games supports a far more trustworthy projection than one with four.

Book coverage. How many sportsbooks are pricing the market. A market only one book offers has no cross-book consensus to de-vig against, so the fair price itself is less reliable.

Consistency. How stable the underlying numbers are. A player averaging 60 yards with a tight spread is more predictable than one averaging 60 with weekly swings between 5 and 150, even with identical sample sizes.

The gate

Below a threshold, the grade is suppressed and the bet is published as THIN DATA with no rating attached.

This is the most important behaviour in the whole system, and it is worth being precise about why. A model handed eight games will happily produce a probability, and that probability will frequently disagree wildly with the market. Presented as a big edge, it looks like the best bet on the board. It is almost always the model being wrong.

How it changes what you do

Confidence is not a size multiplier. It is a trust multiplier.

• High confidence, meaningful edge. The edge means roughly what it says.
• Low confidence, meaningful edge. Treat it as a prompt to look at the game yourself, not as a recommendation. The engine is telling you it noticed something and cannot vouch for it.
• High confidence, no edge. A genuinely efficient price. Pass with some certainty rather than looking for a reason.
• Below the gate. No grade at all.

Its relationship to shrinkage

Confidence also feeds back into the probability itself. Model output is blended toward the market's fair number in proportion to how much history sits behind it — deep samples move the number a long way, thin ones barely move it.

So confidence does two jobs: it decides whether a grade is published, and it determines how much the model was allowed to disagree with the market in the first place. Early in a season, when confidence is low across the board, the published probabilities sit very close to the market by design.

What it does not capture

Confidence measures evidence, not correctness. A deep, consistent sample from a team whose situation has fundamentally changed — a new coordinator, a key injury, a personnel overhaul — produces high confidence in a number that is nonetheless wrong.

The engine has no variable for "this team is different now". That limitation, along with the others, is set out on known limitations.