Start by loading the NFL schedule. No sample picks or results are displayed.
Marty Gwan’s plays of the week
Gold, Silver and Bronze are Marty’s order of preference. Each selection includes his reason, saved sportsbook price and kickoff time.
Explore the saved Hot & Falling Knives board · QB, RB, WR and TE.
Owner studio · choose and publish three plays
Save player selections as drafts in your private scan, then assign each frame. Publishing makes only these three selections public. Games that have started lock the published board.
Reversal Board
Automatically saved · refreshes begin at 9am, noon, 3pm and 7pm Eastern. The last completed scan stays visible while the next scan builds.
Hot Knives: consecutive OVER results. Falling Knives: consecutive UNDER results. Each position and market list is sorted by Knife Sharpness, highest first. Scores measure reversal evidence, not win probability.
Measured scores run down to 0. Unavailable ratings appear last as — and are never converted to zero.
Loading player markets…
Preparing automatic market scan…
Each streak compares completed game stats to the displayed CURRENT line. Historical weekly betting results are not implied. Main lines only; one book/line per player and market. Upcoming eight days.
Historical observations begin when you collect them. Missing archived lines stop the streak.
Sharpest knives of the week
Ranked by supporting evidence, not streak length alone. Usage compares L3 vs prior L3, or L2 vs prior L2 when only four current-season games are available.
Prop knives · trend evidence
Scan coverage
Data coverage: awaiting source audit.
90% means at least 15 of 16 defined checks plus every required check. It is input coverage, not prediction accuracy. Unknown inputs fail; fresh injury-feed access does not confirm a player is healthy.
No scan completed yet.
Your Watchlist
0 saved marketsSaved on this browser and device. These are markets you follow, not locked picks. Quotes follow the sportsbook selected on the board; stale book timestamps remain visible. Projections are experimental estimates.
Marty’s Board
Selections authored by Marty. Saved lines, prices and units stay fixed.
Load published selections to view the board.
Publisher studio · local drafts
These drafts stay on this browser. Assign Gold, Silver and Bronze on the Reversal board, then publish from your signed-in owner account.
Matchup lab
Compare position-group production allowed across NFL defenses.
Last five completed current-season games. Rank 1 allows the most in the selected category; ranks cover only loaded teams with 3+ complete samples.
These are position totals per game, not per-player projections or schedule-adjusted defense grades. Positions use exact athlete IDs in box scores or current rosters. Unknown classifications exclude that game/category. Opponent roles and historical positions may have changed.
Touchdown drought lab
Droughts and TD Knife Ratings are saved together. Refreshes begin at 9am, noon, 3pm and 7pm Eastern; no visitor scan is required.
Games played without a rushing or receiving touchdown. Passing, return and defensive touchdowns are excluded.
Roster players with completed appearances; ratings rank research scoring estimates, conditional on playing. Drought length adds no rating points. Unknown TD totals stop a drought. A drought alone does not increase scoring probability.
TD drought knives · scoring likelihood
Prop Sharpness measures reversal evidence; TD Knife Rating measures an experimental scoring estimate. These scales are ranked separately.
Team trends
SAME NUMBER · RECENT GAMESHow often each team’s recent points and scoring margins cleared this week’s number.
Current sportsbook follows the Reversal board. Scores can span seasons.
Exactly what is a “trend” here?
Team points: compare each final team score with today’s posted TT. An implied TT uses the same-book spread and game total and is labeled separately.
Scoring margin: add today’s signed spread to each past final margin. Positive clears that handicap, negative misses, zero is equal. Historical closing spreads are not used.
These are repeatable stat comparisons at today’s number. They are not historical wager records or guaranteed future outcomes.
Trends to watch this week
Situational records with source links and sample sizes. Verify that the condition applies before treating a pattern as relevant.
Research sources and coverage
Model Lab
Track saved forecasts against final results, then explore the assumptions in a player's Why forecast and Scenarios tabs.
Loading saved forecasts...
Mean absolute error is in yards or receptions; lower is better. Baseline uses the same player's pregame historical average. Bias is forecast minus actual; positive means overprediction. Each market has its own units. Fewer than 20 results is early evidence, not proof of improvement. This is projection accuracy, not betting return.
Recent verified results
Changes on this device
Same game, player, market and sportsbook only. Alerts require a line move of 0.5 or more, or a workload change of at least 10%, between fresh snapshots. First visit establishes a baseline. These are saved changes, not live push notifications.
The 10–0 trend screen
0 verifiedEvery one of the last ten games played must land strictly on the same side. Missing data and equal values stop the streak.
Current-number comparisons and historical pregame observations are distinct. Historical mode requires a dated pregame observation for every game; it is not a wager record. No scan invents missing lines.
Weekly research recap
Fresh pregame candidates are captured on this device after completed scans. One first observation per game/player/market/book. Different books may duplicate a player; this is an observation count, not a unique-player rate.
