# College football pick research

September 8, 2026 · 25-game card · SP+ comparison and market evidence

This report evaluates the original card against the predictive SP+ table supplied in this conversation. It also reviews reported betting action and the local pick engine. The spread calculations use the original quoted numbers, which have not been verified as currently executable sportsbook prices.

**Fourteen picks retain a positive rating-based edge; eleven change direction.** The largest positive discrepancies are Buffalo, Jacksonville State, UTSA, UNLV, and Kennesaw State. The strongest numerical objections concern Auburn, Clemson, East Carolina, and Ole Miss.

These findings identify research priorities. They do not establish a profitable strategy or a consensus among professional bettors. A ten-point model disagreement is a reason to investigate the inputs before increasing confidence.

## What the comparison measures

The predictive table contains 138 teams, with separate overall, offensive, defensive, and special-teams ratings. It was supplied by the user from [Bill Connelly’s ESPN SP+ article](https://www.espn.com/college-football/story/_/id/49868647/2026-college-football-sp+-rankings-all-138-fbs-teams). The arithmetic was verified against the supplied text. The article’s complete table was not independently retrieved.

The earlier table headed “Résumé SP+” is excluded from spread calculations. It describes performance against the schedule already played. Its values should not be substituted for predictive team ratings.

The calculation follows the project’s 2.5-point home-field assumption:

```text
Expected home margin = home rating − away rating + 2.5
Edge for the original pick = expected margin for that team + quoted spread
```

For Ohio State at Texas, the calculation is 25.4 − 30.0 + 2.5 = −2.1. That makes Ohio State a 2.1-point favorite in this estimate. Receiving 1.8 points creates a 3.9-point difference from the estimated fair spread.

The estimate assumes each matchup takes place at the named home team. It does not add independent injury, weather, travel, or matchup adjustments. It is not an official Connelly game projection. Those adjustments could matter, and some may already be reflected in SP+.

## All 25 picks

The fair spread below applies to the same team as the original pick. Positive edge supports that pick at the supplied number. Negative edge favors the other side. “Supports” describes direction, not the probability of covering.

| Original pick | Estimated fair spread | Edge for original pick | Direction |
|---|---:|---:|---|
| Buffalo +10.3 | +0.6 | +9.7 | Supports |
| Kennesaw State -9.5 | -17.8 | +8.3 | Supports |
| UNLV -3.5 | -13.0 | +9.5 | Supports |
| Marshall -13.5 | -8.9 | -4.6 | Opposes |
| Ohio State +1.8 | -2.1 | +3.9 | Supports |
| Western Kentucky +40.5 | +44.3 | -3.8 | Opposes |
| Michigan +5.5 | +7.0 | -1.5 | Opposes |
| Hawaii -8.3 | -10.5 | +2.2 | Supports |
| Jacksonville State +2.8 | -6.9 | +9.7 | Supports |
| Georgia Tech +12.8 | +15.7 | -2.9 | Opposes |
| Bowling Green +28.8 | +27.4 | +1.4 | Supports |
| California +3.3 | +8.6 | -5.3 | Opposes |
| East Carolina -7.5 | -1.6 | -5.9 | Opposes |
| Vanderbilt -20.8 | -22.2 | +1.4 | Supports |
| Fresno State -18.3 | -14.0 | -4.3 | Opposes |
| Louisiana Tech +35.3 | +29.6 | +5.7 | Supports |
| Rutgers +3.3 | -1.3 | +4.6 | Supports |
| Minnesota +0.3 | -2.7 | +3.0 | Supports |
| Air Force +2.5 | -1.0 | +3.5 | Supports |
| Oklahoma State +22.5 | +20.3 | +2.2 | Supports |
| Clemson -20.5 | -13.5 | -7.0 | Opposes |
| Auburn -33.5 | -24.4 | -9.1 | Opposes |
| UTSA +1.5 | -8.2 | +9.7 | Supports |
| Ole Miss -47.5 | -41.8 | -5.7 | Opposes |
| Missouri -5.5 | -2.9 | -2.6 | Opposes |

Source: the user-supplied predictive table and original card, calculated by [the audit script](../analysis/sp_plus_card_audit.py). The underlying [CSV](sp_plus_card_audit_20260908.csv) retains both team ratings, the signed home margin, and the change from the original displayed edge.

