BallKnowMe college football, by the numbers

The record power-l2-v1

Hit rate counts decided picks only — wins and losses. A push (exactly on the number) and a void (game postponed, cancelled or forfeited) are neither, so they are shown separately and never counted as a win. Pending is a game that hasn’t been played; it is not a loss either. Picks are written to an append-only log before kickoff with the line they were made against, and they are never edited afterwards — one pick per game per market, made once, graded once.

Open picks logged, waiting on kickoff

The picks above aren’t a promise — here they are. Each row was written to an append-only log at the timestamp shown, against the line shown, before the game was played. Nothing here can be edited afterwards, so you can check these against the results yourself rather than take our word for the record later. Showing the next 20 of 140 priced picks.

What the model did on history not the record

This is a backtest, run point-in-time on 2023 — the model re-fitted week by week on prior weeks only, then graded against closing lines. It is shown for calibration, and it is deliberately kept out of the record above, because a backtest is what we knew afterwards and a record is what we said beforehand.

Read it honestly: picking winners straight up looks strong, but that is mostly agreeing with the favourite, and the favourite is already priced. The two markets you would actually bet — spread and totals — land below the 52.4% needed to break even at −110. That is the same conclusion as the full audit across 10,777 games from 2016–2025: 49.5% against the spread, 49.5% on totals, every season under break-even. We have no edge on these markets and we don’t claim one.

Experiments running not picks

Ideas being tested forward with real prices attached, kept completely out of the record until they earn their way in. Logged before kickoff and graded the same way, so if an idea fails it fails in public.