Current research constitution · Sept. 12, 2026

How Rabby DIII is being built.

The methodology is established enough to run research rankings and frozen experiments, but it is not yet production-frozen. Historical walk-forward validation decides the final specification.

2021–2025 development window2026 live seasonPREVALIDATION
Universe

Season-specific identities first

The 2026 live universe is 244 active DIII programs. Historical teams remain in their season graph even if they later closed, moved divisions, restarted, or left DIII. Source division labels never override verified canonical identity.

Graph

DIII inside a football-wide bridge

DIII-DIII games form the divisional graph. Actual crossover games connect DIII to DII, FCS and FBS underneath the separate published divisional rankings. Non-NCAA opponents are tracked separately when useful.

Data quality

Verify, repair, quarantine

Mirrors are deduplicated. Malformed labels are normalized. Forfeits without a played game are excluded from score fitting. Conflicts are quarantined instead of guessed. Postseason status is independently repaired when source metadata is wrong.

Model selection

Out-of-sample performance wins

Rabby does not tune toward polls. Candidate models must survive 2021–2025 walk-forward testing with no future leakage. Accuracy, calibration, score/margin error, stability and failure modes determine the winning specification.

Current 2025 model

R0.2 hierarchical cross-division ridge

The audited 2025 graph contains 1,226 DIII-DIII model edges across 241 DIII programs, plus 9 NCAA crossover edges. Team effects are ridge-regularized and connected through division effects estimated from actual crossover games. The 10-team NESCAC component is prior-anchored rather than falsely treated as directly connected.

2.06R0.2 fitted HFA
12.74Training RMSE
10.13Training MAE
2412025 DIII teams
Today’s frozen experiment

Sept. 12 predictions stay exactly as issued

Today’s 94 modeled games use the frozen formula recorded at lock time: 2025 R0.2 DIII-relative ratings, +1.29 home-field points, logistic win-probability scale 10.5, 52-point combined-score baseline, and a ±45 margin cap. Ten cross-division/association games remain track-only.

Important model-development discrepancy: the frozen 1.29-point home-field value differs from the later audited R0.2 full-graph estimate of 2.061757. We will not rewrite the lock. The difference becomes evidence to test during grading and walk-forward validation.

Remaining validation

What still has to happen before “certified” means certified

Build 2024–2021 graphs, run historical walk-forward candidate testing, select the best out-of-sample specification, freeze that methodology, rebuild 2026 YTD under the frozen standard, and only then publish the first certified live Rabby DIII 244 ranking.

2026 live ranking experiment

Baseline stays frozen; live evidence moves separately

The website now preserves the complete 2025 R0.2 ranking as an immutable comparison baseline while a separate 2026 experimental ranking re-solves team strength from current DIII-DIII score margins. The live solver uses the 2025 rating as a four-game-equivalent ridge prior and the audited R0.2 2.061757-point home-field estimate. Finalized 2026 games enter by default; an optional in-progress mode is explicitly preview-only. This live specification is observational and may be replaced after walk-forward testing. It does not alter the frozen 2025 ranking or the Sept. 12 prediction lock.