CFB Methodology

• Game results & scores:CollegeFootballData.com (cfbd) API, 2010–2025, FBS only, regular season + bowls + CFP. ~10,000 games.
• Team-season aggregates:success rate, explosive rate, finishing drives, havoc, line yards, ppa — all derived from cfbd play-level data per team-season.
• Market lines: closing spreads + totals from cfbd bookmaker feeds. Used for backtest grading and live edge calc on futures.
• 2026 schedule + early lines: cfbd 2026 endpoint, 779 games loaded, ~10 with early lines posted. Reloaded weekly.
• Scouting overrides: 68 spring scouting deep-dives + 12 conference tier reports, parsed into structured signals (returning production, transfer portal grade, coaching changes, schedule difficulty). These adjust the team-strength prior before simulation.

Two-model ensemble that predicts margin (home minus away points) for a single game, given team-season features for both sides plus venue, rest, and conference.

Final pick = 0.55 × ridge + 0.45 × xgb. Backtest MAE on out-of-sample 2023–2024 games: ~13.2 points per game. (College football variance is structurally higher than NFL — expect that.)

Returning production and recruiting features lag what spring practice actually shows. So after the ensemble predicts a margin andbefore the simulation rolls dice, every team gets a per-game points adjustment from three independent sources:

unit_margin = Σ (tier - 3) × weight

weights: QB=2.5  OL=1.5  DL=1.5  back7=1.0  skill=1.0  ST=0.3

spring_margin = signal × 0.5     # signal ∈ {-2, -1, 0, +1, +2}

big_game flags (fire only in CFP / conf championship / rivalry):
  big_game_qb_concern        -3.0    big_game_qb_boost        +3.0
  big_game_qb_minor_concern  -1.5    big_game_qb_minor_boost  +1.5
  championship_hangover      -1.5    new_oc / new_dc          -0.75 each

regular-season flag margin = 0   (key_injury / portal_heavy / qb_uncertainty
                                  widen σ instead of moving the mean)

team_margin = unit_margin + flag_margin + spring_margin
team_margin = max(team_margin, -2.0)         # negative cap only

# in big games (bracket / rivalry / top-12 vs top-12):
#   team_margin uses big_game flag set
#   unit boost is damped by 0.5 (model already knows bracket teams are good)

σ_baseline = 17.0 points

widening flags (take max of any active):
  qb_uncertainty       22.0
  key_injury           21.0
  portal_heavy         20.0
  big_game_qb_concern  22.0    big_game_qb_minor_concern  20.5

game_margin = base_prediction + (home_team_margin - away_team_margin)
game_σ      = sqrt(home_σ² + away_σ²) / sqrt(2)

Every team's intel JSON (units, flags, spring_signal, schedule tier, qb tier, confidence) is in version control and rendered on the team page — nothing is hidden.

The unit-margin number is head-to-head team strength. Conference championship probability depends on three things, only one of which is team strength:

• Win-total distribution— how often does this team finish 9-3 or better? Driven by their schedule, not just talent. Two teams in the same conference can have very different schedules.
• Tiebreakers and standings shape— if multiple teams reach 9-3, head-to-head and division records decide who plays in the title game.
• Title-game win probability— conditional on making the championship game, what's the team's chance of winning it (often vs the conference's top dog).

A team that's objectively stronger but plays the conference's three best opponents will often have a lower championship probability than a slightly weaker team that misses all three. Talent beats schedule head-to-head; schedule beats talent over a 12-game path. The simulation captures both. If a team-vs-team intuition disagrees with a conference-title number, both can be right.

For 2026: 50,000 full seasonssimulated. Each game's outcome is drawn from the model's predicted margin distribution (point estimate + a residual sample sized to its position in the strength gap). Across all 50K seasons we count:

• Win totals: regular-season wins distribution per team.
• Conference championships:probability of appearing in & winning each conference's title game.
• CFP berths:12-team field, conf-champ autobids + at-large bids by simulated CFP committee proxy (resume score = wins · SOS).
• National title: probability of winning the full bracket given seed.

For futures markets where books have posted lines, the displayed edge is:

edge = model_implied_prob - market_implied_prob

market_implied_prob = 1 / american_to_decimal(odds)   # de-vigged via
                                                       # proportional method

Anything above ~3% is worth looking at. Anything above ~6% deserves a closer look at the team page — small edges from weak priors aren't bets.

• Not in-game: no live win probability, no quarter- by-quarter. The pipeline runs weekly.
• Not bet tracking:no login, no wallet, no log of user picks. We publish the model's view; what you do with it is on you.
• Not certainty:CFB has more variance than the model wants to admit. Treat the top edges as “the spots worth thinking harder about,” not as locks.