Bowl Championship Series

BCSrankings

The book

How a number gets into the standings.

From 1998 through 2013 the Bowl Championship Series mixed human polls with computer ratings and used that average to pair teams for the big bowls. Six computers survived to the end. Two of them, Colley and Massey, described their methods in public. The other four did not. This site recomputes the two that can be recomputed, approximates the four that cannot, and then keeps going with ratings the BCS would have rejected for using the score.

The composite

The formula is the one used from 2004 through 2013. Each computer ranks FBS teams. First place is worth 25 points, second 24, down to one point for 25th. Anyone outside the top 25 scores zero. For each team the single highest and single lowest of those six point totals are discarded — one of each, even if several computers agree — and the remaining four are added and divided by 100. That is the computer percentage.

computer = (sum of the middle four point totals) / 100

A human poll’s percentage is the points a team received divided by the maximum, which is the number of voters times 25. Voters are counted as the sum of first-place votes, including teams that only received votes. When both the AP poll and the AFCA Coaches poll have usable totals, the standings are an equal blend of three pieces:

BCS = (AP + coaches + computer) / 3

That is not “polls plus computers, split in half.” The Harris Interactive poll held the AP’s seat from 2005 to 2013 and is no longer published, so the AP is back where it sat in 2004. If only one poll has vote totals, that poll and the computers split the score evenly. If neither does, the page shows the computer percentage alone and says so. Ties in the ordering are broken by wins, then losses, then name, so the 25-point scale always has a single team on each rung.

One famous quirk is kept on purpose. A team that is No. 1 in five computers and unranked in the sixth drops the zero, and the computer percentage can still be 1.000. A second zero sticks.

Colley Matrix

Published

Wesley Colley published the method. It uses wins and losses only. A 3-point win is the same event as a 40-point win, and the venue is ignored. Each team begins as if it were 1–1, which is Laplace’s rule of succession, and the rating is then adjusted for who those opponents were.

(2 + tᵢ) rᵢ − Σⱼ nᵢⱼ rⱼ = 1 + (wᵢ − ℓᵢ) / 2

tᵢ is games played, nᵢⱼ is meetings between i and j, and a tie is half a win and half a loss for both teams. The ratings are solved as one linear system. Their average is one half. Nothing in the matrix knows the score or which team was at home. The implementation is checked against that algebra on a three-team season whose ratings are 0.7, 0.5 and 0.3.

Massey BCS

Described

Massey published two different things, and they are easy to mix up. The textbook Massey method is least squares: each game says the winner’s rating minus the loser’s equals the margin. For a win/loss-only version of that system, the margin is just 1. That closed form is in the test suite — the same three-team season rates ⅔, 0 and −⅔ — but it is not what he submitted to the BCS.

His BCS page calls the rating a binary censored score. Only the winner is observed. Teams start with no preseason opinion. Home field counts. Early games are discounted slightly. Games between well-matched teams carry more information, which is what a probability model does and what equal-weight least squares does not. The fit here is a probit model,

P(i beats j) = Φ(rᵢ − rⱼ + h · venue)

with venue equal to +1 at home, −1 on the road, and 0 at a neutral site. A week-1 game is weighted 75 percent of a game in the current week; he did not publish the exact schedule. A wide Gaussian prior, standard deviation 2 on the latent scale, keeps an undefeated team from running off to infinity in September. It is a numerical rail, not a preseason poll.

The four private computers

These are labeled approximations everywhere they appear. They follow what the authors said in public about ingredients, and they obey the BCS rule that margin of victory was forbidden and venue was allowed. They are not reconstructions of private code.

Sagarin ELO-CHESS

Jeff Sagarin published a margin-aware predictor and, separately, a win/loss Elo he called ELO-CHESS. The BCS used the second. Here every team starts at 1500 and the season is replayed once in date order with K = 32. Home field is 65 Elo points, about a 59 percent edge for an otherwise even team. A one-point win moves the rating exactly as much as a 40-point win. Same-day order follows the game id. His real K, home edge, and initialization were not released.

Billingsley

Richard Billingsley described a rating built from record and opponent strength, with recent games mattering much more than September and with venue as a trace. Margin, once about five percent of his formula, came out after 2001. The approximation gives each game a five-week half-life. A win is worth more against a strong opponent; a loss hurts less against one. Road wins are nudged up by about five percent. The ratings are iterated until they stop moving.

Anderson & Hester

Contemporary accounts describe record, schedule, and location, without margin. A road win is scored 1.15, a home win 0.85, a road loss 0.15, and a home loss −0.15. The rating then solves

r = ½ · (average game credit) + ½ · (average opponent rating)

which is a linear system. It is not their code.

Wolfe

Peter Wolfe’s ratings were a Bradley–Terry maximum likelihood model: the chance i beats j is a logistic function of the rating gap. Home field is estimated here because the BCS allowed it. The same wide prior used for Massey keeps separated teams finite. There is no recency weight; the whole season is one likelihood.

Ratings the BCS would not have used

These sit beside the composite and never enter it. The point of the separation is so a blowout cannot quietly change who is No. 1 in the standings.

