THEDARKHORSEMODEL ///
Find the underdog with a case. Explore the matchup, check the history, and see the evidence behind every pick.
Flags explain the matchup; they are not extra point adjustments. Related statistics are grouped to avoid counting the same weakness twice. “Upset” uses the market underdog when a spread is posted. Without a posted line, upset probability is unavailable. Pick’em games have no market underdog.
Current season through the latest checked inputs. “Successful plays” gain enough yards to keep a drive on track. For defensive success allowed, lower is better. Rankings compare the teams in the loaded upcoming schedule.
Team strength rating: a rating based on previous wins and losses, with more credit for beating stronger opponents. Higher means stronger. This is also called Elo.
Successful play: gains at least half the yards needed on first down, 70% on second down, or enough for a first down on third or fourth down.
Successful-play advantage: your offense’s successful-play rate minus the rate your defense allows. A positive number is favorable. “pp” means percentage points: 45% minus 40% equals +5 percentage points.
Big-play rate: the share of counted offensive plays gaining at least 20 yards. Small samples are blended toward a common baseline so one game does not dominate the rating.
Verified history and player performances. Interesting context—not automatic prediction bonuses. Current coach and player facts are available only where verified.
Scores, upset signals and the stories behind the matchup.
Conference membership is current. Cross-conference games are grouped separately when viewing all conferences. Forecasts use the latest successfully checked data shown above.
Ranked within your filters by different types of warning signs, then the model’s win probability for the market underdog. Posted, non-pick’em spreads only. This is a screening rule, not a validated betting recommendation.
Select any two teams. Explore their meetings, final scores, last wins and stadium records.
Historical results are descriptive. The score model’s training window is still 2025–26; it does not treat old wins under different coaches as automatic point adjustments.
Loading saved picks…
| Factor | Status | Use |
|---|---|---|
| Team strength / opponent quality | Included | Elo initialized equally in 2025, updated chronologically; 25% regression between seasons. |
| Net success rate / explosive plays | Included | Derived from 2026 play-by-play. Not opponent-adjusted individually; Elo is a separate feature. |
| Offense / defense / home & road | Included | Points, yards, success rates, split residuals. Recent scoring features use up to 12 prior games; displayed splits use 2026 only. |
| Post–big win / look-ahead | Experimental | Big win = beating a team with pregame Elo ≥1600. Look-ahead = scheduled opponent within 14 days rated ≥100 Elo above current opponent. |
| Evening performance | Experimental | Kickoff ≥7 p.m. Eastern; not necessarily nighttime at the venue. TBA times excluded. |
| Head-to-head | Experimental / short history | 2016–2026 meeting results and stadium filters are available in each matchup report. The experimental model still uses only recent 2025–26 residuals; older results are descriptive, not automatic point adjustments. |
| Specific stadium / crowd hostility | Partial | Home-field and team home/road residuals included. Exact-stadium records from loaded 2016–2026 games are displayed; stadium-specific effect and crowd intensity are not separately estimated. |
| Injuries / QB experience / motivation | Reported context | Public availability and QB evidence are refreshed where available. Numerical injury effects and motivation adjustments are not fitted. |
| Weather / travel / special teams | Reported context | Travel and available weather reports are refreshed. They are not numerical score adjustments; special teams modeling remains incomplete. |
| Market spread / TV channel | Displayed snapshot | ESPN game listings and provider odds, with retrieval timestamp. Market spread defines the underdog for the upset meter but does not change the underlying score prediction. |
Missing play-by-play is excluded from rate denominators, not counted as failed plays. Incomplete game coverage and excluded plays mean these are our calculated rates, not an official provider’s success-rate measure. No garbage-time filter or opponent adjustment is applied to the displayed rates.
Two standardized ridge regressions predict home margin and combined points (fixed penalty 30). The core model uses Elo, net success rate, scoring margin, scoring environment, home field, explosive rate and yards. The context model adds home/road residuals, rest, post-win, look-ahead, evening and series effects. Team-specific residuals are heavily shrunk. Rankings are descriptive and are not added again as separate predictors.
Pregame snapshots use only results from earlier calendar dates. Training begins with 100 eligible games; each later date is held out together. Eligibility requires at least three prior games and 80 eligible offensive plays for each team. Model weights are then refit on all eligible completed games for upcoming forecasts. Source corrections and retrospective schedule availability mean this is a reconstructed backtest, not a fully archived point-in-time production trial.
Win and 21+ margin probabilities use a normal distribution of margin error. Held-out intervals use each training fit’s residual standard deviation (floor 10 points); upcoming forecasts use held-out error spread when enough observations exist. These intervals do not include all parameter, injury or data uncertainty. A normal approximation misses some football scoring structure.
Team streaks, last wins, one-score finishes, venue records and extreme results are calculated from loaded 2016–2026 regular-season schedules. They are not all-time claims; postseason completeness is not guaranteed. Coach trivia is sourced from official reporting. Player facts match athlete IDs in current 2026 rosters to prior 2023–26 matchup box scores, with explicit sample counts. These are descriptive performances, not proof a player performs better against that opponent than against other schools. Facts do not add points to the forecast.
The meter shows the model’s win probability for the market underdog. Missing spreads and pick’em games have no market underdog and no upset percentage. Risk labels are editorial: below 20% lower, 20–34% live possibility, 35–50% elevated, above 50% model favors the upset. The top three games in your filters are ranked by number of different warning-sign types, then win probability, and require a posted nonzero market spread. This rule is not itself backtested.
Sportsbook spread and model spread are separate. A negative number means that team is favored by that many points. Provider odds are retrieved snapshots, not live prices or bookmaker-update timestamps. TV listings can change, and unconfirmed kickoff times display TBA.
We group warning signs into independent evidence families: competitive baseline, efficiency, location/schedule, and recent series. Three aligned families trigger “Multiple upset signals.” This is a transparent explanation rule, not a validated storm classifier. It cannot certify a perfect storm while major injury or QB inputs are missing. Blowout probability is shown separately; high variance can increase both upset and blowout tails.
ESPN schedules, results and game play-by-play supply this snapshot. CollegeFootballData authenticated API access is a possible next source for more historical advanced data. Historical regular-season schedules cover 2016–2026; postseason completeness is not guaranteed and missing team-seasons are disclosed. No API credential is configured here. The daily updater retrieves schedules, results and public availability reports. Numerical injury effects and a comprehensive historical motivation database are not yet fitted.
A lower error means predictions were closer to the results. “Brier score” measures how close win probabilities were to what happened: lower is better. A model promising 80% ranges should contain the real result roughly 8 times out of 10.
For example, teams given a 60% chance should win about 6 times out of 10 over many games. Small groups can vary a lot.
These picks were calculated afterward using earlier-game information. “Correct winner” does not necessarily mean the predicted score was close.