Sports
NFL betting models built around fair price
Football is a low-possession sport with a discrete scoring system, and that makes its margin distribution lumpy in a way no other major American sport's is. A model that ignores where the mass sits will be wrong in the places that cost the most.
Analytics only. quant+ never places bets or holds funds.
Not all half points are equal
Every other property of an NFL model is downstream of this one. Football scores in discrete chunks, so final margins pile up on a few numbers instead of spreading smoothly.
3 and 7
The two numbers where NFL final margins cluster hardest, because the scoring system produces them constantly: a field goal is three points, and a converted touchdown is seven. A spread of minus three is not one point away from minus four. It sits on a mass point, and moving off it changes the set of outcomes that win.
Halves are not small points
Because the distribution is lumpy, a half point crossing a mass point is worth far more than a half point in open space. When a line moves from minus three to minus two and a half, the change is not a rounding adjustment. Treating all half points as equivalent is the single most common pricing error in football markets.
Totals cluster too
Game totals inherit the same lumpiness through the two-point conversion and the range of realistic scoring combinations. A total sitting just above a common combined score behaves differently from one in a gap.
Middles live here
A window that pays both sides requires two numbers with space between them, and the value of that space depends entirely on how much probability sits inside it. Football, with its mass points, is where middles are most structurally interesting and most often mis-estimated.
The practical consequence is that an NFL model's most valuable output is not a single fair number. It is a set of probabilities: the chance the favourite wins by more than two and a half, the chance it wins by more than three, by more than three and a half. Those adjacent numbers look almost identical and are not.
It also means that the model's error tolerance is asymmetric. Being slightly off on a spread sitting in a flat part of the distribution is harmless. Being slightly off near a mass point can flip the recommendation, because the mass point is exactly where a small probability difference changes which side of the number the bet needs.
Six inputs, ordered by how much they earn
Each of these is specific to how football actually generates points. A model built on generic sports features will underperform one built on these.
Drive-level efficiency
Points per drive and yards per play mask the fact that football is a sequence of possessions. Passing efficiency on a per-dropback basis, success rate by down, and red-zone conversion separate a team that moves the ball from a team that scores. Raw yardage rankings are close to useless on their own because they ignore field position, tempo and game script.
Availability, weighted by replacement
Not all injuries matter equally. A starting quarterback, a left tackle and a shutdown corner carry different replacement costs, and the market usually moves on the news before it moves on the number. The adjustment a model should apply is the difference between the starter and the specific backup, not a penalty proportional to how famous the injured player is.
Rest, travel and the week's shape
Thursday games, a bye week, a long road trip across time zones and a Monday night turnaround all sit inside a season-long schedule that gives teams unequal preparation. These are schedule-derived and known in advance, which makes them among the cleanest inputs available.
Weather, where it actually bites
Wind affects kicking and downfield passing far more than temperature does. A model that applies a generic bad-weather penalty to both teams is adding noise, because the effect is concentrated in specific phases of the game and depends on how each team is built.
Market information
Opening lines, the direction of movement, and where the number settles. A model that ignores the market throws away the most concentrated forecast available. A model that follows it too closely has no independent opinion and therefore no edge.
Line type and half-point awareness
A model must express its output on the actual grid the market trades on. Producing a continuous fair spread of minus three point one is less useful than producing the probability of winning at minus three and the probability at minus two and a half, because those are the numbers that trade.
An NFL week has a shape, and the price follows it
Football is the only major American sport with a week-long gap between slates, and that gap is where most of the market's information arrives. Lines open, sit for a stretch with limited liquidity, then move as participation reports come out and again as inactives are confirmed shortly before kickoff.
Those phases are not equivalent. An early-week price is a market opinion with very little money behind it. A price after the final injury report reflects materially more information. A model that produces the same confidence in both windows is overstating itself in the first one.
This is why football rewards a workflow with a scheduled re-price rather than a one-time pick. The question is not only whether the model disagrees with the market, but when it disagrees and whether the number has moved since.
- Window one: line posted. Thin liquidity, wide disagreement between venues, and a number that has not yet been tested. Model disagreements here are largest and least reliable.
