Where our models and angles get tested. The picks come from the panel.
Every pick on the weekly boards comes from a three-seat analyst panel (panel v2.5 from the boards of Tuesday, September 29): three analysts work the same evidence pack for each game on their own, each names a side, and the pick locks at the number it publishes at and grades there, in public, on the record page. This page is the research track beside it. Our models price every game for that evidence pack, and angles are tested on five seasons of college football and then forward on the current season before we trust any of them.
The database
Five seasons of FBS games with real closing lines and graded results, plus roster and recruiting context. Every number we publish traces to a producing query against this database.
The backtest
Every angle is fit on older seasons and measured on a held-out season the model never saw. A frequency that only shows up in-sample is noise, and it stops here.
The live proof
Clearing history earns a forward test, not a bet. An angle is trusted only after it holds in a forward window on the current season, graded at the number we posted, misses included. An angle is never a pick on its own and never sets a tier.
proposed or under test in our college football research log, counted when this page loads; none of them is a pick
One angle cleared our historical gate. It failed its 2026 forward test: 47.3% under on n=131, against a 52.38% gate. It is off the board.
Early season unders on college football totals of 47 or higher went under 54.75% of the time in the backtest (n=884, 2018 to 2025). The rule was frozen before the season and graded in public through week 3: 62 under, 69 over, 2 push. It is in the testing log below with everything else that did not hold.
Read the frozen rule, the backtest and the final result →The number we keep · closing-line value
What is CLV, and why do we track it?
The forward metric we build around is closing-line value (CLV): did the market move toward a number after we flagged it? CLV proves an edge long before a win rate possibly could · a bettor who consistently beats the closing line is, by definition, finding value the market hadn't priced yet.
CLV belongs beside the research process, not in place of it. We pair it with transparent samples, clear assumptions, and re-testing as new CFB seasons land.
Every pick and every forward log appears graded at the number we posted, misses included, on the record page.
The testing log
Angles that came off the board, each one as what we tested and what the test came back with. Every row is a real record from the research database, kept in full and added to, never edited down. A validated angle that stops holding lands here too.
TestedCFB full-game totals in regular-season weeks 1-3 (games played in August through approximately September 14), for games where the closing total >= 47, produce actual combined scores UNDER the closing no-vig total at >=54% in 2015-2025 FBS regular season.
FoundIt cleared the historical gate and earned a live forward test, graded in public. The forward test came back at 47.3% on n=131, below the 52.38% gate it had to clear, so it came off the board.
The detail
Why we thought it might work. Early-season totals are shaded upward by peak-season public over-action (Paul & Weinbach 2007) while the information gap between preseason projections and actual Week 1-3 team quality prevents sufficient sharp correction before game time.
What the test showed. The pre-registered 2026 forward test (FBS regular season weeks 1 to 3, closing total >= 47, side UNDER, graded against the closing no-vig total) finished at 47.3% under on n=131 non push games (62 under, 69 over, 2 push), below the pre-registered gate of 52.38%. By week: week 1 went 19 under, 21 over, 1 push; week 2 went 24 under, 20 over; week 3 went 19 under, 28 over, 1 push. 133 games qualified, and 8 of 157 completed FBS games had no captured pre-kickoff closing line and were skipped, not graded.
Read the article that introduced this angle, with the final count →
TestedFBS-vs-FBS games where the EL CFB Totals v1 predicted total differs from the consensus closing total by at least 4 points cover the model's side of the total (over if the model is higher, under if lower) at better than the 52.38 percent -110 break-even.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
TestedFBS teams facing an opponent at least 10 prior-season SP+ points weaker, in the week immediately before a game against an opponent at least 10 prior-season SP+ points stronger than that current opponent, cover the closing no-vig spread at 46% or lower in the current game, 2013-2025 FBS regular season.
FoundThe test came back at 48.92%, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Train cover rate 48.92% (CI [47.06%, 50.74%]) does not clear pre-committed 46% gate; out of sample 53.9% moves opposite direction. No sandwich-game ATS edge.
TestedFBS teams with a new head coach in their first season at the program playing at home in regular-season weeks 1-2 cover the closing no-vig ATS at <=44% (n=227, observed 42.3%, 2021-2025 excl. 2020).
