Football Model vs. Quant Model vs. the Blend: Which Was Most Accurate?

The football model builds projections from the ground up. Team scoring, coaching tendencies, depth charts, player opportunity, injuries, offensive line play, and expected game flow all have to fit together. If the team numbers don't add up, the player projections don't mean much.

At Sportshack, that football work still has to answer to the results. A projection is supposed to be tested. If another method sees a player or statistic better, the correct response is not to defend the original number. The correct response is to find out why.

I now have enough historical weekly projections and player-game results to compare two different approaches:

The goal was simple: determine which method assigned the best probabilities to passing yards, passing touchdowns, and receiving yards during the 2025 season.

What Was Actually Tested

The test joined the 2025 weekly projections to player-game results. The scored period covered Weeks 6 through 18:

The last number is the one that matters when judging confidence. A quarterback going over 225 passing yards and over 250 passing yards creates two threshold outcomes, but they come from the same game. They are not two independent pieces of football evidence. Every threshold for one player, game, and statistic stayed together in the same chronological fold.

The primary measure was player-start-balanced Brier score. Lower is better. It measures the squared error between the probability assigned before the game and what actually happened.

The Overall Results

The 50/50 blend produced the best overall Brier score. It lowered the error by about four percent compared with either individual model.

The learned blend finished almost dead even with the simple 50/50 approach. To me, that is an important result. Complicated does not automatically mean better. With only 13 independent weekly slates, a simple blend can be more stable than a weight that tries to learn too much from a small sample.

The learned blend improved on the standalone football probabilities by 0.0056 Brier points. A weekly-slate bootstrap put the 95% interval between -0.0069 and -0.0043. That is useful evidence that the combination helped in this test, but it remains one retrospective season.

The Best Answer Changed by Statistic

The overall winner was the blend. The family results tell the more useful football story.

The football model's passing-yard projections were clearly better than the Quant model. The point-projection results told the same story: the football model's passing-yard mean missed by 56.00 yards on average, compared with 67.15 for Quant.

That result fits the football. Quarterback passing volume is tied closely to team structure. A projection builder can account for a new offensive coordinator, the expected run/pass mix, injuries at wide receiver, the strength of the offensive line, and whether the defense is likely to force the offense to keep throwing. A historical distribution alone can be slow to recognize those changes.

Passing touchdowns were different. The football point projection remained better—0.84 mean absolute error versus 0.95—but the 50/50 blend produced better threshold probabilities. Touchdowns are volatile. The Quant distribution helped describe the uncertainty around the mean, while the football projection helped anchor the expected offensive opportunity.

Receiving yards gave Quant a slight win on point accuracy, 19.18 yards versus 19.61, while the blend delivered the best probabilities. Receivers live in a more complicated weekly environment. Targets, routes, quarterback play, game flow, coverage, and competition for catches all interact. The two approaches were seeing different parts of the same problem.

What We Would Use in 2026

We would not force one model across every statistic.

For passing yards, the football projection should remain the primary starting point. For passing touchdowns and receiving yards, the historical test supports using a blend. The weights should stay conservative until the 2026 forward sample grows.

Our working approach is:

1. Build the team and player opportunity projections first.

2. Use the Quant distribution to translate those expectations into threshold probabilities.

3. Apply the family-specific model choice established by the historical test.

4. Freeze every weekly projection before the games.

5. Grade every result, including the misses.

6. Revisit the blend only on a scheduled, forward-looking update—not after one bad Sunday.

The value of the blend is not that one model is weak. It is that the two models make different mistakes. The football model tends to react faster to football changes. Quant is often better at describing the range of outcomes around the expected result. When those strengths are genuinely different, combining them can make the final probability more stable.

What This Test Does Not Prove

This was a forecast-quality test using synthetic thresholds. The workspace did not contain the complete historical NFL market snapshots needed to reproduce the exact lines and prices that were available before every game.

That means the research does not establish financial performance or whether a particular sportsbook or prediction-market price was wrong. It tells us which forecasting approach produced better probabilities on the tested outcomes.

In the end, the result made the process more credible. Passing yards showed that football-based opportunity work can beat a statistical baseline. Passing touchdowns and receiving yards showed that the same projection can improve when a distribution model handles the uncertainty.

The best 2026 model should not be football-only or Quant-only. It should use the best part of each, by statistic, and keep proving the decision one week at a time.

Research note: This article describes a retrospective 2025 forecast test. The thresholds were synthetic, and the result is separate from Sportshack's live forward record. Analytics only. No bet placement.