- Window two: participation and inactives. Availability information arrives and the number adjusts. A model that already priced the specific replacement player is ahead of one that prices the headline.
- Window three: forecast and stadium. Wind and precipitation forecasts firm up. Totals respond more than spreads, because the effect lands hardest on kicking and downfield passing.
What a football model has to survive before it earns trust
A model that has only been checked against a season of results has not been checked at all.
Point-in-time inputs
Every feature in a historical test must have been available at the moment the bet would have been placed. Season-long statistics used to predict week three are a leak, and it is the most common way a football backtest reports an edge that does not exist.
Market-relative scoring
Grading a football model on win-loss record is nearly meaningless at a sixteen or seventeen game sample per team. The measure that carries information is whether the model's number beat the closing number, evaluated across hundreds of markets.
Season and rule drift
Rule changes, offensive philosophy shifts and officiating emphasis all move the distribution of scoring. A model fitted across many seasons at once blends regimes that are not the same, and the blend can look stable while being wrong about the current one.
Four numbers per market, in a fixed order
Fixed order matters. Reading edge before reading the price that produced it is how a model output becomes a story instead of a measurement.
| Output | What it is | Why it comes in this position |
|---|---|---|
| Projection | The model's expected margin and total, in points | It is the raw estimate, before any market has been consulted |
| Win probability | The chance of covering each of the numbers that actually trade | Because a single fair spread hides the mass points that decide the bet |
| Book-implied probability | The price on offer, converted and with the margin removed | It anchors the comparison to something you can actually take |
| Edge | The signed difference between the two probabilities | It is only interpretable once both inputs above are on the table |
This four-field structure is what the model surface renders for every football market, alongside the inputs that produced it.
Where a football model fails
Football has the smallest sample per season of any major American sport. Sixteen or seventeen games per team means single-season statistics are dominated by opponent quality and schedule luck, and a model fitted on one season is fitting noise. The discipline of shrinking toward prior expectations is not optional here, it is the difference between a model and a spreadsheet of last year's results.
The sport also has the highest rate of unmodelled variance. A single turnover, a blocked kick, a dropped interception or an officiating decision on a catch can swing a game by more than a touchdown. A model that predicts the correct probability will still lose a large fraction of individual games, and no amount of feature engineering changes that.
Finally, an NFL model competes against the most heavily traded markets in North American sports. The most liquid football markets are efficient enough that a modest model has no edge at all, and any approach that reports a large edge on a widely traded main line should be assumed to have a data leak until proven otherwise.
quant+ does not claim to solve any of this. It publishes fair probabilities derived from stated inputs, shows the inputs, and measures the result against closing prices. For the general method, see the guide to building a sports betting model. For the player-level side of football markets, see NFL player prop research.
Model the margin, not the narrative
Import the football slate, price every tradable number against fair value, and see where the distribution says the market is wrong.
Analyze the NFL slate ยท NFL player props
Questions people ask
What is an NFL betting model?
A set of estimated relationships that converts team and situational inputs into a probability distribution over the final margin and the total. The useful output is that distribution expressed at the numbers the market actually trades, so the model says something about minus three and about minus two and a half separately rather than producing one continuous number.
Why do the numbers 3 and 7 matter so much in NFL betting?
Because football's scoring system is discrete and heavily weighted toward three and seven point increments. Field goals produce three, converted touchdowns produce seven, and those are the most common ways a game ends. Final margins therefore pile up on those numbers, which means a half point that crosses one of them changes the set of outcomes that win by far more than a half point somewhere in a flat part of the distribution.
How much does weather matter?
Less than the coverage implies, and unevenly. Wind is the variable that matters most because it affects both kicking and downfield passing accuracy. Temperature and light precipitation are often priced into the market long before the game and add little once accounted for. Applying a blanket bad weather adjustment to both teams usually adds noise rather than information.
Can an NFL model beat the market?
Rarely on the most traded markets, and modestly where it does. Football main lines are among the most liquid in North American sports and are priced by well-resourced participants. Any model reporting a large edge on a widely traded number should first be checked for look-ahead leakage, because that is a far more common explanation than genuine sustained outperformance.
Sources
- What a sports betting model outputs | Rithmm, rithmm.com