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. Scheme installation deficits in the first two live game exposures are large enough to neutralize the standard 2.5-3.0 point home-field-advantage spread credit that closing lines apply uniformly regardless of coaching tenure, minimum gate n>=100.
TestedCFB bowl game full-game totals close under the no-vig closing total at >=54% in postseason bowl games (2015-2025 excluding 2020) where the Bovada closing total is between 53 and 57.5 inclusive.
FoundThe held out seasons came back at 53.7% on n=41, short of what the angle needed.
The detail
Why we thought it might work. The moderate-total range concentrates public OVER action (bettors see an offensive showcase) without triggering the aggressive sharp UNDER correction that activates at very-high totals, leaving closing totals systematically inflated above fair value by the public-sentiment mechanism documented by Paul & Weinbach and confirmed operationally in the regular-season validation.
TestedFBS teams with prior-season SP+ >= 10 AND current-season returning production <= 25% cover the closing no-vig ATS at <= 44% in FBS regular-season weeks 1-6.
FoundThe sample came in under our size minimum and the cover rate never got down to the 44% the angle needed.
The detail
Why we thought it might work. Early-season lines are constructed entirely from prior-season SP+ with no in-season performance data yet available, maximizing the informational gap between the composite model's roster-quality signal and the market's backward-looking anchor, the edge should be front-loaded before weekly SP+ updates partially correct the mispricing.
TestedFBS teams with prior-season SP+ <= -5 AND current-season returning production >= 70% cover the closing no-vig ATS at >= 56% in FBS regular-season games.
FoundThe sample cleared our size minimum, but the cover rate never got up to the 56% the angle needed, so it came off the board.
The detail
Why we thought it might work. Closing lines anchor on the poor prior record while high roster retention signals that prior underperformance reflected scheme or experience deficits now largely resolved, the composite model captures the returning-production uplift that backward-looking SP+ anchoring misses.
TestedFBS teams with prior-season SP+ >= 10 AND current-season returning production <= 20% cover the closing no-vig ATS at <= 44% in FBS regular-season games (2022-2025, excl. 2020).
FoundThe sample cleared our size minimum, but the cover rate never got down to the 44% the angle needed, so it came off the board.
The detail
Why we thought it might work. Closing lines anchor on prior-season SP+ while near-total roster turnover signals unsustainable prior performance that the composite model correctly identifies as regression, the individual player production driving the high SP+ has departed, and the market has not adjusted the price.
TestedCFB FBS road teams in the first season of a new head coach (hired January through August of that season) cover the closing no-vig ATS in regular-season road games played in weeks 1-3 at <=44% (home team covers >=56%), 2015-2025 (excl. 2020).
FoundThe test came back at 46.8% on n=201, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Closing lines anchor on prior-year SP+ without a scheme-installation penalty, while new coaching staffs produce maximum communication errors under road crowd noise in their first competitive game exposure with players learning new assignments.
What the test showed. Holdout 46.8% (201 games, 2022-2025) fails <=44% gate. In-sample signal not replicated. Full 45.2% is close but below minimum standard for candidate.
TestedFBS road teams whose home campus is at elevation below 1,000 feet cover the closing no-vig ATS at <=44% in Mountain West regular-season home games at venues above 4,000 feet elevation (Air Force, Wyoming, Utah State, Colorado State, Nevada, Boise State; excl. Colorado), 2015-2025 (excl. 2020).
FoundThe sample never got big enough to trust. 61 qualifying games against the 80 we require, so it came off the board.
The detail
Why we thought it might work. Altitude physiology (Levine and Stray-Gundersen 1997) reduces aerobic capacity 5-10% in unacclimatized visitors while closing lines set on SP+ do not apply a geography-adjusted quality penalty and non-Colorado Mountain West altitude venues receive less market recognition.
What the test showed. n=61 below gate 80 for full period. Holdout 46.8% directionally correct but above 44% threshold. Altitude mechanism may exist but insufficient Bovada game coverage at these matchups.
