2026 · Week 6

A private NFL workbench

NFLViz is a desk for looking at player props and 11-on-11 matchups before kickoff. It is not a tout sheet, not a live-odds board, and it will not pick the winner of the game.

Start with a game or the field. The prop board is there when you want the model's ranked TDs and receptions — and it will tell you when the sportsbook is already right.

Browse this week's games Open the 11-on-11 field

What you can do here

This week's games

Pick a matchup. You'll see the player props the model will rank — plus a preview of both 11-on-11 units.

Open the slate →

11-on-11 workbench

The field. Swap formations, edit who is on the grass, leave notes. Skill colors are Madden ratings with a vintage stamp — not live grades.

Open the field →

Prop board

Ranked touchdowns, passing TDs, and receptions. Yards markets stay off this list on purpose — the book already priced them.

See this week's picks →

Watchlist

Save a prop to check later. Line and model movement show up here. Still not a ticket to a sportsbook.

Open watchlist →

A simple walk-through

Scan Picks for model-ranked player props.
Open Matchup to see 11-on-11, try a formation, or swap a player.
Open a card for comps and sample size before you treat a Take as an action.
Add ticketable legs to Parlay when you're ready.

How it was built

Features first. Then the process that earned them.

Assemble the tape. Years of official play-by-play, roster depth, and a stamped Madden ratings vintage — not a live scouting service.
Train leak-safe. A Ridge model on 2023–2024, tested on 2025. Only information you could have had before kickoff is allowed in.
Ask where the book still misses. Continuous yards (rush / rec / pass) are already in the number. Discrete scoring — who scores, who catches enough — still leaks player-role structure.
Keep only what survived the tests. Weather, injuries, line movement, and depth features were tried. None of them earned a yards-market promotion. Yards now go through the same calibrated TAKE / FADE rule as every market, and each yards card states the measured reason it is not a pick (live model AUC ≈ 0.50, or k ≈ 0).

That is why yards cards read NO EDGE or SKIP today, with his yards in similar games beside them — the product will not pretend it beat the yards line.

The research notes

Sportsbooks already priced the yards. The residual lives in who scores.

Read the full methodology and glossary →

NFLViz is not a “beat every line” engine. It is a leak-safe, frozen probability layer that asked a blunt question: where does the book still leave player-usage structure on the table? The answer held through three independent attacks. That is the finding — and the reason yards, which now go through the same calibrated rule as everything else, rarely if ever reach this feed.

What the model actually is

A per-market logistic regression on leak-safe pregame features (baseline_logistic) gives the hit probability; a Ridge model gives the projected stat. Validated by training on 2023–2024 and testing on 2025; refit on every past season for the current one. The same model covers QB markets — nothing here is calibrated to the book. Edge is model hit probability minus the implied probability of the posted American price — not a gut feel, not a recency narrative.

Similarity comps explain a pick; they are not a second model. The 11-on-11 field is roster structure (skill hue + pair rating Δ), never a scored market. Nothing on this page is a moneyline or spread call.

Why the theory is worth cheering

Continuous yards are the book’s home field. Volume, snaps, and scheme are already in the number. Discrete scoring — who finds the end zone, who catches enough to matter — still leaks player-role and matchup structure the line does not fully eat. Anytime TD ranking sits near AUC 0.77. That is not magic. It is the shape of football: scoring is rarer, noisier, and more role-driven than “will this back go over 64.5.”

We tried to steal the yards markets anyway. Weather, injuries, line movement, and 2026 depth features were added leak-safe and tested. No AUC lift survived bootstrap. No confident-decile ROI cleared zero. The line won. Since 2026-10-01 yards are decided by the same calibrated rule as every market instead of a blanket label, and each card says the measured reason it is not a pick — plus his yards in similar games, never anchored to the book.

Live honesty table

Test-set AUC from the serving payload · 2026 season. Every market goes through the book-calibrated rule: below AUC 0.52 the model cannot back a TAKE or FADE, and the k measured on 2025 shrinks the rest toward the book's price.

Market	Test AUC	What this page does
Anytime TD	0.77	Calibrated TAKE / FADE
Passing TDs	0.59	k = 0.00 — shrunk onto the book
Receptions	0.58	k = 0.01 — shrunk onto the book
Passing Yards	0.55	k = 0.00 — shrunk onto the book
Receiving Yards	0.50	Below AUC gate — no TAKE / FADE
Rushing Yards	0.51	Below AUC gate — no TAKE / FADE

Comp-space closed. Retuning similar-situation features did not produce a new serving model. Comps stay as labeled context (sample size always shown).
Line-movement leak fixed, numbers stood. After dropping post-kickoff “close” snapshots, every AUC moved ≤ 0.0008 and no market changed side. WS2 had already shown movement is not edge — cleaning a corrupted feature should not flip the table, and it did not.
New information, same gate. Injuries, weather, and line movement as pregame features failed the honesty gate (bootstrap AUC lift and priced ROI). Depth-chart retrain: all markets HOLD. We do not wire a HOLD study into this feed.

Ranking skill is not a promised bankroll. Open a card for comps and sample size before you treat a Take as an action. The theory we will keep defending: say where the book is already efficient, and only rank the residual. That is how this app stays useful in Week 1 without pretending it found alpha it did not measure.

Skip to this week's ranked picks →