Wal-Bel AI

Wal-Bel solves team strength from net EPA per play across every game since 2015, using ridge-regularised least squares weighted toward recent form. Ask it the same question twice and you get the same answer — there is no sampling, no persona, no improvisation. That is a deliberate trade against the flexibility of a chat model: less conversational, more repeatable.

A chat interface invites questions the underlying data cannot actually answer, and a language model will answer anyway rather than say so. Wal-Bel's Ask page is built the other way: every question maps to a specific, checkable query against real data, so what comes back was always going to be accurate to what it measured — never a plausible-sounding guess.

Against the closing spread, Wal-Bel runs at 48.7% — a coin flip, and below the 52.4% needed to break even. It is not offered as a betting edge. The full accuracy breakdown is on the Method page.