Edge Labs EDGE LABS Market intelligence
CFB NFL
← The slate
CFB · Week 2 · Sat, Sep 12 · 3:30 PM ET · Memorial Stadium (Champaign, IL) · FINAL Duke 31, Illinois 27

Duke at Illinois. Our picks: Duke +5.5 (medium). Under 51.5 (medium).

The picks come first on this page. The market's number, our number and the gap between them sit under them as MODEL CONTEXT: that is the input our analyst panel argued from, not a call of its own. Tiers are how decisively the evidence agreed, never a win probability.

WIN We had Duke +5.5. It cashed.
ATS pick
Duke +5.5
medium WIN
locked Thu 1:05pm ET
locked +5.5 · closed +6
Duke 31, Illinois 27
Total pick
Under 51.5
medium LOSS
panel 3-0

Panel reasoning

On the spread

Duke at Illinois, Illinois -5.5. The number opened as high as Illinois -8.5 in preseason and has closed steadily toward -5.5/-6 across five books, even with public tickets running near 69 percent on Illinois, a reverse-line-movement signal toward Duke. Our model has Illinois -5.8 against a -5.8 market, a zero gap the model reads as no edge, effectively a pass. The injury column tips it: Illinois is missing its only returning offensive line starter, Brandon Henderson, plus WR Jayshon Platt and four tight ends lost for the season, while Duke's only real absence is a backup tackle. Duke also held Tulane to three points and leaned on the ground game, Nate Sheppard carried 35 times for 227 yards as the team ran 49 times to 28 passes, a downhill profile that worked against a leaky Illinois run defense in week one. The counter-case is real: Illinois scored 42 on UAB, gets extra rest, and hosts a sold-out home crowd. All three panels sided with the dog anyway. Take Duke +5.5 (FanDuel or Bovada).

On the total

Under 51.5. The total opened near 53.5 across the major books on 9/6 and fell to 51.5 everywhere by 9/10, a uniform two-point drop even as public tickets ran roughly 79-83 percent on the Over, a reverse-line-movement signal toward the Under. Weather is a non-factor: warm, mostly sunny, light wind, no rain. Our totals model sits at 53.48, a 2-point gap that is far inside its 12.86-point lean threshold, so it names no side. Duke's defense held Tulane to three points and Illinois just lost four tight ends for the season, thinning short-yardage looks, though Illinois' offense did score 42 on UAB. Under 51.5 (Pinnacle).

Model context

The numbers the panel argued from. Our models price a game, the panel makes the call, and the picks above are the call. Nothing in this block is a pick.

Locked line
Duke +5.5 · Illinois -5.5
the spread we locked at, it does not move
Our model number
Illinois by 5.8
the calibrated EL CFB rating number, home field included
Spread gap
0.2 points
the distance between our model number and the market. Context for the panel, not a call of its own.
Market total
51.5
consensus of 5 books at the model run, Sep 12, 12:38 PM ET
Our total
53.5
EL CFB Totals v1
Total gap
+2.0
the distance between our model total and the market total. Context for the panel, not a call of its own.

Context, not a call. The spread number is our rating difference plus home field. The totals number is a registered candidate that loses to the closing total: across 2,264 held out games it missed the combined score by 12.862 points on average and the closing total missed by 12.635, ahead of us in all 3 of those seasons, so it never posts to the board and never enters the record on its own.

Over and under splits

Descriptive context on this game, every line with the games it counts. None of it is a reason a pick wins and none of it enters a tier.

  • Early-season total of 51.5 fits the validated under window, unders 55% historically (n=943)
    total · under · moderate · n=943
    H64 (validated): Aug through mid-Sep games with a closing total of 47+ landed under 55% of 943 games, 2015-2025. Recent seasons: 2022: 58% (n=92) · 2023: 53% (n=81) · 2024: 57% (n=103) · 2025: 52% (n=119). 2025 ran soft at 52.1%, below the -110 breakeven, and the 2026 forward test is pre-committed on the early-totals page. Early-season totals are shaded up by public over-action; the market corrects as real form data arrives.

The panel

Panel A

Duke +6

Under 51.5

Take Duke +6 at BetMGM. EL's Illinois -5.8 provides essentially no numerical edge; my modest preference comes from matchup evidence. Illinois allowed UAB 194 rushing yards, while Duke's Nate Sheppard opened with 227. That suggests Duke can sustain possessions and keep this within one score. Illinois' 42 points also included a pick-six, a Hail Mary and a touchdown following a long punt return, an unusually favorable scoring mix to extrapolate. Brandon Henderson's expected absence adds an offensive-line concern missing from the dossier. The counterargument is substantial: Katin Houser started efficiently, Illinois has two additional recovery days and a sold-out home crowd, and Duke's passing game offered little threat. The move from Illinois -7/-8.5 toward -6 has already removed the best underdog prices. DraftKings' Illinois-heavy ticket share does not establish sharp action on Duke. With no meaningful model advantage and incomplete availability verification, this is a moderate lean, not a strong position.

