One hundred and thirty-six FBS teams, their 2025 results, a deliberately simple Elo baseline, and opponent-adjusted team efficiency from cfbfastR 3.0. Results, Elo and efficiency stay separate because they answer different questions. Everything on this page is drawn from the same dated files Data Dawgs serves to machines.
Opponent-adjusted EPA/play is the clean first gain from the cfbfastR 3.0 release. Better offenses move right; better defenses move up because the vertical scale is EPA allowed and lower is better. Use the sample columns before trusting an early-season dot.
Loading…
Quadrants are split at the FBS medians, not at zero. Upper right means above-median adjusted offense and above-median adjusted defense for this snapshot. This is a season description; it does not estimate what happens next.
Click a column to sort. Defensive values are EPA or success rate allowed, so lower is better. Rank 1 is best in every adjusted-rank column.
| Sample | cfbfastR modelled efficiency | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Team | Conference | Games | Plays | Net rank | Adj net | Adj off | Adj def allowed | Raw off | Raw def allowed | Success diff |
| Loading… | ||||||||||
Pick two teams. This applies the ratings registry’s published transform
to their end-of-2025 strengths — the same formula, scale and venue adjustment the
dd_project_cfb_matchup tool uses, read from the file rather than retyped here.
Transform, as published:
This is a hypothetical, not a forecast. Both ratings are frozen at the end of the 2025 season and no game on any 2026 schedule has been priced with them. Expected margin, spread and total stay empty rather than being reverse-engineered out of a win probability — the rating does not produce them.
One filter across the data sheets: the team table here, the expected-versus-observed chart on the Elo sheet and the divergence table on its sheet all redraw against the same slice.
Loading…
Click any column to sort. The left band is observed — what happened, counted off the canonical schedule. The right band is modelled — one Elo system, fitted to nothing, graded against nothing. They are kept apart here because they are kept apart in the source file.
| Observed · 2025 results | Modelled · dd-cfb-elo · retrodictive, ungraded | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Team | Conference | Record | PF/G | PA/G | Diff/G | Elo rank | Elo | Exp W | Obs W | Obs − Exp |
Elo rank orders this one system’s numbers. It is not a poll, a committee ranking, or a claim about who is better now. Exp W sums that team’s pregame Elo win probabilities across the 2025 walk-forward backtest, so Obs − Exp is model residual — not luck, not team quality, and not a forecast.
Every FBS team, plotted where the Elo expected it against where it finished. The diagonal is agreement. Distance from it is how wrong this rating was about that team across 2025 — which is the only claim the number supports.
Hover, or focus the chart and use the arrow keys, to read a team. Every point here is also a row in the team explorer above, where Exp W, Obs W and the difference are readable as text.
A baseline only earns the name if it is scored against something. These are the published 2025 backtest numbers, read from the model’s own output file. Lower Brier is better; higher favourite accuracy is better.
A team’s win-percentage rank against its point-differential-per-game rank. A positive gap means the record ranks better than the scoring margin. This is a descriptive baseline: the published file carries a null label on every row, and nothing here calls a team overrated, underrated or a fraud.
| Observed · 2025 results, descriptive | |||||||
|---|---|---|---|---|---|---|---|
| Team | Conference | Record | Win% rank | Margin rank | Gap | Direction | One-score |
| Loading… | |||||||
The complete identified CFB opportunity set — roadmap ideas, not automatically tools. Shipped data and model artifacts carry evidence-backed lifecycle state; every remaining card records what we could build, why, what it depends on, and how we would know whether it worked.
The roadmap cards below are ideas. These are the ones an assistant can
actually call today. The split is read at runtime by intersecting the live tool roster in
/data/surfaces.json with the candidate names
on the cards, so it cannot drift from the data the way a typed count can.
“Live” here means the roster says so —
node tools/validate-data.js checks that roster against every per-surface claim,
but this page does not call the Worker and does not independently confirm anything answered.
/data/surfaces.json — the live Worker tool roster, including the College Football tools above.
Ordering: the 12-step sequence stored with this data is directional, not an irreversible commitment — dependencies ultimately determine the actual build order. Shared governance (source provenance, snapshotting/receipts, model cards, baseline requirement, incremental-information testing, correlation awareness, uncertainty) applies across every idea and is published in the same file.
It is one rating, not a consensus. A single system cannot be a consensus, and this one has no blend weights because there is nothing to blend it with. Registry membership is not evidence of skill.
It is retrodictive, not prospective. The 2025 season was already
complete when the backtest ran. Ratings only ever updated from games already played, which is what
makes the numbers meaningful at all — but no row here was frozen before a kickoff, so no row
is a receipt. The prospective ledger at /data/cfb-model-receipts.json is
empty by design and will stay empty until a scheduled 2026 game is priced before it is played.
There are no market prices on this page. The historical CFB market file exists, but every quote in it shares its game’s kickoff timestamp rather than an observation time. That means they are not closing lines and no closing-line value can be computed from them, so showing them beside a model number here would imply a comparison the data cannot support.
What the rating cannot see. Recruiting, the portal, rosters, injuries, weather, travel, play-by-play efficiency, opponent adjustment and coaching. It knows who won, by how much, and where. That is the entire feature set, on purpose — it is the floor a better CFB model has to clear.
The efficiency sheet is a different object. It uses cfbfastR’s play-level EPA model and upstream opponent adjustment, but Data Dawgs has not refit or independently validated either. It describes the published season sample. It does not turn the Elo into an EPA model, and it does not turn an efficiency rank into a forecast.