The paper season

The rule

• Spreads only, home-and-away games only. No totals, no moneylines, no neutral-site games in the test (those are tracked in a side column).
• One pull a day, at 11:00 Eastern. Every game where the model's margin differs from the market's consensus line by three points or more is a pick; nothing else is.
• Recorded, not placed: the pick is written down at the best line posted by four named books at that minute (BetMGM, DraftKings, FanDuel, theScore Bet), one flat unit each, with the consensus line beside it. Two more books (Caesars, bet365) are read as extra places the line could be taken, never as part of the consensus.
• Settled on Mondays against the final score and the closing line.
• Closing-line value is the test. Winning a coin-flip season proves little; the market moving toward the model's side between 11:00 and tip-off, pick after pick, is the measurable thing. The rule for real stakes next season reads that record through its lower confidence bound, never the point estimate, and needs about 500 settled picks before it reads anything.

What the model is up against

Against eight seasons of closing lines the same rule hit 54% of the time and the market moved toward it by a tenth of a point on average, which is a thin, real-looking edge that one bad season could erase. The 2026 season, the last one before this record began, was a losing one at the closing line. The paper season exists because that is not good enough to bet on and not bad enough to stop measuring.

Picks are a public record of a model and not advice. Lines are the ones posted at the recording minute and move afterwards. No one connected with Roster Math is paid by a book. The full ledger with every recorded field is published as bets.csv once there are picks.