About Insularity NFL Forecast
Insularity NFL Forecast is a sports-analytics portfolio project: a reproducible pipeline that forecasts every NFL game and publishes its forecasts before kickoff, then scores them honestly. The aim is prediction quality and transparency. The site shows projected outcomes, uncertainty and how each forecast compares with the market line, all graded openly; it is for analysis and education and gives no advice.
Design decisions
- Benchmarks first. Market-only, football-only and naive models were built and evaluated before any combined model, on identical games and cutoffs.
- Honest validation. Walk-forward folds, a tuning period, a development period and a locked test season evaluated once. The headline finding is that the market line is very hard to beat; that result is reported rather than hidden.
- As-of data. Every feature is built from what was available before the cutoff, and raw data snapshots are append-only.
- Frozen releases. Forecasts are immutable JSON files with model, data and code versions; the website only renders them.
Stack
Python (nflreadpy, Polars, scikit-learn, NumPy) for data, features, models and evaluation; versioned JSON as the interface; Next.js and TypeScript for this static site.
Inspiration
The weekly board layout is inspired by the clarity of davidsasser.com/nfl. This project is independent: it does not reproduce that site's model or assets and makes no claim about how it works.