This page publishes our model's number for every upcoming NFL game beside the market's own number for the same game, and it names the side those two numbers imply. That side is arithmetic, not a selection, and it is not a pick. On the games graded so far the model trails the market on both the moneyline and the total, and this page publishes that gap.

Model v2, stamped 2026-09-13 and rebuilt every Tuesday. For the market's full board see the NFL odds board, the per book prices on the price grid, and the full ledger on the track record.

+0.0138how far behind the market we areBrier score gap on the moneyline, 1,404 graded games. Lower is better, so a positive number is us trailing.

Where this model stands

Our NFL model is measured against the market on every graded game. It currently trails the market on both moneyline and total points, and we publish the gap.

That sentence is the whole standing disclosure and it is not softened anywhere on this page. The size of the gap is set out further down, measured on graded games with the count printed beside it.

This week's games

GameDateModel home winMarket home winModel totalMarket totalOur projectionStatus
Detroit Lions at Buffalo BillsThu Sep 1756.6%60.4%54.452.5BUF 56.6%Over 52.5Not played yet
Carolina Panthers at Atlanta FalconsSun Sep 2060.9%52.8%44.743.8ATL 60.9%Over 43.8Not played yet
Cincinnati Bengals at Houston TexansSun Sep 2058.1%55.9%44.345.8HOU 58.1%Under 45.8Not played yet
Cleveland Browns at Tampa Bay BuccaneersSun Sep 2063.0%73.7%41.540.5TB 63.0%Over 40.5Not played yet
Green Bay Packers at New York JetsSun Sep 2025.3%31.9%45.742.5GB 74.7%Over 42.5Not played yet
Indianapolis Colts at Kansas City ChiefsSun Sep 2054.0%70.2%44.447.5KC 54.0%Under 47.5Not played yet
Jacksonville Jaguars at Denver BroncosSun Sep 2050.6%56.8%41.644.5Near evenUnder 44.5Not played yet
Las Vegas Raiders at Los Angeles ChargersSun Sep 2066.5%77.4%40.942.5LAC 66.5%Under 42.5Not played yet
Miami Dolphins at San Francisco 49ersSun Sep 2064.1%84.5%45.646.5SF 64.1%Under 46.5Not played yet
Minnesota Vikings at Chicago BearsSun Sep 2046.1%59.4%41.145.5MIN 53.9%Under 45.5Not played yet
New Orleans Saints at Baltimore RavensSun Sep 2063.7%76.6%44.046.5BAL 63.7%Under 46.5Not played yet
Philadelphia Eagles at Tennessee TitansSun Sep 2032.8%32.2%44.042.5PHI 67.2%Over 42.5Not played yet
Pittsburgh Steelers at New England PatriotsSun Sep 2067.5%64.6%40.943.5NE 67.5%Under 43.5Not played yet
Seattle Seahawks at Arizona CardinalsSun Sep 2028.3%19.7%45.144.5SEA 71.7%Over 44.5Not played yet
Washington Commanders at Dallas CowboysSun Sep 2047.8%65.4%52.251.5WAS 52.2%Over 51.5Not played yet
New York Giants at Los Angeles RamsMon Sep 2168.1%80.1%50.048.5LAR 68.1%Over 48.5Not played yet

16 games. None of these games has been written into the record below yet. The weekly Tuesday update is what writes them down, and until it does these are the model's current numbers rather than logged projections. The side is read off the two numbers to its left and nothing else: the team our win probability puts above a coin flip, and Over or Under depending on which way our projected points fall against the quoted line. In 1 of these games our win probability is inside two points of a coin flip, so it says near even rather than naming a team it does not actually separate. Our side is the other team from the market's favourite in 2 of them. That is worth knowing and it is not worth acting on, for the reason set out below. "Model" is this week's published number for the home team winning and for the two teams' combined points. "Market" is the consensus of each book's latest archived quote before kickoff, with the bookmaker margin removed. It is not a closing line.

What the model's side is, and what it is not

The side is arithmetic. Whichever team our win probability puts above a coin flip is the side, and whichever way our projected points fall against the quoted line is Over or Under. There is no judgement in it, no game we single out and no game we leave alone, and a win probability inside two points of a coin flip gets no side at all, because a number that close has not separated the two teams.

What it is not is a recommendation. It is tempting to read our side disagreeing with the market as something to act on. On our own measured record it is not. Over 1,404 graded games our moneyline probability scores 0.2237 against the market's 0.2099 on the same games, and lower is better. Where the two disagree, the market has been right more often than we have. We print the side because doing the subtraction in your head is our job and not yours, and we print the record beside it for exactly the same reason.