TestedFBS regular-season G5 home teams receiving >= 10 points on the closing no-vig ATS in weeks 1-2 vs. P4 road opponents (2015-2025, excl. 2020) cover the spread at >=56%.
FoundThe sample never got big enough to trust. 13 qualifying games against the 80 we require, so it came off the board.
The detail
Why we thought it might work. Early-season public money maximally inflates the P4 road-favorite spread while G5 home openers carry peak opening-day crowd advantage and P4 programs integrate new transfers and system installations in their first live reps.
What the test showed. n=13 non push vs min gate 80. G5 vs P4 matchups in weeks 1-2 with 10+ spread rare in Bovada data. Direction wrong (46.2% home cover vs predicted 56%+).
TestedCFB: opening total >= 60 AND absolute spread >= 14 (lopsided favorite) lifts the high-total UNDER hit rate by about 5 points.
FoundThe test came back at 47.9% on n=557, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Out of sample gate failed. Pre-discovery teamrankings 2013-2018: 47.9% UNDER, log-loss 0.728 > baseline 0.693. 2019 out of sample positive (59-61%, both sources) but single-year and small. Season variance extreme (range 44-64% across 6 seasons).
TestedFBS home teams in regular-season games scheduled on Tuesday or Thursday cover the closing no-vig ATS at <=44%, 2015-2025 (excl. 2020).
FoundThe test came back at 48.3% on n=145, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Standard home-field spread credit (~2.5 to 3 points) is misapplied to mid-week games where reduced crowd attendance and compressed preparation time eliminate the home-environment advantages that justify the credit, while road teams benefit from simpler game plans suited to a 6-day turnaround.
What the test showed. Holdout 48.3% home cover fails <=44% gate. Away covers 51.7% in holdout, essentially random.
TestedCFB team-total UNDER bets when one team's implied score (game total minus spread, divided by 2) falls in the lowest quintile of the historical same-season distribution yield positive ROI net of vig.
FoundThe test came back at 50.1% on n=1,620, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Bookmakers price team totals using symmetric score distributions that systematically underestimate the probability mass below the line when scores approach the zero floor.
What the test showed. 50.1% under rate, completely flat. Censoring-bias mechanism not observable when using implied scores derived from the same closing lines being tested.
TestedCFB breakout-team fade: FBS teams that improved their prior-season conference win total by 4+ wins open the following season as home favorites of -10 or greater (weeks 1-6) and fail to cover the spread at rates above 55%.
FoundThe test came back at 47.1% on n=70, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Betting markets underestimate season-to-season mean-reversion in program strength.
What the test showed. Fade win rate 47.1%. In-sample opposite direction (home covers 62.5%). Mean-reversion mechanism not observable in this filter.
TestedFBS road teams from Pacific time zone schools playing at Eastern or Central time zone venues with kickoffs at or before 12:30 PM ET cover the closing no-vig ATS at <=44%, 2015-2025 (excl. 2020).
FoundThe sample never got big enough to trust. 13 qualifying games against the 60 we require, so it came off the board.
The detail
Why we thought it might work. A 9:00 to 9:30 AM Pacific kickoff places players in active competition at their circadian low point for alertness and fine-motor performance, while sportsbooks set lines on SP+ quality metrics without applying a time-of-day or timezone adjustment factor.
What the test showed. n=13 below gate 60. Directionally correct (30.8%, predicted <=44%) but n too small. Pacific TZ schools rarely get early ET kickoffs; sample may never reach gate without expanded team list.
TestedCFB FBS teams whose prior-season final game was played on or after January 5 (CFP quarterfinal round or beyond) cover the closing no-vig ATS in their season Week 0 or Week 1 opener at <=44% (opponent covers >=56%), 2016-2025.
FoundThe sample never got big enough to trust. 17 qualifying games against the 40 we require, so it came off the board.
The detail
Why we thought it might work. CFP deep-run programs absorb maximum January-onwards roster disruption, accelerated portal departures, early NFL declarations, and staff poaching, while closing lines treat CFP pedigree as a quality-continuation positive rather than a roster-stability risk.