Total. Under 51.5, narrowly. EL's 53.48 is only 1.98 points higher, far below its historical error. Duke's run-heavy opener and limited passing efficiency suggest clock-consuming possessions; Illinois' scoring benefited from events unlikely to recur together. Henderson's expected absence adds a small under adjustment. The over case is credible: Houser was efficient, Illinois struggled against the run, and Duke left points behind through failed fourth downs and a missed field goal. Warm, partly sunny weather supplies no persuasive under edge. Moreover, the dossier's decline from 53.5 means we are buying a worse number, and DraftKings' under-heavy handle may reflect information already priced. I favor fewer sustained scoring drives, but keep conviction modest rather than treating Duke's 17-3 opener as its true scoring level.

Where our model looks wrong

The spread is not demonstrably wrong, but final-margin inputs conceal Illinois' favorable scoring events and run-defense concerns. Henderson's expected absence is missing entirely. Conversely, treating all 13 Illinois injury rows as important losses would exaggerate the evidence. The totals estimate of 53.48 may overstate Duke's passing offense after the quarterback change; Eget's debut supports a more conservative expectation, although one game cannot establish a stable baseline.

Panel B

Duke +5.5

Under 51.5

Our number is Illinois -5.8 against a -5.8 consensus, so we have no gap to trade. More telling: Pinnacle opened -6 on 9/6 (after both week-1 games) and sat at -6 on 9/10, so the DraftKings -8.5 (9/2) and FanDuel -7.5 (9/3) opens were stale preseason numbers catching up, not steam. I take the dog on matchup and price. Illinois' new Bobby Hauck 3-3-5 allowed UAB 413 yards and 194 on the ground, and Bielema said tackling was a known issue after a light-contact camp (National Football Post preview, 9/9). Duke ran 49 times to 28 passes and Nate Sheppard leads FBS after week 1 with 227 yards on 35 carries (goduke.com, 9/5) - the downhill power profile that scheme leaks. Illinois plays without starting RG Brandon Henderson (lower leg, 2-6 weeks) and WR Jayshon Platt (Bielema presser, 9/8) against a front whose Bryce Davis had 3 TFL and 2 sacks. A run-heavy dog shortening the game is a cover profile, and +5.5 at FanDuel/Bovada buys the half-point through 6.

Total. Our model says 53.5, the market says 51.5, and the market got there honestly: every book we track opened 53.5 on 9/6, after both week-1 games, and all five sat at 51.5 by 9/10 - a genuine two-point drop inside the sharp window, including at Pinnacle. Weather does not supply the under; NWS Champaign (issued 11:36am CDT 9/10) has Saturday mostly sunny, high near 85, southwest wind about 6 mph, no rain, and ESPN's game page listed 87F. The case is possession count. Duke ran 49 times to 28 passes, opened the Tulane game with a drive that ate nearly 10 minutes, and scored 17 while allowing 3. Illinois' 42 came on 13-of-17 passing that will not repeat against a Manny Diaz front. I project Illinois 28, Duke 20 - 48, under the number.

Where our model looks wrong

The rating lands on the market number by coincidence, not process. It has one game of in-season data on top of a preseason prior, so it cannot see that Duke's 2025 offense (34.6 ppg with Darian Mensah, now at Miami) has been replaced by Walker Eget throwing for 131 yards on 26 attempts - Duke's 2.639 in-season rating is being carried by a 17-3 defensive result against a Tulane team that beat itself with penalties, which flatters Duke's true offensive level. In the other direction it reads none of the injuries: Illinois losing a starting right guard and a starting receiver, plus a defense in its first year of a new 3-3-5 with self-declared tackling problems, are all invisible to a rating that learns only from final margins. The flat 2.5 home field also looks light for a Big Ten home opener against an ACC visitor traveling to Champaign. Two large errors of opposite sign cancelling to a zero gap is not the same as a confident 'pass'.

Panel C

Duke +6

Under 51.5

Our model and the market agree almost exactly (our number -5.8 home vs market -5.8 at capture), so the model itself calls this a pass, it has no proven ATS edge in held-out testing anyway. The market has since drifted further toward Duke: several books opened Illinois -7.5/-8.5 and now sit at -6/-5.5, a real 2-3 point move, while public tickets still lean roughly 69% Illinois, textbook reverse line movement suggesting sharp money on Duke. That movement lines up with the news: Illinois lost its only returning offensive-line starter, RG Brandon Henderson, for this week (2-4+ week recovery per Bielema, 9/10), on top of four tight ends already out for the season and a cluster of questionable skill-position players. Duke's only real absence is a backup tackle. Duke's defense also looks the part of the number it posted, 3 points allowed to Tulane, while Illinois's D gave up 23 to UAB. None of this is overwhelming, and Illinois still has the better week-1 offensive output and a home crowd, so I'm not going bigger than a lean. Take Duke +6.