## Where the largest discrepancies come from

**Buffalo +10.3:** Buffalo’s rating is −11.6 and FIU’s is −13.5. The home adjustment makes FIU only a 0.6-point favorite. This is a large conflict with the supplied market number, but Buffalo’s offensive ranking of 125th makes sustained scoring an obvious question. Its defensive ranking of 80th is the stronger part of its profile. The next useful evidence is opponent-adjusted performance and a verified price, not another repetition of the same SP+ rating.

**UNLV −3.5:** UNLV is rated 15.5 points above North Texas before the home adjustment. North Texas’s defense ranks 123rd, versus UNLV’s offense at 44th. The rating estimate is UNLV −13.0. That favorable comparison coexists with contrary market movement, discussed below. It does not explain away that disagreement.

**Jacksonville State +2.8:** Jacksonville State’s −7.0 rating exceeds Ohio’s −16.4 by 9.4 points. Even after home field, the estimate favors Jacksonville State by 6.9. Ohio’s offense ranks 130th. Before treating this as a major opportunity, verify that “JSU” maps to Jacksonville State, that the line belongs to the correct game, and that both ratings reflect the same update.

**UTSA +1.5:** UTSA’s 2.9 rating exceeds Texas State’s −7.8. The estimate favors UTSA by 8.2. Texas State’s offense ranks 42nd but its defense ranks 132nd, so its scoring ability alone does not imply a balanced team. UTSA’s own defense ranks 93rd, a reason to avoid equating the overall edge with low game-to-game uncertainty.

**Kennesaw State −9.5:** The rating difference is 15.3 points, increasing to 17.8 at home. Georgia State’s defense ranks 137th. Kennesaw’s own offense ranks 94th, so the argument is primarily the gap between two teams, rather than an elite favorite. Current personnel and the quality of each opening opponent remain relevant.

These are explanations of the supplied model inputs. They are not independent confirmations of those inputs.

## Picks the updated calculation opposes

The reversals are substantial enough to change the research order. Auburn’s original favorable edge of 3.5 becomes −9.1. Marshall moves from +8.6 to −4.6. East Carolina moves from +5.1 to −5.9. Those changes compare two displayed calculations at the original lines; they are not evidence that bettors moved the market by those amounts.

Auburn’s rating advantage over Southern Miss, including home field, is 24.4 points. Laying 33.5 asks for much more. Clemson’s equivalent estimate is 13.5 against a quoted 20.5. Their offensive rankings of 76th and 66th help explain why these estimates fall short of the large spreads.

Michigan receives 5.5 in the original card, but the updated estimate requires seven. Oklahoma’s neutral-field rating advantage is 9.5. Home field narrows the gap without eliminating it.

East Carolina’s rating is slightly below App State’s. Home field makes ECU a modest 1.6-point favorite, leaving little numerical support for laying 7.5. California, Fresno State, Western Kentucky, Georgia Tech, Ole Miss, and Missouri also lose their original positive edge.

A reversal does not automatically authorize betting the opposite side. The opposing team’s actual spread and odds still determine the wager’s value.

## What can be said about sharp action

The evidence is uneven. A named sportsbook director describing respected wagers is stronger attribution than an odds tracker showing a price change. Neither represents every professional bettor.

| Matchup | Available evidence | Interpretation |
|---|---|---|
| Oklahoma at Michigan | Borgata director Tom Gable reports respected Oklahoma bets at −5.5. | Direct reported opposition to Michigan +5.5. |
| Ohio State at Texas | Circa initially moved Texas −2 to pick’em, then back toward Texas. Borgata subsequently described early side action as relatively balanced. | Early Ohio State interest, followed by two-sided action. No broad sharp consensus established. |

These bookmaker reports were checked on September 8. They identify the book and number involved, which makes them more useful than an unattributed claim that “Vegas likes” a team. [Patrick Everson, VegasInsider](https://www.vegasinsider.com/college-football/college-football-odds-week-2-2026/)

VSiN reports Air Force moving from approximately +6/+6.5 to +3/+3.5, supporting the original side at better earlier prices. It also reports UNLV shortening from −5.5 to −4.5 at some books, indicating interest in North Texas. East Carolina moved from −9.5 to −7.5, while Syracuse attracted support against California. These are reported market changes, not identified professional wagers. [Adam Burke, VSiN](https://vsin.com/college-football/week-2-college-football-betting-report-and-odds/)

For Buffalo, Burke points to FIU outgaining USF despite losing, while Buffalo struggled against Albany. That offers a football explanation for FIU’s market standing, but not proof of sharp FIU action. [VSiN](https://vsin.com/college-football/week-2-college-football-betting-report-and-odds/)

For the other games, this report does not establish confirmed professional support. Generic tracker labels and unattributed betting percentages are insufficient. Earlier exploratory comments about those games should remain provisional.

