I was listening to the Tony Kornheiser podcast a while ago and he was raising a suggestion of going after a Sabermetrics coach for the Redskins. NFL scouting is less data-oriented than baseball and basketball, and Kornheiser was suggesting maybe a Billy Bean like figure for the Redskins to shake things up.
The suggestion got me thinking. Sabermetrics for scouting is easy. But could a machine do play calling better than an offensive coordinator? Not something like the play calling in Madden but a real data analytics based solution. The nature of football with stoppage between plays makes it a natural fit.
Here is how it could work. It requires a lot of data that isn't readily available today, but maybe one day. First, offline, gather as much historical data on individual plays as possible. Each play contains features such as:
- Yards to go
- Field position
- Defensive formation
- Defensive personnel (maybe even down the individuals)
- Defensive coordinator/team
- Offensive formation
- Offensive personnel
- Offensive play
Let's say you could gather the above for every single play for your team in the past X seasons. You could in theory learn a function that takes the above input features and outputs yards gained. There is a question of how much data you have or need (especially if you want to build a different model per opponent). But let's suppose you get it done.
Then, at runtime (aka play calling time), some assistant coach can enter in all observable features. They would have to watch the defense line up to get the formation and personnel. This is a bit tricky since the defensive usually waits for the offense. Anyway, let's assume it's doable. Then it's a simple matter of applying the model to every single play in the playbook and seeing which one returned the most predicated yards gained.
This offense would be pure data-driven, no emotions. Will probably go for it on 4th down more than a human being. But somehow I feel like it could revolutionize the game, even more than Sabermetrics did to baseball.