Methodology
A plain-language walkthrough of how a prediction on this site actually gets made - and what it doesn't mean.
1. Quantitative models
Every predicted score, win probability, and spread comes from versioned statistical models (Elo, Ridge regression, LightGBM) trained and backtested on historical NFL data. We keep an independent model, which never sees betting lines, structurally separate from a market-aware model, which can.
2. Backtesting
Before any model goes live, it's walk-forward backtested on sealed, previously-unseen seasons - never trained and tested on the same data. We publish that record honestly, including when the market beats our models.
3. Market comparison
We pull real sportsbook lines and compute a no-vig consensus across multiple books. The gap between our models and the market is one input to our decision engine, never something we chase on its own.
4. Research
A research step checks for information the structured data might miss - injuries, coaching changes, weather - and separates verified facts from unconfirmed reports and opinions. It never produces its own competing prediction or overrides a model number.
5. Decision engine
A deterministic, rules-based engine - not a human, not an LLM - combines the model, the market, and the research into one of five outcomes: No Bet, Watch, Lean, Best Bet, or Veto. Zero qualifying bets in a week is a valid, expected result.
6. Verified tracking
Every published pick is recorded in an immutable, append-only ledger the moment it's made - line, price, and reasoning included - and is never edited after the fact, win or lose.