Methodology overview
CFBD models answer different questions. PPA values an individual play, WEPA adjusts season efficiency for opponent strength, Elo updates a team's rating after each game, SRS evaluates point differential across a schedule, CORE combines play context with opponent adjustment, and win probability estimates the chance of winning from a specific point before or during a game.
Those outputs are related, but they are not interchangeable. A rating is not a probability, a descriptive efficiency measure is not necessarily a forecast, and values from different models should not be compared on the same scale.
Choose the model that fits the question
| Model | Primary question | Unit or scale |
|---|---|---|
| PPA | How much value did a play create or lose? | Expected point value |
| Win probability | How likely was the home team to win? | Probability from 0 to 1 |
| WEPA | How efficient was a team or player after opponent adjustment? | Model-adjusted efficiency |
| Elo | How has a team's game-by-game strength estimate changed? | Relative rating points |
| SRS | How strong was a team's scoring margin after schedule adjustment? | Points relative to average |
| CORE | How efficient was a team after play-context and opponent adjustment? | Points above average per 100 plays |
If you are building an analysis, start with the question rather than the model name. For example, use PPA for play value, in-game win probability for game state, and a team rating for season-level comparisons.
What CFBD publishes
These pages explain what each model measures, the broad methodology, how to read the output, and the most important limitations. PPA, win probability, WEPA, Elo, SRS, and CORE are proprietary CFBD models. Exact formulas, fitted coefficients, training artifacts, and every implementation detail are not part of the public documentation.
That boundary matters for interpretation. The pages provide enough context to use the outputs responsibly without presenting a simplified formula as if it were the production model.
Comparing results over time
Always retain the season, week or snapshot boundary, season type, and model version when the endpoint provides them. Model changes can affect the comparability of values across periods. Coverage also varies by endpoint, so a missing value should not be treated as a zero.
See Metrics and definitions for the shared terms used across these models and the API reference.