The scoreboard
doesn't lie.
We track every prediction we make. No cherry-picking. No revisionist history. Just a running record of how our models perform against the market, updated continuously.
The PRIME spread accuracy that appeared here and in the pre-registration was produced by an audit harness with full-season feature lookahead and early-stopping on the test season. It was withdrawn on 2026-05-17 and the finding was confirmed by an independent leak-free re-run on 2026-06-10. Graded leak-free against the closing line, the same tier is 51.1% (95% CI 45.2% to 57.0%, ROI -2.5%); PRIME_TOT is 44.3% straight-up. Neither is a validated edge. The pre-registered tiers, success criterion, and checkpoints are unchanged, and the forward test below proceeds as a test of a break-even-or-below signal. Full notice: /insights/nfl-2026-27-pre-registration-correction.
The backtest tables previously shown in this section (ATS by edge threshold, totals by edge threshold) were produced by the same harness and are withdrawn pending a leak-free recompute. The model's straight-up winner accuracy, about 66% across the walk-forward seasons, is unaffected; it is a forecasting result, not a betting edge.
Pre-registered before NFL Week 1 2026 under the Honest Validation Protocol. Signed and dated at /methodology/2026-27-predictions; corrected 2026-09-08. Predictions are evaluated against the closing line at the primary sharp market; all figures are out-of-sample.
1,283 games
simulated.
Michigan over UConn (69-63). Simulator had Michigan -2.9
Season accuracy: 853/1,283 straight up, 469/704 ATS, 348/704 O/U. NCAA Tournament: 52/66 correct predictions across all rounds.
Every fight,
graded.
The full record of the production UFC model since the live record began on May 30, 2026: every bout it priced before the bell that has since settled with closing odds. Straight-up accuracy is the skill measure; the Brier score says whether the model's probabilities beat the market's. Nothing here is a betting recommendation.
The pre-June UFC tier ROI validation was withdrawn: the audit substrate recorded the red corner as the winner in 100% of fights, so tier win rates equalled the red-bet fraction. What survives is straight-up winner accuracy. On the live record so far the market's probabilities are better calibrated than the model's, and no betting tier has a positive return. VAR does not present UFC as a validated betting edge. Honest Validation Protocol.
Card by card
Card names link to the research post published before the event.
| Date | Card | Bouts | Model SU | Market SU | Brier model | Brier market |
|---|---|---|---|---|---|---|
| Sep 5, 2026 | UFC Fight Night: Hooker vs Parnasse | 13 | 7/13 (54%) | 9/13 (69%) | 0.293 | 0.169 |
| Aug 29, 2026 | UFC Fight Night: Nurmagomedov vs Song | 13 | 7/13 (54%) | 7/11 (64%) | 0.259 | 0.224 |
| Aug 22, 2026 | UFC Fight Night: Hernandez vs Rodrigues | 12 | 5/12 (42%) | 7/11 (64%) | 0.287 | 0.273 |
| Aug 15, 2026 | UFC 330: Makhachev vs Machado Garry | 12 | 7/12 (58%) | 8/12 (67%) | 0.262 | 0.262 |
| Aug 8, 2026 | UFC Fight Night: Gamrot vs Salkilld | 9 | 9/9 (100%) | 9/9 (100%) | 0.143 | 0.101 |
| Aug 1, 2026 | UFC Fight Night: Medić vs Rodriguez | 13 | 9/13 (69%) | 10/11 (91%) | 0.193 | 0.112 |
| Jul 25, 2026 | UFC Fight Night: Ankalaev vs Guskov | 12 | 8/12 (67%) | 10/12 (83%) | 0.220 | 0.113 |
| Jul 18, 2026 | UFC Fight Night: Du Plessis vs Usman | 12 | 9/12 (75%) | 8/11 (73%) | 0.146 | 0.127 |
| Jul 11, 2026 | UFC 329: McGregor vs Holloway 2 | 13 | 10/13 (77%) | 10/13 (77%) | 0.186 | 0.182 |
| Jun 27, 2026 | UFC Fight Night: Fiziev vs Torres | 12 | 9/12 (75%) | 8/11 (73%) | 0.204 | 0.152 |
| Jun 20, 2026 | UFC Fight Night: Kape vs Horiguchi 2 | 11 | 7/11 (64%) | 9/11 (82%) | 0.219 | 0.146 |
| Jun 15, 2026 | UFC Freedom 250: Topuria vs Gaethje | 7 | 5/7 (71%) | 5/7 (71%) | 0.197 | 0.195 |
| Jun 6, 2026 | UFC Fight Night: Muhammad vs Bonfim | 12 | 10/12 (83%) | 9/12 (75%) | 0.171 | 0.168 |
| May 30, 2026 | UFC Fight Night: Song vs Figueiredo | 10 | 6/10 (60%) | 7/9 (78%) | 0.237 | 0.152 |
Betting-tier audit (underpowered)
| Tier | Bets | Win rate | ROI | 95% CI on win rate |
|---|---|---|---|---|
| Model edge 5 to 7 pts (vig-free) | 17 | 29.4% | -55.8% | [13.3%, 53.5%] |
| Model edge 7 to 10 pts (vig-free) | 21 | 33.3% | -5.3% | [17.2%, 54.9%] |
| Model edge 10 to 15 pts (vig-free) | 24 | 33.3% | -20.1% | [18.0%, 53.5%] |
| Model edge 15 to 20 pts (vig-free) | 20 | 15.0% | -60.8% | [5.5%, 36.3%] |
| Model edge 20+ pts (vig-free) | 39 | 23.1% | -27.7% | [12.7%, 38.5%] |
| Any underdog lean | 103 | 19.4% | -38.3% | [13.0%, 28.1%] |
| High-confidence underdog (model 65%+) | 8 | 37.5% | -15.0% | [13.7%, 70.1%] |
| Underdog lean (model under 65%) | 73 | 16.4% | -45.1% | [9.7%, 26.6%] |
| Favorite lean | 9 | 55.6% | -3.3% | [26.2%, 81.3%] |
Underpowered. Same tier logic as the internal audit; 8 fight(s) skipped for odds mismatch.