Total. All five books we track have the total pinned at 51.5, down from an open near 53.5-56.5 (DraftKings had no open captured; FanDuel/BetMGM/Pinnacle opened 53.5). That's a big, uniform drop, and it's happening while public tickets run heavily Over (79-83% per VegasInsider/SBD, 9/9), another reverse-line-movement tell, this time toward the Under. The results support it: Duke's defense held Tulane to 3 points in week 1, and Illinois just lost four tight ends for the season, thinning its red-zone and short-yardage looks. Illinois's own offense torched UAB for 42, so this isn't a lock, and our totals model still reads a touch higher (53.48) than the market, but that gap (2.0) is nowhere near its own 12.86-point lean threshold. Weather is a non-factor: warm, dry, light wind. Under 51.5.

Where our model looks wrong

The spread model doesn't see personnel: it has no read on Illinois losing its only returning O-line starter this week, its four season-ending TE losses, or the cluster of Illinois skill-position question marks, nor Duke's clean defensive showing in week 1. It's also explicitly non-predictive against the spread in its own holdout testing (46-50% in every gap bucket), so treating its 0-point gap as neutral rather than a signal is correct here, the real edge is the injury news and reverse line movement, neither of which the model can see at all.

The write-up

The case

Duke at Illinois, Illinois -5.5. The number opened as high as Illinois -8.5 in preseason and has closed steadily toward -5.5/-6 across five books, even with public tickets running near 69 percent on Illinois, a reverse-line-movement signal toward Duke. Our model has Illinois -5.8 against a -5.8 market, a zero gap the model reads as no edge, effectively a pass. The injury column tips it: Illinois is missing its only returning offensive line starter, Brandon Henderson, plus WR Jayshon Platt and four tight ends lost for the season, while Duke's only real absence is a backup tackle. Duke also held Tulane to three points and leaned on the ground game, Nate Sheppard carried 35 times for 227 yards as the team ran 49 times to 28 passes, a downhill profile that worked against a leaky Illinois run defense in week one. The counter-case is real: Illinois scored 42 on UAB, gets extra rest, and hosts a sold-out home crowd. All three panels sided with the dog anyway. Take Duke +5.5 (FanDuel or Bovada).

Confidence: medium

Three independent analyst panels (A, B and C) worked the same evidence pack and the week's news independently, each returned a side, and each then ranked its five strongest picks for the week. Every panel was on this side. Medium is a decisiveness label for how firmly the panel agreed, never a win probability: high means every panel was on this side and ranked it among the week's best; only high picks count on the record and go free.

The total: Under 51.5 (medium)

Under 51.5. The total opened near 53.5 across the major books on 9/6 and fell to 51.5 everywhere by 9/10, a uniform two-point drop even as public tickets ran roughly 79-83 percent on the Over, a reverse-line-movement signal toward the Under. Weather is a non-factor: warm, mostly sunny, light wind, no rain. Our totals model sits at 53.48, a 2-point gap that is far inside its 12.86-point lean threshold, so it names no side. Duke's defense held Tulane to three points and Illinois just lost four tight ends for the season, thinning short-yardage looks, though Illinois' offense did score 42 on UAB. Under 51.5 (Pinnacle).

Our number

edgelabs EL rating (preseason v1 until a team has played, then the in-season v1 Elo-style refit, Sun 23:00 and Tue 02:00 PT) run through the 2026-09-07 CFB calibration map; home field 2.5. Home perspective: negative = home favored. Ours: Illinois -5.8; market at synthesis: -5.8. Held-out 2024-2025 reconstructions (9,927 FBS games): 46-50% against the spread in every gap bucket, calibrated margin error 12.7 points. No against-the-spread edge at any gap size. Treat the number as context, not as a priced edge.

written before kickoff, frozen at kickoff

Trends

Descriptive context, every line with the games it counts. Filters like these did not hold up as predictors in our testing, so none of this is in the confidence read and none of it is a reason a pick wins. Situational splits look back 3 seasons and need at least 8 decided games to be shown at all. Records against the number and on the total join these once our closing-line history is restored; the straight-up splits are live now.