Diminishing margin

Least squares, as in the textbook Massey method, but the target is the sign of the margin times log(1 + |margin|). Three, fourteen, and forty points land near 1.39, 2.71, and 3.71. Home field is estimated on that scale. A 17–10 game and a 45–38 game look the same; the efficiency rating is the one that notices pace.

Adjusted efficiency

Each team has an offensive rating and a defensive rating. Points scored by the home team are modeled as offense minus the opponent’s defense, plus a home-field constant. Both sides of every game are equations. Zero means an average unit. The team rating is offense plus defense: expected margin against an average opponent on a neutral field, in points. A higher defensive number means fewer points allowed.

Bayesian Bradley–Terry

Same likelihood as Wolfe, with a tighter prior (standard deviation 1) so a two-game sample cannot pretend to be a ten-game sample. The 90% band is a Laplace approximation: the curvature of the log posterior at its mode, inverted. The rank band then asks how far this team could slide if everyone else stayed put. It understates how much the whole ranking can shuffle, and it is wide on purpose in October.

Walk-forward predictor

Before each week, Elo and the efficiency rating are whatever the previous weeks support. A logistic regression maps the Elo gap, the efficiency gap, and a home indicator to the chance the home team wins. That regression is fit only on games already played. The week is scored, and only then are the ratings allowed to see it. Reported accuracy, log loss, and Brier score start in week 4 and count FBS-versus-FBS games. The published rating applies those weights to the ratings after the week. It is a forecast, so it will disagree with a résumé ranking, and that disagreement is the feature.

Where the games come from

If CFBD_API_KEY is set, the update uses the CollegeFootballData.com API: games, FBS membership, and the AP and coaches polls. That is the preferred source, because a team’s conference is the one it actually played in that season.

Without a key, the pipeline walks ESPN’s public scoreboard for the FBS and FCS groups, day by day, and reads polls from ESPN’s core API. This copy of the site was built that way: no key was available in the environment. ESPN’s group filter and standings tree decide who is FBS. A conference name that clearly belongs to FBS is enough when a team is missing from the standings payload. FCS opponents stay in the matrix so that beating a strong FCS team is not the same event as beating a weak one, and they are left out of the published tables.

Bowls and the postseason are stored and then ignored. The historical BCS standings stopped before the bowls, on championship weekend. Weeks follow ESPN’s calendar. The front page prefers the latest week that added at least twenty games, so a Tuesday game does not replace Sunday’s rankings.

Polls are aligned by date: the ballot dated within four days after the last game of the week, which is the Sunday release. The 2013 AP poll in ESPN’s archive has first-place votes and zero point totals, so that season uses the coaches’ poll alone and says so. When ESPN still carries the official BCS standings, they are shown as a reference column. They are not an input.

What will not match the old printout

  • Colley’s published matrix reaches below Division I. Games against Division II, III, and NAIA are missing here, and so are the rest of those teams’ schedules. FBS order can still agree while the raw ratings sit a little higher, because the missing weak teams are not in the average.
  • Massey’s early-season weights and home-field parameterization were described, not specified. The probit is the model his description points to, not a byte-for-byte copy.
  • Sagarin, Billingsley, Anderson & Hester, and Wolfe are approximations. Treat agreement among them as a hint, not as their historical ballot.
  • The Harris poll is gone. Substituting the AP changes the human third relative to 2005–2013.
  • Conference labels on older seasons come from ESPN’s standings for that year when the payload is complete, and from conference ids on the games otherwise. A short or renamed conference can be labeled coarsely.

Checked against Colley’s own table

After each build the Colley ratings are set next to the HTML he publishes. The comparison lines up our week with whichever of his weekly files shares the most top-25 names, because his week index and ESPN’s are not guaranteed to be the same number.

  • 2013. Our week 16 against his week 13: 24 of 25 names in common, Spearman 0.95 on the matched top, mean absolute rating gap 0.080. Ours: Auburn, Florida State, Stanford, Alabama, Missouri. His: Florida St, Alabama, Ohio St, Auburn, Missouri.
  • 2024. Our week 16 against his week 15: 24 of 25 names in common, Spearman 0.99 on the matched top, mean absolute rating gap 0.094. Ours: Oregon, Georgia, Notre Dame, Texas, Ohio State. His: Oregon, Notre Dame, Georgia, Ohio St, Texas.
  • 2025. Our week 16 against his week 17: 24 of 25 names in common, Spearman 0.96 on the matched top, mean absolute rating gap 0.094. Ours: Indiana, Georgia, Ohio State, Oregon, Oklahoma. His: Indiana, Georgia, Ohio St, Oregon, Mississippi.
  • 2026. Our week 12 against his week 11: 23 of 25 names in common, Spearman 0.98 on the matched top, mean absolute rating gap 0.082. Ours: Utah, UCLA, Texas, Georgia, Miami. His: Utah, UCLA, Texas, Georgia, Miami (FL).

The algebra tests do not need the network. From the repository root: pytest.