MODEL PREVIEW · COLLEGE · NOT A PICK

Inside Rene Nunez's College Football Model: Honest Ratings, a Market Blend and a Momentum Idea That Didn't Survive

Numbers from Rene Nunez's college football model at renenunez.dev, as shown on Oct. 14, 2026 (Week 7 ratings and predictions; performance graded through Week 6), plus details Rene shared with us on Oct. 15, 2026.

Most model sites lead with how good they are. Rene Nunez leads with where his model falls short. His methodology page says plainly that his pregame model does not beat the closing spread, then shows exactly how he builds around that. We reached out to Rene about featuring his work, and he was glad to let us walk through it.

How the ratings work

The model rates all 266 Division I teams, FBS and FCS, on one scale: how many points better or worse a team is than an average FBS team on a neutral field. Zero means average.

It's built on possessions rebuilt from CollegeFootballData.com play-by-play, with garbage time thrown out. Each team gets an offense number and a defense number in points per possession, plus a shared home-field term that is refit every week from a starting point of 2.5 points. Because raw scores are noisy, the fit blends the scoreboard with an EPA-based "process" score. That's then averaged half and half with a simpler points-only rating that also covers FCS games without play-by-play.

From those ratings the model projects a spread and a total for every game, with a range around each. On the untouched 2023 to 2025 holdout (2,258 games), its 50%, 80% and 95% ranges caught 51.6%, 80.1% and 94.2% of results, very close to where they should be.

The market blend

Here's the part that stood out to us. Across 3,853 backtested games from 2021 to 2025, the model's own projections missed the final margin by 13.40 points on average. The closing line missed by 12.14. Rene's table shows the market winning every single season.

So instead of pretending otherwise, the site publishes a blended line: the model's number moved halfway toward the sportsbook market. The model's own numbers stay on the site labeled "Pure." Rene's view is that the model's value is coverage and calibration. It rates every D1 team, including FCS programs the market barely prices, and its ranges are honest.

The same thinking runs the in-game win probability. It starts from the published pregame line and leans more on the live scoreboard as the clock runs. In an offline test, starting from the closing line instead of the pure model improved accuracy, and Rene writes that this shows the pregame starting point matters most, not that his model beats the market.

Momentum: tested and rejected

The project is named after momentum, and Rene admits the section was "a little painful to write." He tested whether things like sustained stops, turnovers, field position, fourth downs, missed kicks, success rate and tempo carry signal the scoreboard hasn't caught up to yet.

Two versions were tried. Both helped slightly on the seasons used to build them and both got a little worse on the holdout. His read: whatever is real in a hot streak shows up on the scoreboard quickly, and the scoreboard is already in the model. Momentum work is paused. It's rare to see a creator publish a dead end like that, and it makes the rest of the page easier to trust.

How picks get graded

Picks come from a separate, versioned policy. The current one is v11, published Oct. 12 for Week 7. Every edge is measured after shrinking the model toward the market it's betting into. A pick needs the priced line to sit at least 2 points past the break-even line for that price, plus positive expected value. If fewer than 15 picks clear that bar in a week, the strongest remaining positive-value offers fill the slate and are labeled as exploratory.

Every pick records the line, price, book and time it was made, and settles at a flat one unit against those frozen numbers. Nothing gets re-priced after the result is known.

The 2026 record so far

Rene told us the number he cares about most right now is forecast accuracy, not market performance. Specifically, how far the published line misses the final margin, and how that compares with the closing line.

Through Week 6, over 748 graded games:

That's a gap of 0.62 points between the published line and the closing line, and the published line came closer than the market in 49% of games. In FBS vs. FBS games the gap shrinks to 0.11 points. In Weeks 3, 5 and 6, the model's average miss was lower than the market's. The historical backtest gap was 1.26 points, so the live season is running closer to the market than the backtest did.

There are still weak spots. Week 1 was rough (15.49 vs. 12.75), and FBS vs. FCS games are the hardest, with a 15.42-point miss against the market's 12.58.

Research pick record. The picks are a separate layer, and Rene frames them as research picks at one unit each. Through Week 6 they are 60-67-2, down 5.98 units at a flat one unit per pick (129 settled, 13 pending for Week 7). By market:

These are observed results, not a guarantee of future returns, and not betting advice. The full breakdown is in the Recommendations view on his performance page.

Week 7 ratings

Rene's Week 7 top 10, in points above an average FBS team:

The biggest gap between the model and the market is Alabama, which the model rates 6.6 points higher. Kentucky (+5.0), Utah (+4.9) and BYU (+4.3) follow. Rene notes this compares team strength, not a betting edge on any game.

Why Alabama? Rene explained that the 6.6 is a ratings comparison, not a game spread: 28.0 in the model against a market-implied 21.4. The market-implied rating is fit to this season's closing spreads, so it only moves as fast as the lines move. The model rating is fit on possession-level scoring, and Alabama has beaten the closing spread by about 12 points a game over six games, so the model gives credit for those margins faster. It doesn't translate into a pick this week. For the Tennessee game, the published line is Alabama -8.1 against a market of -8.5, so there is no spread pick there.

Week 7 research picks

These are the five Week 7 picks that cleared Rene's edge threshold, all at DraftKings prices recorded Oct. 12. They are research picks from a public model at one unit each, not betting advice, and they sit alongside the 60-67-2 (-5.98 units) record above.

• UAB vs. East Carolina, Over 47.5 (-110)
• Buffalo vs. Massachusetts, Over 41.5 (-110)
• Michigan vs. Penn State, Over 42.5 (-108)
• Houston vs. Oklahoma State, Under 61.5 (-110)
• Oregon vs. Nebraska, Over 52.5 (-110)

Eight more Week 7 picks are labeled exploratory on his site because they fall below the edge threshold, so we've left them out. All of them will be graded on his predictions page.

Beyond the spreads: the Heisman board and unit ratings

Rene pointed us to two parts of the site he finds most interesting. The unit ratings split each team into offense and defense, and then into rush and pass. Those unit ratings feed the Heisman board, a separate player layer. Every credited play is measured against what an average player produces against that opponent's rush or pass unit, pulled toward the average for the player's position, with a 30th-percentile replacement baseline subtracted. The board itself is a conditional logit model fit on 2010 through 2025 ballots, so the predicted vote shares add up to 100% the way a real ballot does.

Credit

All numbers here come from Rene Nunez's model at renenunez.dev, as shown on Oct. 14, 2026, plus details Rene shared with us on Oct. 15, and will change weekly. See the methodology page for the full build, the ratings for every D1 team, and the performance page for the live record.

These are model projections and research picks, not guarantees or betting advice. Rene's own numbers show how hard the closing line is to beat, and how close his published line is getting.

Thanks to Rene for letting us feature his work.

Back to news