What the test showed. n=17 vs min gate 40. Strong directional signal (70.6% fail rate) in recent seasons but 2022 reversed. CFP expansion in 2024 changes qualifying-team pool.
TestedFBS teams playing their second consecutive regular-season road game (no home game between the two away games) cover the closing no-vig ATS at <=44% in that second road game, 2015-2025 (excl. 2020), regular season weeks 2-13.
FoundThe test came back at 50.1% on n=1,956, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Cumulative travel fatigue from hotel sleep disruption, unfamiliar practice facilities, and consecutive opponent-environment exposure is not priced into lines that apply a single-game road disadvantage adjustment independently for each game.
What the test showed. 50.1% away cover, flat across all windows. Cumulative road fatigue mechanism not priced-in vs closing line.
TestedFBS road teams playing their first conference regular-season road game of the season after at least two prior non-conference games cover the closing no-vig ATS at <=44% in that game, 2015-2025 (excl. 2020).
FoundThe test came back at 50.4% on n=752, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. The home conference opponent has accumulated 2+ games of live film on the road team while the road team faces conference-caliber defensive and offensive schemes for the first time that season, creating a preparation asymmetry that SP+-based line-setting does not capture.
What the test showed. 50.4% away cover, in-sample inverted. Preparation asymmetry mechanism not present in closing ATS data.
TestedIn FBS non-conference regular-season games (weeks 1-2, 2015-2025 excluding 2020), when the preseason SP+ rating differential diverges from the Bovada closing spread by >= 7 points in direction favoring one team, that team covers the closing no-vig spread at >= 62% of qualifying games.
FoundThe test came back at 45.8% on n=273, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Preseason SP+ ratings systematically incorporate team-quality signals that early-season betting lines underweight due to narrative anchoring and public-sentiment pricing.
What the test showed. Combined SP+-favored team cover rate 45.8%, below coin flip, far below claimed 62%. Strong directional asymmetry: home-direction 76.1% is confounded by P5-vs-cupcake scheduling (market conservatively undersets spread for unknown early opponents, SP+ correctly rates P5 much higher); away-direction 31.3% is inverse of the claim (market is right, SP+ overcredits away teams by ignoring HFA). Note: original claim included +2.5 HFA adjustment not applied here, and used Bovada not consensus. Even accounting for methodology difference, the combined rejection appears robust.
TestedIn FBS non-conference regular-season games (weeks 1-3, 2021-2025, excluding 2020) where the closing total is >=47 AND the preseason SP+ differential exceeds the closing spread by >=8 points in the same direction as the favorite, the UNDER covers the no-vig closing total at >=56%.
FoundThe test came back at 48.8% on n=41, short of what the angle needed, so it never reached the board.
The detail
What the test showed. SP+ divergence interaction adds no out of sample value. Interaction model log-loss 0.7010 > no-vig baseline 0.6931. out of sample under rate 48.8% misses 56% gate. Signal inconsistent in training (2021 and 2024 reversed).
TestedFBS regular-season home teams at outdoor venues >= 4,500 feet above sea level.
FoundThe test came back at 42.9%, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Aerobic capacity decreases ~2.4% per 1,000 feet above sea level for unacclimated visitors (Fulco et al. 1998), compressing fourth-quarter competitive output in ways the standard 2.5-3.0 point home-field-advantage adjustment does not capture.
What the test showed. Altitude physiology mechanism does not explain the ATS pattern. Wyoming, the highest-elevation FBS venue at 7,218 ft, covers only 42.9% overall and 42.9% in even-strength games, directly contradicting the hypothesis that aerobic VO2max impairment at altitude compresses visitor 4th-quarter output. All 8 venues excluding Colorado and Utah produce 48.5% ATS (breakeven). Colorado and Utah drive the aggregate effect for program-specific reasons unrelated to altitude: Colorado consistently plays as the weaker team vs above-average visitors yet still covers (pricing anomaly), while Utah is simply a better team.
TestedIn FBS non-conference regular-season games (weeks 1-3, 2022-2025, excluding 2020) where the home team is a Power 4 program [ACC, Big 12, Big Ten, SEC] and the closing spread is between -2.5 and -14 (competitive home favorite), the home P4 team covers the no-vig closing spread at >=56%.