Duke
  • won 5 of their last 5 (n=5, 2025 to 2026)
  • won 5 straight
  • 6-2 straight up in road day games (n=8, 2024 to 2025)
Illinois
  • won 4 of their last 5 (n=5, 2025 to 2026)
  • won 3 straight
  • 13-2 straight up at home (n=15, 2024 to 2026)
The rest of the splits we can compute on this game
Duke
  • won 7 of their last 10 (n=10, 2025 to 2026)
  • 6-2 straight up in road divisional games (n=8, 2024 to 2025)
  • 8-3 straight up in day divisional games (n=11, 2024 to 2025)
  • 9-4 straight up at home (n=13, 2024 to 2026)
  • 11-5 straight up in day games (n=16, 2024 to 2026)
  • 11-5 straight up in divisional games (n=16, 2024 to 2025)
  • 8-4 straight up in night games (n=12, 2024 to 2025)
  • 6-3 straight up after a straight up loss (n=9, 2024 to 2025)
  • 7-4 straight up on the road (n=11, 2024 to 2025)
  • 5-3 straight up in home divisional games (n=8, 2024 to 2025)
Illinois
  • won 7 of their last 10 (n=10, 2025 to 2026)
  • 7-2 straight up in home day games (n=9, 2024 to 2025)
  • 7-2 straight up in home divisional games (n=9, 2024 to 2025)
  • 13-4 straight up in day games (n=17, 2024 to 2025)
  • 6-2 straight up after a straight up loss (n=8, 2024 to 2025)
  • 7-3 straight up in night games (n=10, 2024 to 2026)
  • 9-4 straight up in day divisional games (n=13, 2024 to 2025)
  • 3-5 straight up in road divisional games (n=8, 2024 to 2025)
  • 11-7 straight up in divisional games (n=18, 2024 to 2025)
  • 4-5 straight up on the road (n=9, 2024 to 2025)

The signals and the work

Members

3 independent signals on this game

The signal tally per side, the totals read, every angle with its sample size, and the line movement from open to now. 3 picks are released free on the slate every week. $24/mo, founder rate.

The two numbers underneath

TeamEL ratingConference
Duke 2.639 ACC
Illinois 5.458 Big Ten

Rating difference plus 2.5 points of home field is the rating read above. Full board: the power ratings.

Conditions at kickoff

Partially cloudy. 86F · wind 7 mph

Provenance: schedule and finals from the Edge Labs database (2026 season); the line is the latest capture for this game with the book named; our number is the EL rating read (in-season once a team has played, preseason before that) plus 2.5 home field, none at neutral sites, and it is context on this page rather than the call. Signals come from the trend engine, each with its real sample; a tier is how decisively the evidence agreed and is never a win probability. Once the game kicks off this page stops computing and reads our pick lock ledger instead: the pick and the spread we were locked at, written once at kickoff, never updated, graded against that same number. Line movement comes from our permanent capture log, one book across both ends and never a capture taken at or after kickoff, so the second number is the last pregame line and after kickoff it is the close; a game the log has captured only once shows no movement rather than an invented one. Weather is the captured forecast for the venue, and the chip appears only when it is worth saying (wind at 12 mph or more, a 50 percent or better chance of rain, or 35F or colder), never on an indoor venue. Injury counts are our latest daily scan, real report rows only, and a team the scan does not cover is left out rather than shown as zero. The write-up is assembled from those same stored rows, never written around them: the case comes from the lock row, the signals from the trend engine with their own samples, the series from our game database from 2013 forward, and the scoring profiles from completed games only, each with the number of games it averages. A section with no data behind it is left out instead of filled in, it refreshes while the game is pregame, and it can never be edited once the game has kicked off. Trends are descriptive only, computed by the same shared module the write-up uses so the two cannot disagree: straight-up splits come from completed games in our database, situational splits look back three seasons and need at least eight decided games, records against the number and on the total arrive with the closing-line restore, and every line carries the games it counts. Filters like these did not hold up as predictors in our testing, so none of them enter the confidence read. The totals block is a read and not a pick: our number is EL CFB Totals v1 (edgelabs.el_cfb_totals, season 2026, model_version v1, docs/el-cfb-totals-v1.md), a registered candidate that loses to the closing total, so it never posts to the board or the record; the market total beside it is the same capture the spread comes from before kickoff and the frozen consensus the model priced against after it, and the gap is measured against whichever number is printed. The totals trends under it are the trend engine's own rows for this game, deduped on the headline, each carrying its sample. On the games the analyst panel works each week (the games our model prices off the market plus the ranked and Power Four games, at most twenty), the pick and the total call are the panel's synthesis: three independent analyst panels, the same evidence pack, their own research, one side each, tiered by how firmly they agreed, never a win probability. A panel majority on a side is what makes a pick; without one the game is a published no-pick, printed as No pick rather than left blank. Research and context, never a guarantee, 21+.