Stats Glossary
All metrics are opponent-adjusted where noted. Green values indicate better performance; red indicates worse. For defensive metrics, lower EPA allowed is better (displayed green).
How the Projected Ranking Is Built
PROJECTED mode blends two things and shifts weight between them automatically as the season plays out —
no manual switch, no hard cutover week.
The blend
Power Score = (preseason weight) × Preseason Composite + (1 − preseason weight) × this season's actual stats so far. Preseason weight = 3 / (3 + games played) — 100% preseason at kickoff, roughly 50/50 by game 3, and effectively pure in-season data by week 6-7. The "3" was backtested against 6 real seasons of outcomes, not guessed.
What the in-season half weighs
EPA offence and defence (0.14 each), Elo (0.12), SP+ (0.10),
strength of schedule (0.10), success rate offence and defence
(0.08 each), yards per play allowed (0.073) and gained (0.041),
havoc (0.004). Renormalised over whatever is available, so a
partially-ingested season still ranks.
Updated 2026-10-03 — see below
Two weights changed on 2026-10-03
The in-season weights had been set by judgement rather than fitted.
Fitting them against a neutral test — whether the rating
predicted who won the following week's games, across 64
season-weeks and 4,800+ games — produced two changes that stood
on their own. Havoc fell from 0.06 to 0.004: it earned
under 1% of fitted weight and was negative in 3 of 5 test years.
Yards per play entered at 0.114 combined, second only
to EPA; it had been listed as an input since the ranking was built
but was never actually wired up, so it had always counted zero.
Median movement was 2 places.
research/cfb-autoresearch/log.md, iteration 18
What is still not fitted
Elo, SP+ and success rate — 46% of the in-season weight —
keep hand-set values, because no week-by-week history of them exists
to fit against: those tables stored one row per season and overwrote
it nightly. Weekly snapshots now accumulate, so this becomes
answerable next season. Stated plainly because "fitted" describes
half this measure, not all of it.
Preseason Composite
A weighted blend of five signals, all knowable before Week 1: last season's SP+/SRS/FPI (40%), returning production – overall (20%) and QB specifically (8%), 247Sports talent composite (12%), and a coaching/trajectory adjustment (20%, see below). Weights are reasoned, not fit by regression, except the trajectory term.
Coaching / Trajectory Adjustment
Every team's rating pulls slightly toward its own ~20-year historical SP+ baseline (teams that had an unusually good or bad year tend to partly revert) — plus an extra, larger pull for teams with a new head coach this season. Backtested against 788 real team-seasons (2020-2025): explains ~20% of year-over-year SP+ swings, more than any single other input here.
Methodology: docs/power_ranking_methodology.md in the CFBAnalytics repo
Model Metrics
Power Score
Composite z-score across all efficiency, talent, and ratings metrics. Higher = stronger team overall. In PROJECTED mode this is the blended score described above; in FINAL mode it's pure current-season stats with no preseason blend.
Source: CFBAnalytics model
Pred Margin
XGBoost-predicted point margin vs. an average FBS opponent on a neutral field. Positive = favored.
Source: CFBAnalytics XGBoost regressor
Elo
Rolling strength estimate updated after each game based on win/loss and margin relative to expectation. Roughly centered at 1500 for an average FBS team.
Source: CFBAnalytics model
Talent
247Sports Composite recruiting rating aggregated over recent classes. Reflects roster quality ceiling independent of coaching or scheme.
Source: 247Sports via CFBD API
External Efficiency Ratings
SP+
Opponent-adjusted, tempo-free efficiency rating. Splits into offense (higher = better) and defense (lower = better). Designed to be predictive, not descriptive.
Source: ESPN / Bill Connelly via CFBD API
FEI
Fremeau Efficiency Index — scoring advantage per non-garbage possession a team would expect against an average opponent on a neutral field. Accounts for field position and game situation.
SRS
Simple Rating System — margin of victory adjusted for strength of schedule. Iteratively solved so a team's rating equals its average margin adjusted for opponent quality.
Source: CFBD API
FPI
Football Power Index — ESPN's predictive power index combining efficiency stats, recruiting, and injuries to forecast future game outcomes.
Source: ESPN via CFBD API
EPA — Expected Points Added
EPA OVR
Average Expected Points Added per play across all play types. Measures how much value each play generates relative to the expected points from that down-distance-field position.
Source: CFBD API advanced stats
Rush EPA
Average EPA per rushing play. Positive values mean the run game consistently gains more than expected; negative means it loses value on the ground.
Source: CFBD API advanced stats
Pass EPA
Average EPA per passing play (including sacks). The pass game typically drives EPA more than the run game — elite offenses often exceed +0.3.