## Why following the move may not pay

The repository already contains a study of 4,179 matched games with opening and final archived spreads. Among 3,306 non-push bets following moves of at least half a point at the final number, 1,669 covered: **50.48%**. The same direction graded at the earlier opening number reached 54.94%, but selecting that direction required knowing the later line. That opening-price result is a hindsight diagnostic.

The practical distinction is price. A bettor who took Air Force +6.5 and a bettor who takes +2.5 own different wagers. A three-point loss wins the first and loses the second. Agreement with the earlier bettor’s team does not reproduce the earlier bettor’s opportunity.

The study itself uses retrospective snapshots and imperfect coverage. It cannot prove that every timely movement strategy fails. It does show why a simple “follow the sharp side” rule needs a timing-aware test. [Local line-drift study](../analysis/output/line_drift_study.md)

## What the model labels actually mean

The pick engine can fall back to SP+ projections while retaining historical bucket statistics from the in-season 2025 EPA-based model. Consequently, the bucket labels are not calibrated probabilities for these current SP+ picks. The saved payload contains:

| Absolute edge bucket | Historical sample | Recorded cover rate |
|---|---:|---:|
| 3 to under 7 points | 170 | 51.18% |
| 7 to under 14 points | 101 | 43.56% |
| 14 points or more | 9 | 77.78% |

The nine-game final bucket is especially weak evidence for a confidence label. More generally, moving a historical percentage from one forecasting method to another requires validation. The figures above were read from the saved payload; a fresh full backtest was not completed during this review. [Pick engine](../src/cfb_model_lab/picks.py), [saved payload](../public/api/picks.json)

At −110, a wager needs to win 110 / 210, or 52.38%, of decisions to break even before other costs. A point difference is not a win probability. Neither a 9.7-point disagreement nor a 51% historical label supplies a validated expected return for an individual game.

The code also takes the median spread across providers. With an even number of quotes, that median can lie between available spreads. For example, −10 and −10.5 produce −10.25. A one-decimal display can then show an unusual number. This explains how such numbers can arise, but the exact original snapshot is needed to prove the origin of each pasted line. [Spread aggregation](../src/cfb_model_lab/grade.py)

The local static picks file inspected during the review was a Week 1 snapshot dated August 15. It does not establish the exact inputs behind the user’s later card. The missing connection between the original prediction, its rating timestamp, and its odds snapshot remains an open audit item.

## The next research that would change a decision

The highest priority is to resolve the four roughly ten-point discrepancies: Buffalo, UNLV, Jacksonville State, and UTSA. Each needs a timestamped sportsbook quote, verified game identity, current player availability, and a check against a genuinely independent forecast. If the original pick already used SP+, another SP+ calculation is an update comparison, not a second vote.

Ohio State and Air Force deserve price-specific review because their rating support intersects with reported early market interest. Rutgers deserves a separate matchup review, but this report does not claim professional agreement. The eleven reversed picks need their original input snapshots explained before the original card is trusted.

CFBD can provide provider-specific lines, results, ratings, and advanced statistics. Its documented betting schema does not identify bettors or supply ticket and handle percentages. Historical coverage varies by field. A rigorous test must retain what was known before kickoff rather than substitute final-season ratings. [CFBD betting reference](https://api.collegefootballdata.com/api/betting), [coverage notes](https://api.collegefootballdata.com/data-availability)

For a future validation, record the model version, rating capture time, book, quoted spread, wager odds, and observation time together. Evaluate preseason SP+ picks separately from the in-season model. Report sample sizes, cover rates, realized returns, and performance by season. A strategy discovered after examining outcomes needs subsequent unseen games before its apparent advantage can be trusted.

## Evidence and reproducibility

The [comparison CSV](sp_plus_card_audit_20260908.csv) and [metadata](sp_plus_card_audit_20260908.json) preserve the calculations and source fingerprint. The [audit script](../analysis/sp_plus_card_audit.py) validates 138 distinct rankings and generates all 25 comparisons. This report’s table is generated from that CSV.

No live CFBD pull or authenticated sharp-money feed underlies this report. The numerical card uses the user’s supplied ratings and lines. Market observations are attributed reporting, checked separately. The report does not modify the production prediction logic, assign bet sizes, or claim a demonstrated profitable system.