Every fight the production model priced before the bell since the live record began, graded against the official result. A prediction counts only if it was stamped before the event's start time; nothing is back-filled and nothing is dropped.
Straight-up accuracy is the corner-invariant skill measure. The Brier score compares the model's win probabilities with the de-vigged closing consensus on the same fights; lower is better.
Validation note: the betting-tier ROI figures VAR cited for UFC before 2026-06-14 were withdrawn. The audit data had the red corner recorded as the winner in 100% of fights, so every tier's win rate was an artifact of corner assignment, not skill. This table is the honest replacement: live, pre-fight, and checked to be non-degenerate (red wins about half the time).
The tier audit is severely underpowered: a handful of bets per tier and wide intervals. It accrues one card at a time and is shown so that the negative reads are on the record too.
Until 2026-09-08 this section showed a hand-picked list of winning picks from March and April 2026. It was removed because it was selective. This is the full record.
1,176 games.
28.9M simulations.
The 63.2% headline that previously appeared here is superseded. The April 2026 Honest Validation Protocol audit re-litigated every published win-rate VAR had ever cited and retracted any that didn't pass the eight-rule bar (walk-forward across at least three independent test seasons; Beta-Binomial 95% credible-interval lower bound on every cite; pre-registered ship gates; independent samples; empirical CLV haircut; block-bootstrap bankroll sim; memory hygiene; production-code-path verification). The pre-registered NFL PRIME spread tier above is what survives.
We publish this correction because it's the kind of transparency that separates serious analytics operations from black-box pick services. The full protocol is at /methodology/protocol; the live application is the pre-registration linked above.
All performance metrics are evaluated against closing lines, the sharpest available market signal, rather than opening lines. This is the honest test.
Closing Line Value (CLV) is the gold standard for measuring predictive edge because it captures the final consensus of all market information. A model that consistently beats the closing line is generating real alpha, not exploiting stale openers.
Why we
publish this.
Most analytics companies don't show you their track record. We do, because edge is verifiable, and because the clients we want to work with know the difference between a real model and a marketing claim.
Live NFL
forward proof.
The 2026-27 NFL season is the first publicly-verifiable forward test of VAR's model. The two tiers under public pre-registration are shown below with their leak-free backtest figures (the 2026-04-30 anchors were withdrawn; see the correction) and their forward (live) progress. Every PRIME-tier pick appears here within 24 hours of the game ending. Updates append-only.
The backtest figures cited in the 2026-27 pre-registration (PRIME spread 62.83%, PRIME_TOT 63.5%) were withdrawn on 2026-05-17: the audit harness had feature lookahead. Leak-free recompute: spread |edge| ≥ 6 is 51.1% against the close (95% CI 45.2% to 57.0%), PRIME_TOT 44.3% straight-up. The forward test proceeds and is reported as a test of a break-even-or-below signal. Full notice.
|edge| ≥ 6- Wk4: CI lower bound ≥ 50% at n ≥ 10 → green; below → yellow flag + 24h postmortem
- Wk9: CI lower bound ≥ 52.4% at n ≥ 20 → green; below → amber + tier may move to in-calibration
- Wk18: CI lower bound ≥ 52.4% at season end → green; below → red, claim withdrawn from leaderboard
|edge| ≥ 7- Wk4: CI lower bound ≥ 50% at n ≥ 10 → green; below → yellow flag + 24h postmortem
- Wk9: CI lower bound ≥ 52.4% at n ≥ 20 → green; below → amber + tier may move to in-calibration
- Wk18: CI lower bound ≥ 52.4% at season end → green; below → red, claim withdrawn from leaderboard
Pick log
Matchups link to the research post published before kickoff.
| Wk | Matchup | Tier | Side | Model | Market | Edge | Result |
|---|---|---|---|---|---|---|---|
| 1 | SF @ LA | PRIME_TOT | over | +57.3 | +48.0 | +9.3 | PENDING |
Tier rows show the leak-free backtest recompute alongside the running forward (live) figures; the withdrawn 2026-04-30 anchors are kept struck through for the record. Forward columns are blank until games are graded.
Correction 2026-09-08: the backtest figures originally cited in the pre-registration were withdrawn on 2026-05-17 (feature lookahead in the audit harness). The forward test proceeds as a test of a signal whose corrected backtest is break-even or below. Full notice: /insights/nfl-2026-27-pre-registration-correction.
Every PRIME-tier bet that fires per /methodology/2026-27-predictions appears here within 24 hours of the game ending. No historical edits — all changes are append-only.
Each pick links to its corresponding research post under /research/picks where the pre-game rationale lives. A pick is a post published before kickoff; a game that cleared the filter at some hourly snapshot but never inside the publication window is not a pick and is counted separately as an unpublished signal.
Break-even at -110 closing line: 52.4%.