Indiana finished the 2025 season 16-0. It also ranked first in average scoring margin, first in Simple Rating System, and first in Elo. That part is easy.
The more useful cases are the ones where the methods disagree.
Notre Dame ranked 13th in win percentage but third in both scoring margin and SRS, and second in Elo. Georgia and James Madison both finished 12-2, which tied them for fourth in win percentage. Georgia ranked 10th in SRS and fifth in Elo. James Madison ranked 29th in SRS and 13th in Elo.
Team
Win % rank
Point differential rank
SRS rank
Elo rank
Indiana
1
1
1
1
Notre Dame
13
3
3
2
Georgia
4
15
10
5
James Madison
4
7
29
13
Which ranking is right?
I think that question is too broad. Each method is answering a different question, and each one fails in a different way. The better question is: right for what purpose?
Start with one clearly defined season
I used all completed games involving a 2025 FBS team, including the postseason and games against FCS opponents. Final scores include overtime. Neutral-site games count normally because neither win percentage nor scoring margin needs a home-field adjustment.
The records endpoint and Games API reconciled exactly: all 136 FBS teams had the same game count in both sources. That gave me one consistent set of results for win percentage and point differential.
SRS and Elo are a little different. I pulled the end-of-season values from the CollegeFootballData ratings endpoints and filtered them to the same 136-team FBS population. Those are precomputed ratings and the endpoint responses do not expose every internal choice behind the calculations, so I would not claim that I rebuilt CFBD’s SRS and Elo from an identical game set (both notably exclude games against FCS opponents).
That difference matters. A ranking is not just a formula. It is also a set of decisions about which games count, where teams start, how quickly ratings change, and how schedule strength is handled.
The API requests themselves are short. This example assumes you have already created an authenticated CFBD client:
YEAR = 2025
records = games_api.get_records(year=YEAR)
games = games_api.get_games(
year=YEAR,
season_type=cfbd.SeasonType.BOTH,
classification=cfbd.DivisionClassification.FBS,
)
srs = ratings_api.get_srs(year=YEAR)
elo = ratings_api.get_elo(year=YEAR)
1. Win percentage: what did the team accomplish?
Win percentage is the most transparent measure in the group:
Win percentage = (wins + 0.5 × ties) ÷ games played
There were no ties in the 2025 FBS data, but including them makes the definition complete since ties historically occurred in college football. The corresponding code stays close to that definition:
win_percentage_rows = []
for record in records:
if record.team not in fbs_team_names:
continue
games_played = record.total.games
wins_with_ties = record.total.wins + 0.5 * record.total.ties
win_percentage_rows.append({
“team”: record.team,
“win_percentage”: wins_with_ties / games_played,
})
win_percentages = pd.DataFrame(win_percentage_rows)
The appeal is obvious. Games are played to be won, and a standings or résumé summary should remain connected to that outcome. Indiana’s 16-0 record belongs at the top of that kind of ranking.
The failure mode is just as obvious: teams do not play equal schedules, and a one-point win counts exactly as much as a 40-point win. Win percentage can tell us what happened without telling us how dominant the team looked or how difficult the path was.
Kennesaw State is a useful example. It finished 10-4 and ranked 23rd in win percentage. Seven of its wins came by 14 points or fewer, while losses to Indiana and Western Michigan came by 47 and 35 points. The record properly credits Kennesaw State for winning 10 games. It does not capture the shape of those results.
2. Point differential: how dominant did the team look?
Average point differential adds the score:
Average point differential = total points scored minus total points allowed, divided by games played
I reshape each completed game into one row per FBS team. A positive margin means the team outscored its opponent:
margin_rows = []
for game in games:
if not game.completed:
continue
if game.home_points is None or game.away_points is None:
continue
if game.home_team in fbs_team_names:
margin_rows.append({
“team”: game.home_team,
“margin”: game.home_points – game.away_points,
})
if game.away_team in fbs_team_names:
margin_rows.append({
“team”: game.away_team,
“margin”: game.away_points – game.home_points,
})
point_differential = (
pd.DataFrame(margin_rows)
.groupby(“team”, as_index=False)[“margin”]
.mean()
)
This is why Kennesaw State falls from 23rd in win percentage to 74th in point differential. The method sees both the close wins and the lopsided losses.