Source: CFBD API advanced stats
Succ%
Success Rate — % of plays considered "successful": gaining ≥50% of yards needed on 1st down, ≥70% on 2nd, 100% on 3rd/4th. A consistency metric independent of big plays.
Source: CFBD API advanced stats
Defense note: For defensive EPA columns, negative values are good — they mean the defense is forcing opponents into negative-value plays. The color coding is flipped accordingly (green = negative EPA allowed).
Returning Production
Ret%
Overall % of prior-season PPA (Expected Points Added) returning. Captures how much of a team's total production is back — offense and defense combined.
Source: CFBD /player/returning API
QB%
% of prior-season QB PPA returning. The single highest-leverage returning production signal — a returning starter QB is worth ~3–4 pts/game vs. a new starter.
Source: CFBD /ppa/players/season + /roster
Skill%
% of WR/TE/RB/FB PPA returning. Measures skill position continuity — important for QB-receiver timing and run game familiarity.
Source: CFBD /ppa/players/season + /roster
DL% / DB%
% of defensive line and defensive back PPA returning. DL% tracks pass rush continuity; DB% tracks coverage continuity. Both are strong early-season predictors.
Source: CFBD /ppa/players/season + /roster
OL / Pass Protection
Sack%
Offensive sack rate — sacks allowed as a % of total pass dropbacks. Lower is better. Elite OLs are typically below 4%; poor pass protection exceeds 7%.
Source: Computed from CFBD /plays (play-by-play)
Std/Pass Sk%
Sack rate split by down type. Standard downs (1st, 2nd-and-short, 3rd-and-short) vs. passing downs (2nd-and-8+, 3rd-and-5+). Passing down sack rate is a strong OL quality signal.
Source: Computed from CFBD /plays
Std/Pass LY
Football Outsiders Line Yards split by down type. The FO formula credits OL for the first 4 yards of a carry (100%), yards 5–10 (50%), and discounts beyond 10 (0%). Split by standard vs. passing downs isolates OL performance in expected-run vs. expected-pass situations.
Source: Computed from CFBD /plays using FO Line Yards formula
PFF Grades
Play-by-play grades from Pro Football Focus, 0–100, shown on the
Team Profile tab. Thresholds are not the same from one grade
to the next, so a number means little on its own: the median FBS team grades
80.4 overall but only 63.3 in run blocking. Each bar on a team profile carries a grey
tick at that grade's own FBS median — read against the tick, not against a
remembered rule of thumb.
Overall / Offense / Defense / Special Teams
The four composites. Offense and Defense are near-perfect roll-ups of the unit grades below (R² 0.98 against them across 1,052 team-seasons); Special Teams is not (R² 0.15), which is why it carries information the others cannot. FBS medians: 80.4 / 75.0 / 79.2 / 74.1.
Model inputs — prior season's value, lagged one year
Passing / Receiving / Pass Blk
Offensive unit grades: quarterback play, pass catching, and the line protecting the quarterback. FBS medians 69.7 / 70.4 / 67.1 — all well below the composite medians, so a 70 here is roughly average rather than weak.
Displayed only
Rushing / Run Blk
Ball carriers, and the line blocking for them. The widest gap on the page: rushing's FBS median is 81.5 while run blocking's is 63.3, so the same 72 is poor in one and good in the other.
Displayed only
Pass Rush / Coverage / Run Def / Tackling
Defensive unit grades: pressure on the quarterback, the secondary limiting separation, front-seven run fits, and open-field tackling. FBS medians 73.1 / 75.1 / 76.7 / 68.7. Nothing else on this dashboard decomposes a defense this way — SP+, SRS and FPI are single numbers.
Displayed only
Point Spread Rating
PFF's headline team rating, in points — roughly what they would favour the team by against an average opponent on a neutral field. Correlates r = 0.90 with SP+, so it mostly restates a rating already shown here.
Displayed only
Why last season's grades, not this season's
A finished season's grade contains every game of that season, so using it to predict a game inside that same season means reading the answer. The model sees the prior season's grades, known before kickoff. The profile page shows the current season's, because a profile predicts nothing and the freshest number is the right one there.
On the model: the four composites entered the
prediction model on 2026-09-29 after walk-forward testing, and the gain is small —
Brier improved in 5 of 5 random seeds, but by 0.0004; the AUC change sits inside
reseeding noise. Schedule-strength adjustment of the grades was tested three ways and
rejected, since SP+, SRS and FPI are already opponent-adjusted. Method and full numbers:
research/cfb-autoresearch/log.md, iteration 14.