Results use the exact line and streak saved BEFORE kickoff, not a later line. Only exact ESPN event matches are graded. Both archives contain observed samples; no missing weeks or simulated results are filled in. Device observations are local until the publisher deploys RESEARCH_ARCHIVE_JSON.
Knife Sharpness · results by rating
PREGAME SCORES · VERIFIED RESULTSDoes stronger reversal evidence lead to more reversals? Compare saved scores with final production at the saved line.
The earliest valid pregame capture per event/player/market is used, with ties settled by book name. Different books do not inflate the result count. Only the current score version is included. Equal outcomes and pending results are excluded from the rate. These are observed research outcomes, not wagers or profit.
How to read these results
A reversal ends the captured streak: an UNDER streak reverses if the final stat is above the saved line; an OVER streak reverses if it finishes below. Reversal rate = reversed / (reversed + continued). A score of 75 is not a 75% forecast.
The 95% interval uses the Wilson formula to describe sampling uncertainty. Games and player markets may be correlated, and the saved sample is selected rather than random. The interval does not correct these biases. Bands below 20 decided outcomes are marked provisional; 20 is a display threshold, not proof of reliability.
Published site data comes from the configured archive; device data comes from this browser. An empty archive stays empty. Publishing, refreshing and grading follow the existing recap workflow. No odds-based ROI or TD calibration is claimed here.
Closing-line coverage
Loaded player markets only. Closing snapshots must match the selected book and game; missing closes remain ungraded. Open a player to load eligible historical closes.
| Player / market | Book | Verified closes | Missing |
|---|
Workload projection error by market
Current-season walk-forward checks: each forecast uses only earlier games. Duplicate player/market/book observations are removed. Error is in yards or receptions, not a win rate. Small, selected samples remain provisional.
| Market | Forecasts | Workload mean absolute error | Simple-mean baseline error |
|---|
This week’s matchup projections
Team scoring baselines and projected workloads for the hot/falling knives matched to each game.
Game-log import
Automatic scanning loads game logs. Optional manual JSON corrections can also be imported. Values must use the same market and units as the prop. Dates run across seasons; bye weeks do not count. Historical lines must identify a sportsbook.
Logs and picks stay in this browser. Export regularly to move between devices. Historical odds backfill and cloud tracking remain future integrations.
Import shape
[{"player":"Player name","market":"player_receptions","games":[
{"date":"2026-09-20","value":4,"line":5.5,"book":"fanduel", "targets":8,"routes":31,"snaps":52}
]}]line, book, targets, routes, snaps, carries, and attempts are optional. Only verified completed games should be imported. The example is a schema, not a real player's data.
Loaded market inventory
V4 prospective model tracking
Saved pregame forecasts will appear after the background scheduler captures them.
Locked research picks
Save a candidate from its player page before kickoff. Enter the official final stat to grade it. Saved price and line remain fixed.
Official completed-game stat matching. Unmatched picks stay pending.
The streak is a filter—not a reason to bet
A long streak alone does not increase the probability of reversal. V1 screens streaks and provides a smoothed empirical estimate from your imported game logs. This is not yet a validated NFL projection model.
- Games are sorted newest first, with L10 shown in player charts.
- 100,000 browser simulations draw from a Beta posterior using all imported games, with one prior OVER and one prior UNDER. Each trial samples an outcome from that probability. Pushes are counted separately.
- This estimates the historical chance of exceeding today's line under exchangeable conditions. It does not adjust for opponent, injuries, weather, or role changes.
- The projection is the average of imported game values. L3 versus prior L3 usage is shown only when both windows have complete inputs.
- Fair odds and expected value use the selected reversal side. Push probability is included in expected value.
- No-vig comparison is available only when both sides share the same book and line.
- Missing game logs block projections. Missing historical lines block historical streaks. No line means no priced edge.
- Outputs are research estimates, not recommendations. Calibration and out-of-sample evaluation must precede confidence labels.
Knife Sharpness methodology
Balanced Sharpness (v3): usage, efficiency and current-season window mean contribute up to 40, 30 and 30 points. A score of 50 is neutral; higher values support the reversal side and lower values oppose it. Components use a smooth directional logistic scale, with 25% usage/efficiency and 20% mean-gap reference changes. Incomplete ratings display —, not zero. These research defaults are not calibrated win probabilities. Old saved ratings keep their original scale and remain in overall tracking.
Next model layer
Position-specific opponent allowances, injury/role impact modeling, unattended line collection, multi-device pick tracking, and out-of-sample reversal calibration. The current matchup panel supplies context; it does not adjust the Sharpness formula.
Install Knife Lab on your phone
iPhone: Safari → Share → Add to Home Screen. Android: browser menu → Install app / Add to Home Screen. Offline views contain saved data only; reconnect and rescan for current markets.