NFL Big Data Bowl 2027 First findings · October 14, 2026
Three cones.
Tight coverage.
Can change of direction at the combine tell us who stays close on Sundays? I'm starting with cornerbacks, the three-cone drill, and a new stat: Blanket.
Introducing
Blanket
average separation < 1 yd
One man-coverage assignment. One receiver. Less than a yard of separation on average.
ReceiverCornerbackConcept illustration
A stat built around staying close
What counts as a Blanket?
For this first version, I average the distance between a cornerback and the same receiver from the snap through the throw. A time-weighted average below one yard earns one Blanket. Exactly one yard does not.
A receiver can briefly get more than a yard away and still be blanketed on average. Being close only when the ball is thrown does not guarantee a Blanket.
Blanket rate is Blankets divided by eligible, measured man-coverage assignments. Missing tracking is excluded from the denominator; it never counts as a failed rep.
Try the definition
How tight is tight?
Change the threshold or the averaging window to explore sensitivity. The proposed Blanket definition stays fixed at under 1 yard, snap to throw.
Average separation < 1 yard, from snap through throw.
The separation distribution
Most measured reps sit above one yard.
Each bar counts pairings in a range. Ranges have different widths; this is a count chart, not a density plot. Values use the snap-to-throw time average.
Median rep average: 1.8 yards. Bar widths show a share of all 215 screened pairings.
The first football question
Should starting cushion count?
Omitting the first half-second changes the result from 2 to 12 reps under one yard. Starting cushion can penalize a cornerback before the receiver has made a move.
At the throw alone, 71 pairings are below one yard. That snapshot cannot tell us how tightly the whole route was covered.
Combine → coverage
Does the three-cone drill explain it?
We don't have convincing evidence yet. Among 8 cornerbacks with at least 10 screened pairings each, three-cone time and player-average separation have Pearson correlation r = 0.33. Longer times trend toward more separation in this sample, but the player-bootstrap 95% interval runs from -0.98 to 0.94. It spans strong relationships in both directions.
The sample is the constraint.
The audit found 60 cornerbacks, but only 14 have recorded three-cone times. Repeated snaps don't create extra independent combine observations. The correlation is unadjusted for speed, receiver quality, route, cushion, or scheme.
Use the movement, too.
I extracted a prototype turning-speed feature from three-cone sensor tracking. Only 4 players clear its data requirements and the coverage workload filter. That's a feasibility check, not a validated scouting metric.
The box, W, and two backpedal/break drills have tracking for all 60 cornerbacks in the audit, which could support a broader test of the same change-of-direction idea.
Follow the denominator
What made it into the prototype?
| Reason | Reps |
|---|---|
| Nearest tracked route runner more than 5 yards at snap | 578 |
| No tracked route runner at snap | 568 |
| Less than one second before throw | 2 |
| Ambiguous snap pairing: margin below 1 yard | 8 |
| Missing or ambiguous defender snap/throw | 9 |
The big limitation: only selected players are tracked. 135 of the 215 accepted pairings have one observed route runner at the snap. The closest tracked receiver may not be the CB's actual responsibility. This is why the page reports provisional counts and does not rank players.
Methodology · version 3
How the first pass works
- Choose the cohort. Join combine and NFL players by ID and draft year. Keep CBs with a three-cone time, then regular-season passing plays with CB/SLOT_CB alignment, team MAN_COVERAGE, player assignment MAN, and no nullified penalty.
- Infer a receiver at the snap. Choose the nearest tracked offensive route runner within 5 yards, with a 1-yard margin to the next observed candidate when there is one. Freeze that identity for the rep. Reject receivers paired to multiple measured CBs.
- Check the window. Require unique snap/throw events, at least one second of coverage, both endpoints, at least 90% of expected 10 Hz samples, finite positions, and no synchronized gap above 0.2 seconds.
- Measure separation. Integrate Euclidean distance between simultaneous x/y positions using the trapezoidal rule, then divide by elapsed time. Coordinates are treated as yards under the NFL tracking convention; the final study must confirm units against the 2027 field dictionary.
- Count and compare. A time average strictly below 1 yard is a provisional Blanket. Give each player equal weight in the combine correlation and resample players, not snaps, in 2,000 bootstrap draws.
Three-cone sensor feature
Within each attempt, measure median speed at frames with a circular heading change of at least 120 degrees/second. Require both adjacent speeds of at least 1, timestamps 0.09–0.11 seconds apart, and three qualifying frames. Take the median across attempts. These exploratory thresholds and drill phases need sensitivity checks.
This first pass has not verified assignments with film, fit a predictive model, or tested incremental value beyond speed. The 5-yard filter favors tighter starting alignments. Results describe this screened sample, not all NFL coverage.
Next in the notebook
Verify the matchup.
Then test the movement.
Five tracking examples are prepared for private review. Next: check assignments against play video, test cushion and window choices, and expand to drills with broader participation.