Insights

Correction to the 2026-27 NFL pre-registration: the backtest anchors are withdrawn

Posted 2026-09-08. Applies to the pre-registration published 2026-05-15 at /methodology/2026-27-predictions.

What this page said

The pre-registration cited the following backtest results for the two tiers it commits to, and described both as having passed validation across three walk-forward seasons:

Tier Accuracy 95% CI ROI n
PRIME spread (|edge| ≥ 6) 62.83% [56.36%, 68.87%] +19.95% 226
PRIME_TOT (|edge| ≥ 7) 63.5% [56.2%, 70.2%] +21.0% 178

What we found

On 2026-05-17, two days after this page went up, an internal audit found that the harness which produced those figures had full-season feature lookahead and early-stopping on the test season. The figures are withdrawn. An independent, leak-free re-run on 2026-06-10 confirmed the finding.

The leak-free numbers

Same model, same games, graded against the closing line with as-of features only:

Measure Result
Spread |edge| ≥ 6 vs the close (play-grain, 2026-05-17) 51.1%, ROI -2.5%, 95% CI [45.2%, 57.0%]
PRIME spread, straight-up (drive-grain, 2026-06-10) 49.3%
PRIME_TOT, straight-up (drive-grain, 2026-06-10) 44.3%
PRIME_ML, straight-up (drive-grain, 2026-06-10) 37.1%
Break-even at -110 52.4%

None of the pre-registered tiers has a validated edge. The corrected backtest is a coin flip or worse.

What still stands

The model’s straight-up winner accuracy, about 66% across the walk-forward seasons, is unaffected. That is a forecasting result, not a betting edge, and it is how we use the model.

What changes

  • The tier definitions and the Week-18 pass/fail criteria on the pre-registration page remain on record exactly as written, so that they cannot be revised after the fact.
  • The 2026-27 forward test proceeds. It is now a test of a signal whose corrected backtest is break-even or below, and it is reported as exactly that. Every per-game research post links this correction and carries only the leak-free figures above.
  • The pre-registration page and the live tracking page at /performance state this correction at the top, with the withdrawn figures struck through and kept for the record.

Why we are publishing this

Our validation standard requires that any number we have cited publicly and later found to be wrong is corrected in place and dated, not quietly replaced. This is that correction.