Why a Model-Market Gap Is Not Automatically a Signal

The gap is subtraction: projection minus line. As an illustration, if a model projects a game total of 47.1 and the consensus line is 44.5, the gap is +2.6. The arithmetic is exact. What it implies is not.

The model vs market guide covers the vocabulary — projection, line, gap, price and quote time. This piece is about what happens when those pieces meet in practice, and why each one can change what a gap is worth.

Model vs market: projections, lines and prices

Every projection carries error. Even a carefully built model misses individual games by wide margins, because football outcomes are noisy: one turnover, one missed kick or one late score can move a final total a long way.

So a gap has to be judged against the model’s typical miss, not against zero. For illustration, a two-point difference from a model whose typical miss on a game total is around ten points sits well inside its ordinary range of error. The size of a gap only means something relative to how precisely the model has actually performed — which is what prospective tracking exists to measure.

How prospective tracking measures a model in order

A line is offered at a price on each side, and those prices include the sportsbook’s margin. At −110 on both sides of a total, a side has to be right about 52.4% of the time across many repetitions just to break even, because 110 ÷ 210 ≈ 0.524.

That threshold applies before anything else. A model that picks the right side slightly more often than a coin flip can still fall short of it. A gap therefore has to point in a direction the model is reliably right about, often enough to clear the margin — merely pointing somewhere is not enough.

Sportsbooks do not all quote the same number at the same moment. If one shows 44.5 and another 45.5, the size of the gap depends on which one is used. Choosing the most favourable book after the fact always makes a gap look larger than a fair comparison would.

First Down Intel compares against a consensus built from the books it tracks, and reports how many books contributed, so a comparison never depends on hand-picking a single quote.

A projection compared with a line observed hours earlier or later is not a like-for-like comparison. Lines move with news. A quote taken before an injury announcement can produce a large, meaningless gap against a model that already reflected the news — or a model that did not.

Every market price FDI shows carries its observation time, and a quote that is no longer current is marked that way. When a frozen prediction is later compared with the market, only quotes that existed when the prediction was made are used: a later quote is never treated as something the model could have known.

A projection rests on assumptions about who plays. If a starting quarterback is ruled out, or the conditions for a game change, both the model and the market have to respond. Until both have, a gap can simply mean one side has updated and the other has not.

For player projections this matters even more. A projection for a player who does not take part has nothing meaningful to be compared with. In FDI’s prospective records such a prediction stays visible but is not graded — a missed game is never scored as a zero.

Put together, these points mean the arithmetic size of a gap does not, by itself, establish expected value. Getting there would require knowing:

• the model’s measured error on comparable games, recorded prospectively;
• whether the comparison quote was current and representative of the market;
• whether availability or conditions changed after either number was formed;
• whether the difference is large and reliable enough to clear the book’s margin.

Most of those are properties of a model’s record over many games, not of any single game. A gap on one game cannot answer them.

First Down Intel shows model context beside market context and is explicit about where evidence is thin. It does not turn a gap into an instruction. A Moneyline Lean, where one qualifies, is directional model context rather than a recommendation, and many games have no read at all. The Intel Score describes how strongly the evidence supports the way a read was built — not the chance that it is right.

The honest order of operations is simple: observe the market with its timestamp, read the model with its limits, and let a prospective record — frozen before kickoff, settled after the game — show over time whether the model’s differences with the market have meant anything. Until that record exists, a gap is a question, not an answer.

Explore NFL intelligence from First Down Intel