← All projects

October 19, 2026

Pricing NFL games at halftime

How I built a halftime moneyline model for the NFL, why it needs two models to agree before it bets, and how I tried hard to prove it wrong before trusting it.

NFL / Live betting / Prediction markets / Python
Live trackerNFL Halftime MoneylineFull tracker →

Record

14-8

Units

+0.42u

ROI

+1.9%

Halftime is the strangest few minutes in sports betting. The game stops, the market re-prices the whole second half at once, and then it waits. This model makes its own price for the full-game moneyline in that window and only bets when it's confident the market has it wrong.

The idea

Pregame lines get hammered by the sharpest money in the world. Halftime prices get set fast, on thinner markets, by people reacting to a scoreboard. A first half carries a lot of information beyond the score, and a market in a hurry doesn’t always weigh it correctly. That’s the edge being tested here.

Architecture

  1. InputHalftime state + first-half performancePlus what the market expected pregame
  2. Model ASecond-half score simulationA full distribution, not a point
  3. Model BWin-probability modelBuilt independently
  4. PriceBoth vs. the live marketAfter fees and spread
  5. DecideBet only on agreementMost games are passes
01

Game state

Everything is snapshotted at the end of the first half: the score, who gets the ball, what the market expected before kickoff, and how each team actually played. Nothing from the second half can leak in.

02

Score simulation

The first model projects how many points each team scores the rest of the way and simulates the second half many thousands of times, including overtime, to produce a full-game win probability.

03

Win probability

The second model attacks the same question from a different angle, with a different structure and a different view of the game. It isn't a copy of the first one with tweaks.

04

Market price

Both are compared with the live prediction-market price at halftime. Fees and the bid-ask spread are charged before anything counts as an edge.

The two methods, in plain English

Both models are regressions: they learn a formula from thousands of past halftimes that turns what you can see at the break into a prediction. They predict different things, which is what makes them independent.

A

Ridge regression → points

The score simulator starts with ridge regression, which predicts a number: how many points each team will score in the second half. It's ordinary linear regression (each input gets a weight, and the weighted sum is the prediction) with one important twist.

B

Logistic regression → win probability

The second model uses logistic regression, which predicts a probability: how likely the home team is to win. Instead of a raw number, it squeezes its weighted sum into a value between 0% and 100%.

Why “ridge”?

Football stats overlap. A team that’s moving the ball well also tends to be scoring, converting third downs and winning field position, so many inputs tell the same story. Plain regression handles that badly: it can give huge positive weight to one stat and huge negative weight to its near-twin, fitting the past perfectly and the future terribly. That’s overfitting.

Ridge adds a penalty on large weights. The model is rewarded for fitting the data, but charged for every weight it pushes far from zero, so it spreads credit across related stats instead of betting everything on one. How strong the penalty is gets chosen by testing on seasons the model didn’t train on. The result is a steadier formula that holds up on new games, at the cost of fitting the past slightly less tightly. That’s the trade you want.

Ridge only gives an average, though: the Steelers’ expected second-half points. Games don’t land on the average, so the simulator then plays the second half out thousands of times around those expectations, using how real games have actually unfolded from similar spots, and counts how often each team ends up ahead. That count is the win probability.

Why “logistic”?

A win probability has to stay between 0 and 1, and it doesn’t move in a straight line. Going from down 3 to tied matters a lot; going from up 24 to up 27 barely matters, because the game is already decided.

Logistic regression builds that in. It computes a weighted score, then passes it through an S-shaped curve that flattens near 0% and 100%. In the middle, where games are close, each extra point moves the probability a lot; at the extremes it moves it very little. Under the hood the model works in log-odds, where a probability of 50% sits at zero and every step of evidence adds or subtracts. That’s why it’s a natural fit for “who wins”: evidence stacks up, and the curve handles the rest.

Same data, two different questions (how many points vs. who wins) and two different shapes (a straight-line average plus simulation vs. an S-curve). When two methods that different land on the same side of the market, that agreement means something.

Why two models

My first version found plenty of edges. Too many.

On its own, the score simulator claimed double-digit edges on bet after bet. When I measured it honestly, it was only about as accurate as the market, so most of those edges couldn't be real.

The fix wasn't a fancier model. It was a second opinion. When the two models agree, the bets held up. When they disagree, the first model's call was roughly a coin flip. So the rule is simple: no agreement, no bet.

Both models agree

Two independent methods see value on the same side, after costs. That's a play.

The models disagree

Pass. Disagreement usually means one of them is wrong, and I don't know which.

How it was verified

Most of the work went into trying to break it. Sports backtests lie easily, so every step was set up so the results could be wrong in public.

  • Walk-forward onlyEvery season is predicted by models trained only on earlier seasons. The model never sees the games it's graded on.
  • Untouched holdoutThe first version was tested once on 2026 games it had never seen. It came out positive but inconclusive, and I said so instead of rounding it up.
  • Rules frozen in writingThe final agreement rule was written down and dated before its official results were computed. Anything decided after looking is labelled exploratory.
  • Leakage auditEvery feature was checked to come from the first half or before kickoff, so nothing from the second half can sneak in.
  • Real costsResults are scored at the price you could actually take, after the spread and the exchange's fees, with a check that the visible size could fill the bet.
  • Accuracy, not just profitEach model is scored on how well it predicts every game, compared with the market itself. A model that wins money but predicts worse than the market is getting lucky.

Results so far

Halftime samples are small: about eight bets on a busy Sunday. A real edge here takes a season or more to separate from luck, so the forward record will be judged on accuracy against the market and on how prices move after each bet, not just wins and losses.

What’s next