NewLaunch offer: 50% off for life with code SHARP50. Claim it

Research

The college football halftime model.

Every halftime call in college football comes from one model that bets only its most confident spots. Here's how it was tested, what it found, and what to expect from it.

Precision Sports Mathematics · Nov 14, 2026

The short version

  • It only bets the top 5% of halftime spots it scores, about 20 to 28 a season.
  • On seasons it never saw, those spots won 61.1% in 2007–2014 and about 73% from 2024 on. Expect closer to 60%.
  • Most are Unders in blowouts, where the 2nd-half line still expects too many points.
  • The edge lives only at the extreme. Widen it to the top 10% and it falls to 52.6%, below break-even.
  • Buying points doesn't create the edge. It only helps a pick that's already good, at a fair price.

These are backtest results: hypothetical, not results of actual bets. The live record is on the Results page.

The question we asked

The owner's brief was simple: be right more often than wrong. Older halftime rules, like "bet the Under when the book's new number moved 13+ points," stopped working on recent seasons (see what didn't work). So instead of one trigger, we tried a model that weighs everything known at the half and then bets only when it's very sure.

The model sees the pregame total and spread, 1st-half points, the lead, whether the favorite is the team ahead, how far the book moved, and the 2nd-half line itself. The exact weights and cutoffs stay private.

How it was tested

  • Written down first. The method, the seasons and the pass mark were committed on Oct 19, 2026, before any result was looked at.
  • Fit on one set of seasons (2015–2022), then scored once on two sets it never saw: 2007–2014, and 2024 onward.
  • Graded against real lines, both the market consensus and DraftKings' 2nd-half number, the book the app reads live.
  • Checked for data errors. Rows with impossible 2nd-half lines were removed with the app's own filter. The result barely moved (59.7%, 72.7%, 73.4%).

What it found

Unseen 2007–201461.1%

226 bets, about 28 a season. Range 54.6–67.2% (95%).

2024 on, consensus lines73.1%

67 bets, about 22 a season. Range 61.5–82.3%.

2024 on, DraftKings lines73.8%

65 bets. Range 62.0–83.0%.

All unseen seasons63.8%

293 bets. About +19% return at −115, the price books charge at the half.

It was the first strategy in four rounds of research to clear the full bar we set in advance.

What to expect

  • Closer to 60% than 73%. The 2024+ number rests on 67 bets. The older 226 bets are the more sober estimate.
  • Lumpy seasons. Four of 19 seasons went 37–48%: 2008, 2014, 2018 and 2020. A bad month is expected, not a sign it broke.
  • Few bets. About 1–3 a week during the season. Some Saturdays have none.
  • Mostly Unders in blowouts. The average halftime lead on these spots was about 23 points. Starters sit, the clock runs, and the 2nd-half number often stays too high.

A real example: Wake Forest led California 34–0 at the half. The 2nd-half line was 26.5, the model called Under, and the 2nd half finished with 14 points.

Where buying points fits

Buying points raises any bet's hit rate, and books price that in. So it only helps when the pick is already good. On these spots from 2024 on, buying 3 points won 75.0% (a random college football Under with 3 points bought won 61.2%), and buying 6 won 88.2%. Whether that's worth it depends on what your book charges: 3 points is only worth paying up to −300 here. No source keeps old prices for bought 2nd-half points, so the app's buy points calculator checks the price you're quoted instead of guessing.

Related research

See every live call on the Results page, or how the whole system works on How it works.

Backtest figures are hypothetical. They are not results of actual bets, and past results don't predict future results. No call is guaranteed. HalfSharp is an analytics and information service for adults 21+; we don't accept wagers. Gambling problem? Call 1-800-GAMBLER.