How Sportsbooks Price NFL Totals (And Why the Model Keeps Failing)
Every NFL total starts with a probability model. For most sports, the math behind the total is built on the Poisson distribution, a statistical framework for estimating how often rare, independent events occur in a fixed window of time.
The model works well for soccer. It works reasonably well for baseball. For NFL football, it breaks down in three specific and measurable ways, and those breakdowns produce persistent pricing errors in the market.
This is not about fade-the-public strategy or line shopping. This is about what the pricing model assumes and where those assumptions fail against the structure of football as a sport.
A betting strategy based on adverse-weather totals won at a 59.7% rate in an academic study covering the 2001-2009 NFL seasons. The break-even rate at -110 juice is 52.38%. The gap exists because the model behind the total does not fully price extreme conditions.
The Starting Model: Poisson Distribution
Poisson modeling in sports was formalized in 1982. Statistician M.J. Maher published "Modelling Association Football Scores" in Statistica Neerlandica, fitting a Poisson model to soccer match data by assigning each team an attack strength and a defense strength. The model's core assumption: each goal scored is an independent event. How many goals the home team scores does not depend on how many the away team scores.
Maher found the independent Poisson model gave a reasonably accurate description of football scores. He also found a bivariate Poisson model, which introduces a small correlation (roughly 0.2) between the two teams' scoring, improved fit modestly.
Dixon and Coles extended this work in 1997. Their paper, "Modelling Association Football Scores and Inefficiencies in the Football Betting Market," appeared in the Journal of the Royal Statistical Society Series C. They found one additional weakness in the pure Poisson model: in low-scoring games (specifically 0-0, 0-1, 1-0, and 1-1 final scores), the independence assumption breaks down more severely. Their correction for this low-score dependence improved accuracy at those score lines and, critically, produced a positive return when applied as a betting strategy.
Both papers use "football" to mean soccer. Goals are rare events, averaging 2-3 per match across both teams. Each goal is largely independent of the previous one. The Poisson distribution is designed for exactly this kind of process: rare, uncorrelated events in a fixed time frame.
Why Soccer Fits. Why NFL Does Not.
The Poisson model works well for sports where three conditions hold:
- Scoring events are rare and arrive in single units.
- Each scoring event is independent of the others.
- The variance and mean of scoring are approximately equal (a defining feature of the Poisson distribution).
Soccer satisfies all three. Baseball satisfies them well enough to support similar modeling. NFL football fails all three.
| Condition | Soccer | Baseball | NFL Football |
|---|---|---|---|
| Rare, unit-sized scoring events | Yes (goals: ~2-3/game) | Mostly (runs: ~8-9/game) | No (3-pt, 6-pt, 7-pt lumps) |
| Independence between scoring events | Mostly, with low-score exception | Largely yes | No (game script correlation) |
| Variance ≈ mean (Poisson fit) | Yes | Close | No (overdispersion. Variance exceeds mean) |
The NFL failures are structural, not coincidental. They arise from the specific nature of football as a sport.
First, scoring arrives in discrete lumps. A touchdown is worth 6 points, plus an extra point (1) or two-point conversion (2). A field goal is 3 points. A safety is 2 points. There is no unit scoring event in football. Every score shifts the total by multiple points at once, creating a distribution of combined scores that the Poisson model, which treats events as arriving one unit at a time, does not accurately describe.
Research on collegiate and professional American football scores confirms this. The negative binomial distribution fits NFL scoring noticeably better than Poisson because negative binomial handles overdispersion: situations where the variance in outcomes exceeds what the mean would predict. In Poisson distributions, variance equals the mean by definition. NFL scoring refuses that constraint.
Second, and more important for betting: scoring events in NFL football are not independent.
The Independence Problem: Game Script
In soccer, a goal changes the score by 1. Both teams continue playing largely the same way. The fundamental game structure does not shift dramatically when one team leads by two goals.
In the NFL, a team trailing by 17 points in the fourth quarter operates in a completely different way from a team trailing by 3. The trailing team abandons the run, throws on roughly 85% of snaps compared to a 60% league baseline, and accepts a prevent defense against them. The leading team runs the clock, takes fewer risks, and stops trying to score. The scoring events in the final quarter are not independent of what happened in the first three quarters. They are determined by it.
This is game script. And the data supports its impact on totals.
The result for totals pricing: blowout games inflate combined point totals in ways the opening market did not model as high-probability events. A total of 44.5 for a game that becomes 31-7 by halftime is no longer a bet on those two teams' offenses in a competitive game. The final score reflects a game environment that the model never anticipated when it set the number.
The Poisson model assumes each scoring event is independent. NFL football produces scoring events that are systematically dependent on game context. The model's assumption fails every time a blowout occurs, and blowouts occur in roughly one in four NFL games.
The Baseline Problem: Five Years of Scoring Decline
Before addressing specific failure modes, one number sets the context. NFL teams averaged 21.4 points per game in the 2024 season, the lowest per-team average since 2006 and the fifth consecutive year the figure declined. Combined, both teams in a typical 2024 NFL game produced roughly 42.8 points.
The typical opening total for NFL games in recent seasons has ranged from 44 to 47. Market instinct anchors to historical expectations. The actual scoring environment has moved below those levels for five years running.
This is not a structural inefficiency on its own. Books adjust. But it establishes a baseline: the market's starting assumption about how much two NFL teams will score in a game is already optimistic relative to recent data. The three failure modes below widen that gap further.