FoundThe test came back at 45.8% on n=48, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Early-season home crowd intensity in first competitive non-conference home matchups exceeds the standard 2.5-3.0 point HFA adjustment applied uniformly by bookmakers across all game weeks.
What the test showed. Cover rate=45.8% (vs 56% gate, wrong direction). n=48 (vs 120 minimum). Hypothesis fails both the direction test and sample-size minimum. The crowd-intensity HFA premium mechanism is not detectable in competitive early-season non-conference spreads (-2.5 to -14). Conference heterogeneity (Big 12 vs ACC) confirms noise, not signal. Filter too narrow to reach required n even with 4 seasons of data.
TestedFBS regular-season road teams on their 2nd or more consecutive road game (both the current and immediately preceding games are designated away contests) cover the closing no-vig spread at <=46% in 2015-2025 regular season (excl. 2020).
FoundThe test came back at 50.14% on n=2,850, short of what the angle needed, so it never reached the board.
The detail
What the test showed. n=2850, cover rate=50.14%, z=0.150. Year-by-year oscillates 46.6-53.9% with no directional trend. Market prices consecutive road burden efficiently.
TestedFBS teams on 6-day rest (played prior Saturday, current game on Friday night) cover the closing no-vig spread at <= 47% against opponents on standard 7-day rest, FBS regular season 2015-2025.
FoundThe test came back at 51.16%, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Primary: cover rate=51.16%, z=+0.591 vs gate of <=47% / z<=-2.0. Direction wrong: 6-day rest Friday teams cover at 51.16%, not under 47%. No sub-population (home, away, conf, non-conf) clears both gates simultaneously with adequate sample size.
TestedIn FBS non-conference regular-season games (weeks 1-3, 2015-2025, excluding 2020) where the closing spread is >= 24 points and the underdog has a closing moneyline of >= +600 (implied win probability <= 14.3%): the underdog covers the closing no-vig ATS spread at >= 54% of qualifying games.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. Cognitive probability overweighting inflates the underdog's moneyline action, prompting bookmakers to shade the ATS spread wider than fair value to balance book, leaving the underdog's ATS price more favorable than the true talent gap warrants.
TestedFBS regular-season post-bye teams cover the closing no-vig spread at >=54% vs. non-bye opponents in 2015 to 2025 (n=2,115 games where exactly one team had a bye the prior week).
FoundThe test came back at 50.1% on n=957, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Books anchor on the most recent game result rather than schedule-adjusted team strength, systematically underpricing the preparation and rest advantage the bye team carries into the matchup.
What the test showed. Cover rate 50.1% (n=957 non push, real-book closing lines, 2018-2025). Gate requires 54%; fails by 3.9pp. Z=0.03, 95% CI [46.9%, 53.2%]. 2025 sharply reversed to 41.5%. 2019 also weak at 42.8%. 3/8 seasons above 54%, 3/8 below 50%. Glickman-Stern mechanism does not manifest in CFB ATS closing line markets 2018-2025.
TestedIn FBS regular-season games where one team played their prior game <= 5 days before the current game (short-rest team, typically a Thursday-game team returning Saturday) and the opposing team played their prior game >= 7 days before the current game.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. CFB closing lines anchor on prior-game results and talent differentials rather than schedule structure, leaving the 1.5-2 day rest disadvantage systematically underpriced. Minimum n=300 qualifying game-sides.
TestedCFB: when the opening total is 60 or higher, the UNDER beats the CLOSING total across multiple seasons; the edge is high-total-specific.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
TestedCFB full-game totals close under the no-vig closing total at >=54% in games where the closing spread is >=21 points, 2015-2025 regular season.
FoundThe test came back at 49.12%, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Game totals are set with an implicit garbage-time scoring premium that partially fails to materialize when the leading team's starter is pulled after halftime and both teams' scoring efficiency compresses in the fourth quarter.
