NFL Big Data Bowl 2027
Does how fast a player builds speed in the first second of the 40-yard dash predict how he moves in NFL games, beyond the official 40 time? A reproducible tracking pipeline, cohort-held-out models and an honest answer: mostly no.
Kaggle competition · Oct. 2026
The question
The NFL Combine times every prospect over 40 yards. The official numbers are the 40 time and the 10-yard split. The Big Data Bowl 2027 also gives the 10 Hz tracking of each dash, so the whole speed curve is visible. This project asks whether the shape of that curve, especially how speed builds during the first second, adds information about a player’s later NFL movement beyond the 40 time, the split and his body measurements.
What was built
- A 10 Hz Combine tracking pipeline: early acceleration, later acceleration, speed retention and position-aware drill profiles.
- Data-quality and linkage audits: trace cadence, movement onset, course-axis assumptions, repeated attempts and Combine-to-game player matching.
- Player-level models (ridge regression) evaluated leave-one-draft-cohort-out, with preprocessing fitted on the training cohorts only.
- A nested feature-selection audit, measurement-reliability analyses, outcome and role sensitivities, and a registry of every analysis tried.
- A reproducible Kaggle notebook with pinned requirements and a source bundle.
What it found
The honest headline is mostly negative. For defensive linemen, first-second acceleration is associated with the 90th percentile of rookie peak speed, but that result came out of a feature search and is not confirmatory. The broader selection-aware search showed little average lift over the official 40 time and split, and the direct first-step blitz-speed analysis did not show added predictive value. The project documents these negative and uncertain results instead of turning them into a scouting score.
Limits
- Only three draft cohorts, and 2025 was used in feature screening, so there is no untouched confirmation set.
- Rookie tracking exposure depends on playing time, role, health and assignment.
- Movement onset is a speed threshold, not a reaction time: the data have no start-signal timestamp.
- The competition data and player-level outputs are not redistributed (Kaggle competition rules).
Stack: Python · pandas · scikit-learn · NumPy · Jupyter · Kaggle