How far behind the market we are

MeasureOur numberThe marketGraded games
Moneyline, 2020-2024 out of sample0.22370.20991,404
Moneyline, 2025 held-out season0.22060.2119284
Total points, 2020-2024 out of sample10.5910.301,408
Total points, 2025 held-out season10.5910.43285

Moneyline is a Brier score and total points is an average error in points. Lower is better in every row, so the market's is the better number in every row. On the two 2025 rows the gap is small enough that a single season cannot measure it, which is a statement about how little 285 games can settle and not a claim that we have caught up.

The record so far

Every projection here was written down before kickoff and cannot be edited afterwards, so this fills in on its own as results land, losses included.

How the side we named has done, across every week we have published.

Which version published which week. Week 1 of 2026 was projected by model v1. Model v2 serves from 2026-09-15 and is what the games above are built on. Version 1 receives no new games ever again, and it is still graded, because the games it projected are real and they are owed their results. The record is therefore split by version and a pooled figure can never hide inside a version change.

Model versionStateProjections loggedGraded
Model v1Closed, still graded320
Model v2Serves from 2026-09-1500

Counts are across both markets, the moneyline and the total. A projection counts as graded only when both our number and the market's number exist for that game, so a game missing either one counts on neither side.

Every side we have published

GameWeekVersionOur projectionStatus
NE at SEA2026 week 1Model v1SEA 69.7%Under 44.5Waiting on the result
SF at LAR2026 week 1Model v1LAR 75.0%Under 48.0Waiting on the result
ARI at LAC2026 week 1Model v1LAC 62.2%Under 47.5Waiting on the result
ATL at PIT2026 week 1Model v1ATL 52.8%Over 42.0Waiting on the result
BAL at IND2026 week 1Model v1BAL 67.2%Under 48.0Waiting on the result
BUF at HOU2026 week 1Model v1BUF 57.4%Under 44.5Waiting on the result
CHI at CAR2026 week 1Model v1CHI 61.5%Under 47.0Waiting on the result
CLE at JAX2026 week 1Model v1JAX 70.8%Over 40.5Waiting on the result
DAL at NYG2026 week 1Model v1Near evenUnder 48.0Waiting on the result
GB at MIN2026 week 1Model v1MIN 54.9%Under 46.5Waiting on the result
MIA at LV2026 week 1Model v1MIA 67.8%Over 40.5Waiting on the result
NO at DET2026 week 1Model v1Near evenUnder 50.0Waiting on the result
NYJ at TEN2026 week 1Model v1TEN 59.5%Over 39.5Waiting on the result
TB at CIN2026 week 1Model v1CIN 57.7%Under 50.5Waiting on the result
WAS at PHI2026 week 1Model v1PHI 67.9%Under 44.5Waiting on the result
DEN at KC2026 week 1Model v1DEN 53.6%Under 43.5Waiting on the result

Every side we have published, oldest first, with the model version that produced it. Nothing is removed from this list once a result lands.

What the model is

Each number is the average of two independent estimates of the same game. The first reads each team's recent efficiency, its prior-season strength, rest, travel, kickoff slot, surface and roof, and the announced starting quarterback. The second simulates the game's scoring events directly and plays it twenty thousand times. Averaging two independent estimates is a standard way to reduce error. It does not add information that neither of them had, and neither of them reads the betting market, which is what keeps the comparison on this page honest.

An improved model is a stated goal, not a promise. The version label above is the real one, and when it changes this page will say what changed and show both sets of numbers.

How these numbers are made and checked

Frequently asked questions

Is the model's side a pick?

No. It is arithmetic done for you: the team our published win probability puts above a coin flip, and Over or Under depending on which way our projected points fall against the quoted line. We publish no picks, no selections and no recommended wagers, and our record against the market is printed beside every side.

Your side and the market's favourite disagree. What does that mean?

On our measured record it means we are more likely to be wrong than the market is. Over 1,404 graded games our moneyline probability scored 0.2237 against the market's 0.2099 on the same games, and lower is better.

Why does a game say near even instead of naming a team?

Because our win probability for it is inside two points of a coin flip. A number that close has not separated the two teams, so naming one would be inventing a side the number does not support.

Why publish a model that trails the market?

Because the number is still informative and the honest thing is to show it with its record attached. Every projection is logged before kickoff and graded afterwards, losses included, so the claim is checkable rather than asserted.

What is the market number exactly?

The consensus of each bookmaker's latest archived quote strictly before kickoff, with the bookmaker margin removed. It is not a closing line and we never describe it as one.

Related: NFL odds board · NFL price grid · Public track record

This guide explains how a betting market works and how the arithmetic is done. It is information, not betting advice: Nebula Insights publishes no picks, no selections and no recommended wagers. Sportsbook rules vary by book and jurisdiction, so always confirm the house rules that apply to your account. You must be of legal gambling age in your jurisdiction.