What the test showed. Fails at all 3 spread thresholds (17, 21, 28pt) and all 3 tier controls (P4-vs-P4, G5-vs-G5, mixed) across 2015-2025 (11 seasons). Primary gate (>=21pt): 613/1,248 = 49.12%, 4.88pp below the 54% gate and below the 52.38% breakeven. Market is efficient at large spreads. Closing books correctly price the garbage-time scoring compression. The proposed mechanism (starter pulled, fourth-quarter efficiency collapse) is already reflected in closing totals.
TestedFBS regular-season games where a team won their prior game by >=28 points show opening-to-closing spread movement of >=1.5 points in that team's favor in >=58% of qualifying matchups, and backing the opponent at open (before public sentiment inflates the blowout winner's line) yields positive CLV of >=0.5 points vs. the closing spread.
FoundThe test came back at 27.48% on n=808, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Fails the primary gate: only 27.48% of qualifying blowout-winner games (>=28pt margin, n=808) show a >=1.5pt line move toward the blowout winner, far below the 58% gate. Year-by-year unstable: 2024 showed negative CLV (41.92% positive). Median CLV = 0.5 across subsamples is a CFB whole-to-half-point structural artifact, not signal. Data limited to 2021-2025 (Bovada opening spreads not available pre-2021).
TestedCFB full-game totals close under the no-vig closing total at >=54% in games where either team's prior game had a final margin >=28 points, 2015 to 2025 regular season.
FoundThe test came back at 50.38% on n=1,727, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Book-setters and public bettors anchor the next-game total on the extreme prior outcome, inflating it above the team's sustainable scoring rate.
What the test showed. 50.38% under rate (861/1709 excl. pushes, n=1727), misses 54% gate by 3.6 points. Year-by-year range 48.0%-51.7%, no season exceeds 52%. Market has fully priced blowout prior margins into totals. The blowout-anchoring bias described in Thaler-Ziemba and De Bondt-Thaler is not detectable in CFB totals at this threshold. Efficient market result consistent with prior rejections. Note: Bovada totals available from 2019 only (2015-2018 absent from DB); n=1727 is 8x the minimum gate so underpowering is not a concern.
TestedCFB totals on home P4 teams ranked AP Top 10 at game time produce under win rates >=54% vs. closing total in 2021 to 2025 regular season.
FoundThe test came back at 50.00% on n=190, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Books shade totals upward to capture public over-betting on marquee programs and the resulting inflated total is above sharp-priced fair value.
What the test showed. Fails on directional consistency: in-sample 2021-2023 shows exactly 50.00% under, 2022 reversed to 45.59% under (lines underpriced Top 10 home games that year). out of sample 2024-2025 combined shows 55.80% under, driven almost entirely by 2025 while 2024 is flat. Five-season aggregate: 52.44% under, z=0.88, p=0.38, no reliable signal. The Paul-Weinbach shading mechanism cannot be confirmed in this window.
TestedLate-season CFB home teams at outdoor Big Ten North, Mountain West, or MAC venues (game-date October 15 to December 15) cover the spread at >=56% vs. visiting teams from warm-weather conferences (Sun Belt, Conference USA, SEC West non-conference) when game-time temperature <=40°F, in 2013 to 2025 regular season.
FoundThe filter never produced a single qualifying game in eleven seasons of data, so there was nothing to test.
The detail
Why we thought it might work. The cold-weather home-field advantage is underpriced when the visitor has no cold-weather practice history.
What the test showed. CFB scheduling places non-conference games in September, so warm-weather visitors (Sun Belt/CUSA/SEC) never appear at Big Ten/MAC/MWC venues in the Oct-Dec cold window. Even extending data to 2013 would not change this: the scheduling constraint is structural, not a data gap.
TestedCFB composite power rating differentials (Sagarin + Massey optimal combination) predict opening-to-closing line movement direction: in FBS games where the composite implies a team >=4 points stronger than the opening spread favors, that team's line moves >=1 point in their direction by close in >=55% of cases.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
TestedCFB spread bets where the bettor's projection falls within 0.5 points of a key number (3 or 7) and the line is set on the favorable side of that key number.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. Historical CFB score differentials land on exactly 3 at 2.7x and on exactly 7 at 2.1x the rate predicted by a normal distribution.
TestedCFB spreads on G5 vs G5 games are softer than spreads on P4 vs P4 games.
