If you’ve been following along, I manage my Fantasy Football team with a custom NFL data set that I’ve been compiling over at https://gridirondata.com. As it stands we are 3-0 on the season! Our week 3 matchup was very close and we won by maybe 1 point. Unfortunately I feel as though it might be on the back of player injury.
Overall most of our players did well. Josh Allen was not anywhere near his projection but other players helped him out. We are going to continue the streaming defense strategy for the season as it seems to work pretty well. We also claimed a few waivers so some new faces for week 4.
Vikings defense for week 4 against Miami. There are some bench players not pictured here to cover for injuries if we have any later in the week.
Injuries in the NFL have been running rampant this year. It’s only week 4 and we have starting quarterbacks out for the year. This starts to ask the question about how can we predict the possibility of injury in game. For example, if we were starting Sam Darnold in week 1, was it possible to determine that he was going to go out with injury?
GridIron Data doesn’t ship an “in-game injury” field, but it turns out the data to build one is already there. In-game injuries leave a clear mark in snap share: a receiver who usually plays 85% of snaps logs 22% and then sits the next week. By flagging those sudden drops and confirming them with the next week’s game log, injury status, or injury-tagged news, I can generate thousands of labeled player-games from 2020 through 2026 without a dedicated injury dataset.
From there, the features mostly come for free: workload and special-teams snaps, the player’s injury status heading into the game (a Questionable tag is probably the strongest signal), prior missed games, and recent jumps in usage. Adding nflverse’s historical injury reports, schedules, and birthdates would bring in practice participation, short weeks, turf, and age, which should sharpen things considerably.
The goal isn’t a crystal ball. In-game injuries are rare, so the right output is a well-calibrated risk tier (“this RB is about 2–3x baseline this week”) that can feed into projections as an expected-value adjustment, rather than a yes/no prediction. Scoring would run as a weekly Lambda job alongside the existing injury refresh, exposed through a new risk endpoint.
The bigger opportunity may be real-time: tightening the news pipeline during live games to catch “questionable to return” headlines and push them through the injury changes feed, so Pro users know about an injury before their league does. First step is a backtest of the snap-share labeling to see whether the signal is real. I’ll share what I find.
Follow along and never miss an update by subscribing to the mailing list.
Leave a Reply