Point differential also explains why Notre Dame looks stronger than its 10-2 record alone. The two losses came by three points to Miami and one point to Texas A&M. The Irish then won 10 straight, including eight games by at least 15 points, and finished third in average scoring margin.
That added signal comes with a judgment call. Should a 50-point win count much more than a 30-point win? Large margins can include information about dominance, but they can also reflect garbage time, late-game incentives, opponent depth, and coaching decisions that have little to do with future team strength.
I tested caps of 14, 21, and 28 points per game. The choice was not cosmetic. A 21-point cap moved Florida State from 26th to 62nd in the point-differential ranking and Coastal Carolina from 120th to 94th. If a reasonable parameter changes the ordering that much, the ranking should be presented as a model output rather than an objective fact.
3. SRS: how strong were the opponents?
Simple Rating System starts with scoring margin and adjusts it for schedule strength. In plain language, beating a strong team should count for more than beating a weak team, and losing to a strong team should hurt less than losing to a weak team.
The circular part is the point. We need opponent strength to rate a team’s results, but we also need results to estimate opponent strength. SRS resolves that through a connected schedule network, repeatedly updating teams until the ratings settle.
James Madison shows what that adjustment can do. The Dukes finished 12-2 and seventh in point differential at +18.7 points per game. Most of the wins came against a Sun Belt schedule, while the two losses were to Louisville and Oregon. SRS moved James Madison to 29th after accounting for the complete opponent network.
Georgia moved the other direction relative to raw margin. The Bulldogs were only 15th in point differential, but they ranked 10th in SRS after a schedule that included Tennessee, Alabama twice, Ole Miss twice, Texas, Florida, and Georgia Tech.
SRS is useful when the question is season-long, schedule-adjusted performance. It can still be fragile early in a season, when the schedule network is thin and many teams have not played common or connected opponents. Results also depend on implementation choices: whether to include FCS games, how to handle margin caps, and whether to weight recent games differently.
4. Elo: what did we believe before and after each game?
Elo works sequentially. Each team has a pregame rating. The difference between the two ratings produces an expected result, and the actual result determines the postgame update.
Four choices drive a basic implementation:
Initial rating: where every team begins, or how much prior-season information carries forward.K-factor: how quickly one result can change a rating.Home-field advantage: how much the expected result shifts for the home team.Margin treatment: whether a close win and a blowout produce the same update.
This makes Elo useful for updating an estimate through time and for producing pregame expectations. It also makes Elo path dependent. Two teams with the same final record can finish with different ratings because they started in different places, faced different opponents, and reached those results in a different order.
South Carolina is the sharpest example in this comparison. The Gamecocks finished 4-8, ranking 98th in win percentage and 76th in point differential, but 29th in CFBD Elo. That does not mean Elo believes four wins are better than 10. It means Elo is not a standings table. Its output reflects incremental updates to a prior rating and the strength of the teams involved.
I would not attribute South Carolina’s exact ranking to one parameter without rebuilding CFBD’s complete Elo history. The endpoint result is still useful because it exposes the method’s central tradeoff: prior information can stabilize a rating, but it can also keep the rating far away from the current-season record.
To test sensitivity, I also replayed the 2025 season with a deliberately simple Elo model that started every team at 1500. Changing the K-factor from 10 to 40 moved Duke 18 ranking positions and UNLV 17. Removing a 55-point home-field rating adjustment changed some teams again. None of those settings is automatically correct. Each describes how quickly we want new evidence to replace the old estimate.
Match the method to the question
If you want to measure…
Start with…
Watch for…
Standings or résumé
Win percentage
Unequal schedules and no margin information
Descriptive dominance
Point differential
Blowouts, garbage time, and margin-cap choices
Schedule-adjusted season strength
SRS
Early-season connectedness and implementation choices
Updating estimates and forecasting
Elo
Initialization, path dependence, K-factor, and home field
More sophisticated does not mean free of judgment. SRS and Elo can correct weaknesses in raw standings, but they introduce assumptions of their own. A useful rating makes those assumptions visible and matches them to the decision the reader is trying to make.
If you want to reproduce the comparison, the CFBD API documentation covers the games, records, and ratings endpoints used here. The CFB Starter Pack provides a guided notebook path if you would rather begin with a structured workflow.
Which method do you trust for which purpose: win percentage, point differential, SRS, or Elo? And which assumption would you change first?