FoundThe test came back at 51.97% on n=1,499, short of what the angle needed, so it never reached the board.
The detail
Why we thought it might work. Books spend less time grading the matchup.
What the test showed. G5vG5 underdogs cover 51.97% out of sample, which technically beats the 50% no-vig log-loss baseline by 0.00076. But: ROI at -110 juice is -0.41% (below 52.38% break-even); the log-loss gain is trivially small; season-by-season results are highly inconsistent, G5 dogs covered 42.60% in 2015 and 48.76% in 2025, below 50% in 4 of 11 seasons; the 2022 outlier drives most aggregate signal; the 2024 season saw P4 dogs cover MORE than G5 (53.85% vs 51.50%); the G5vG5 vs P4vP4 difference is z=1.11, not statistically significant.
TestedCFB team-total under bets on teams with implied score <=10 points (derived from game total and spread) beat their no-vig probability by >=3 percentage points.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. Books set team-total lines assuming unconstrained normal score distributions, ignoring the zero-floor censoring effect.
TestedWeather (wind, precipitation) is under-priced into CFB totals more than NFL totals.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
The detail
Why we thought it might work. CFB has more variance in venue-level weather data quality.
TestedNew-HC FBS teams priced as 7-14 point underdogs cover at below-market rates (40.3% ATS, n=278, 2013-2019); fading them, i.e., backing their opponents, yields positive EV in this spread bucket.
FoundThe test came back at 47.71% on n=327, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Out of sample validation fails. Opponent cover rate of 52.29% does not clear the -110 break-even threshold of 52.4%. Signal decayed significantly from 2013-2019 era (opp cover 59.74%) to 2020-2025. 2022 season reversed sharply (HC cover 59.0%, n=61).
TestedFading new-HC FBS teams priced as 7-14 point favorites in regular season games returns 59.4% ATS for the opponent (n=186, 2020-2025 regular season, FBS only).
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
TestedCFB FBS score differentials have sigma >= 15.0 points versus the NFL sigma = 13.86 established by Stern, causing the standard phi(spread/13.86) win-probability formula to overestimate favorite cover probability by >= 2 percentage points at spreads of 14+ points in FBS non-conference games.
FoundIt did not clear our standard on n=3,937, so it came off the board.
The detail
What the test showed. FBS-vs-FBS 2021-2025, n=3,937 games, consensus closing spread. WMCE of Phi(spread/sigma) vs realized favorite win rate: sigma=13.86 -> 1.943pp, sigma=15.30 -> 1.744pp. Reduction +0.199pp, robust in direction (15.30 better in all of q6/q10/q15/q20 quantile-bin schemes, +0.07 to +0.25pp) but an order of magnitude below the >=1.0pp gate. sigma=15.30 is the marginally better global constant, driven almost entirely by 20+ point spreads where 13.86 overestimates favorite win prob by 2-3pp; at 14-20pt spreads the two are within ~1pp and 13.86 is actually slightly better in the [13.5,17) bucket.
TestedVegas under-weights composite roster talent gap in non-marquee CFB matchups. When Team A composite is 0.05+ above Team B AND Vegas implies A as dog, A covers >55% historically.
FoundThe test came back at 49.3%, short of what the angle needed, so it never reached the board.
The detail
What the test showed. Walk-forward backtest on G5-vs-G5 games, 2022-2025 predict window. At the headline 3-point divergence threshold, the M1 model produced 228 bets with an ATS hit rate of 49.3% and mean CLV of +2.5760 points. The coefficient averaged -0.1748 across folds and was consistent in sign across all four folds. M1 improved margin MAE by -0.034 points over M0 on average (12.152 vs 12.186), though the improvement is small relative to residual standard deviations above 14 points. The 95% CI on ATS hit rate spans across or below 50%, meaning cannot be confirmed or rejected at this sample size.
TestedCFB situational ATS edges (HFA overvalue, big-favorite fade, bowl underdogs, G5 softness).
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.
TestedPreseason power ratings (SP+) beat the early-season CFB opening spread.
FoundIt did not clear our standard, so it came off the board before it ever produced a bet.