Three Specific Market Failures
1. Adverse Weather
The most thoroughly documented failure in NFL totals pricing is weather. The Poisson-adjacent model prices totals using team attack and defense ratings derived from historical games. Those ratings represent average-game performance. Extreme conditions are not adequately discounted from those averages in real time.
Richard Borghesi documented this in "Weather Biases in the NFL Totals Market," published in Applied Financial Economics in 2008 (Vol. 18, No. 12). Using data from the 1984-2004 seasons, Borghesi found that heat, wind, and rain all reduce NFL scoring, and the market consistently underestimates the magnitude of the effect under extreme conditions.
His adverse-weather betting strategy produced a 59.7% win rate on unders during the 2001-2009 seasons. At standard -110 juice, bettors need 52.38% to break even. The 7.3-point gap between 59.7% and 52.38% represents a meaningful and persistent inefficiency.
The mechanism: the model behind the total treats quarterback arm strength, receiver route running, and field goal range as constants derived from prior games. Temperature below 25°F and wind above 15 mph compress all three. The model does not discount for conditions it was not built to price. Bettors who understand this structure know where to look. The when is secondary to the why.
The DawgHousePicks NFL wind and totals analysis documented three specific wind thresholds that produce measurable scoring suppression. The reason those thresholds matter is the model failure described here.
2. Blowout Inflation
The game-script problem runs in both directions for bettors depending on when they act.
On the pre-game total, games expected to be blowouts present a structural under opportunity. When one team is favored by 14 or more points, the game is pricing in a dominant performance. The final score will likely include fourth-quarter garbage-time scoring from the trailing team that was not part of the original market assessment. The total's opening price does not fully price these points as likely because, before the game, they were not likely. After a 28-0 second quarter, they become near-certain.
On live totals, the same mechanics create an over opportunity in the opposite scenario: once a blowout is confirmed, the live total should reprice to include trailing-team garbage-time scoring. Books that update slowly leave a window. Sharp live bettors who track possession count, drive efficiency, and pace have an edge in this window because the model the book uses for live pricing faces the same independence-assumption problem as the pre-game model.
The practical application: avoid laying heavy juice on unders in games with large spreads. The garbage-time scoring the trailing team will produce in the fourth quarter was not in the opening total. The market prices competitive offensive output. The final score includes scoring the model never anticipated.
3. Dome Team Displacement
NFL teams that play home games in climate-controlled stadiums develop attack-strength ratings built on games played in a dome environment. Those ratings are what the totals model uses when it prices their road games.
When a dome-based team travels to a cold outdoor game in late season or the playoffs, the model is applying dome-game attack ratings to an outdoor cold environment. The quarterback, receivers, and kicker all perform differently in cold than in a climate-controlled stadium. The model does not know this.
Indoor games produce a higher scoring floor than outdoor games. The rate of indoor games going over vs. under is nearly equal (49.9% over, 49.2% under), reflecting that total pricing for indoor games is relatively accurate. Outdoor cold weather games skew under when the temperature drops significantly. When the visiting offense is a dome-based unit whose ratings were built on indoor production, the outdoor suppression effect on their performance is larger than the model accounts for.
The specific setup to target: dome teams (Rams, Saints, Falcons, Lions, Vikings in their new facility, Chargers) traveling to outdoor games in Buffalo, Kansas City, or Chicago in November or December. The model overrates the visiting offense. The total should be lower than where the market sets it.
Applying the Framework
These three failures share a common structure. The model behind the total uses team attack and defense ratings from prior games. It then applies those ratings to an upcoming game assuming independence between scoring events and an average-game environment. When either assumption breaks, the total misprice follows.
Three questions to run before betting any NFL total:
- What is the weather forecast? Temperatures below 25°F or sustained wind above 15 mph suppress scoring in ways the model does not fully price. Borghesi's evidence shows unders win at above-threshold rates in adverse conditions even after the market became aware of the bias.
- What is the spread? Games with large spreads (14 or more) price in a blowout. Fourth-quarter garbage-time scoring from the trailing team will inflate the final total above what the model assumed when it built the opening number. Avoid laying juice on unders in heavy-spread games unless the weather suppression is severe enough to offset the garbage-time inflation.
- Is a dome team playing outdoors in cold weather? Dome-team ratings overstate outdoor cold-weather production. The total has not adequately discounted this displacement.
When two or three of these factors stack, the gap between the model price and the fair price widens further. A dome team visiting Buffalo in December with wind over 15 mph and the spread at 10 is a situation where all three failure modes point in the same direction.
What the Model Cannot Fix
The Poisson-based framework has been refined significantly since Maher's 1982 paper. Dixon and Coles added the low-score correction. Subsequent researchers have incorporated time-decay weighting, home-field parameters, and injury adjustments. Books have gotten better at weather adjustment since Borghesi's original 2008 findings.
But the independence assumption cannot be fixed within the Poisson framework itself. NFL football is a sport where the state of the game determines how both teams play. Scoring events depend on the scoreboard. A model built on independence between events is applying a tool from a world where that dependence does not exist to a sport where it defines the fourth quarter of every close game.
That structural gap between the model and the sport does not disappear when the market becomes more sophisticated. It becomes smaller. Finding it requires knowing what the model is, what it assumes, and where those assumptions break against actual NFL football. The three failure modes above are the places where the gap is widest and most persistent.
The 2024 league average of 42.8 combined points per game, below market instinct for five straight years, adds one more layer: the market's baseline assumption is too high before any situational factors are applied. When extreme weather, a blowout spread, or dome displacement compounds that baseline mismatch, the total's fair price drifts meaningfully from where the line sits.
Bet the gap, not the headline number.