2026-27 NFL
forward proof.

Pre-registration is what separates analytics from astrology.

The Honest Validation Protocol (HVP) rule 3 requires that any model, threshold, or filter affecting live betting decisions has its success and failure criteria fixed before the test data is looked at. Running an analysis and picking the threshold that maximizes claimed performance afterward is in-sample fitting, and the result is unreliable.

NFL backtest performance was measured on three independent walk-forward test seasons (2023, 2024, 2025). The 2026-27 regular season and playoffs are the first set of games the model has never seen and was not trained on. The numbers from this season are the forward proof that matters.

For institutional buyers, academic citers, and journalists: this document is the contract. Everything below is locked. If any of it changes after Week 1 kickoff, the changelog at the bottom of this page will record the date, the reason, and the new signing commit, with the original copy preserved in git history.

Two production tiers under test.

VAR's NFL model runs across five production tiers internally. For this forward test we are publicly pre-committing to the two highest-confidence tiers: PRIME spread and PRIME_TOT. Both have passed HVP validation across three walk-forward test seasons and carry the same lower-bound discipline.

The other three production tiers (STRONG, PLAYABLE, PLAYABLE_SPREAD) remain operational internally but are not part of this public forward test. If they prove durable through the 2026-27 season, they will be promoted into a future pre-registration document.

The numbers being tested forward.

Both tiers were validated 2026-04-30 against three independent walk-forward test seasons (2023, 2024, 2025), each held out from training. The lower bound of the 95% Beta-Binomial credible interval is the figure we plan and size against. Point estimates are reported for transparency, not as commitments.

Per-season PRIME spread results: 2023 60.32% (n=63), 2024 65.85% (n=82), 2025 61.73% (n=81). Per-season consistency was a pre-registered requirement of the backtest itself; no individual test season was below 60%.

See /performance for the full methodology and per-season breakdowns. The Honest Validation Protocol governing these measurements is version 2026-04-30.

Expected sample sizes for 2026-27.

The backtest figures above span three regular seasons. Scaled to a single season, the model is expected to surface roughly:

• PRIME spread: ~75 graded bets across the 2026-27 regular season and playoffs
• PRIME_TOT: ~60 graded bets across the 2026-27 regular season and playoffs

These are estimates derived from the per-season counts above and may move within a normal envelope (the filter is fixed; the number of games that clear it depends on market line dynamics that the model does not control). Final n will be reported on the live tracking page after the conclusion of each NFL week.

What counts as the forward test passing.

For each tier, the season passes if, at Week 18 (the final week of the NFL regular season plus playoffs), the lower bound of the 95% Beta-Binomial credible interval is at or above the break-even threshold of 52.4%. This is the same bar internal HVP audits hold historical claims to.

Point-estimate parity with the backtest is the secondary goal but not the primary commitment. Realized CLV haircuts, regime drift, and ordinary single-season variance can move the point estimate while the lower bound holds. The lower bound is what we cite.

What counts as the forward test failing.

Three pre-registered checkpoints. Each tier is evaluated independently. Responses escalate with severity.

If both tiers trigger the same flag in the same window, that is a systemic signal (model-wide drift, ingestion fault, regime shift) and the postmortem will treat it as such rather than as two independent events.

What we will publish, on what schedule.

• Every pick before kickoff. Each PRIME spread and PRIME_TOT bet that clears the filter will appear on the live tracking page at /performance before the relevant game starts. Timestamps are recorded; nothing is back-dated.
• Every result within 24 hours. Settled outcomes update the live tracking page within one day of the game ending.
• Weekly running stats. By Tuesday following each NFL week, the running n, win rate, CI lower bound, and ROI for each tier are refreshed.
• Postmortems for every checkpoint trigger. Yellow, amber, and red responses follow the table above.
• Per-game research posts. Each NFL game where a PRIME-tier bet fires is covered in a research post under /research/picks, published before kickoff with the model's rationale.

What we will not do during this season.

• No changing the filter thresholds. PRIME spread remains |edge| ≥ 6. PRIME_TOT remains |edge| ≥ 7. The thresholds were derived from per-market calibration in the backtest and are locked.
• No changing the validation criteria. The 52.4% break-even bar on the lower bound of the 95% Beta-Binomial CI is fixed.
• No selective reporting. Every PRIME-tier bet that clears the filter is included in the running stats. No silent dropping of unfavorable weeks.
• No re-defining tiers mid-season. If new tiers are developed from in-season data, they may be operated internally but cannot be added to this pre-registration. They will appear in a separate pre-registration document signed before the 2027-28 season.
• No editing this document. Substantive revisions are appended to the changelog with a new signing date and commit; the prior copy is preserved in git history.

How to verify this document.

This page is served from the open-source victoryar Next.js application. The signing commit hash above is the git commit that first published this content. Subsequent edits are visible in the repository commit log.

The Honest Validation Protocol referenced throughout is version 2026-04-30. A public companion page is forthcoming at /methodology/hvp. Until then, the eight HVP rules summarized in this document are accurate representations of the standard.

Questions or citation requests: @xVictoryarx on X.