Before the Model Knows Anything
Defining a useful baseline for an NFL Week One prediction project.
Week One is an attractive prediction problem for almost exactly the reasons it is a difficult one.
The rosters have changed. Coaches have changed. Preseason signals are incomplete, and last season’s data is both valuable and stale. Before choosing an algorithm, I need to decide what a correct prediction would demonstrate—and what it could not.
Start with the argument, not the platform
I am building the project in Databricks Free Edition, but the platform is not the lesson. The first useful artifact should be a baseline simple enough to distrust intelligently.
That means documenting which information would genuinely have been available before kickoff, separating prediction from retrospective explanation, and resisting the urge to interpret complexity as rigor.
A public prediction needs a public clock
Predictions will be recorded before the games. The retrospective will include the misses, not only the screenshots that aged well.
If this becomes a useful teaching project, it will be because another person can see where judgment entered—not because the final notebook contains an impressive model.