The rule on this side of the vertical is walk forward, beat the closing line out of sample, or don't publish. The spread and totals model — quarterback and efficiency levers included — ran that test over six seasons and didn't beat the bar. So there are no NFL game picks here, and the report below is the reason why. The DFS optimizer is live now, running off the validated projection engine; props are the next lane.
Against the spread over 1,610 games, 2020-2025 regular seasons (holdout — never touched in training or feature selection). Standard prices need 52.38%to break even; the model landed short. The closing line also beat it on log-loss. That's a failed gate — published, dated, and left up.
Every season in the window was predicted from information available before it started — no peeking, no refitting on the games being graded. Lower log-loss means the better probability. Walk-forward validation of the rebuilt (v2) sides/totals game model: a margin-Elo logistic side model and an OLS scoring-environment totals model, regular-season games only (the rebuild fixed a protocol bug — postseason games had contaminated the previous "regular season" holdout), trained 2010-2019 with every selection decision made on pre-2020 walk-forward folds under a pre-registered adoption rule, evaluated on the untouched 2020-25 regular-season holdout against de-vig closing market prices over play-by-play since 1999. The gate: the model blend must not lose to the market on log loss, and model picks must clear 52.38% at -110. Disclosure: on pre-2020 walk-forward folds the closing market beat every model variant tested, so the pre-registered adoption rule assigned the model zero blend weight — the blend is the de-vig close itself and the log-loss legs tie by construction, not by model skill. Result: FAILED — the pick streams do not clear breakeven on the holdout. No game-line product ships from this model until a version clears this gate on a future holdout. Published because the record is the product: a model that does not beat the market says so, with the numbers.
The closing line was the sharper number — lower log-loss over 1,610 games — and the model's 50.10% against the number sits under the 52.38% a bettor needs to break even at standard prices.
The closing line was the sharper number — lower log-loss over 1,599 games — and the model's 48.80% against the number sits under the 52.38% a bettor needs to break even at standard prices.
The verdict stands until a model beats this bar out of sample, and that report will hang in this same spot. A failed gate isn't an embarrassment — it's the product working: a number that can't beat the close never becomes a pick you paid for.
Our number against the price on every game. The spread and totals model was tested against six seasons of closing lines and lost — the full report is on this page, because the rule is publish either way. Win totals live on the season-sim board, priors only, labeled as such.
GATE FAILED · REPORT PUBLISHEDOpen →Props price player-level questions, and player-level is where a projection engine earns its keep. One projection source feeds props, DFS, and the weekly read. This is the lane when the engine ships — the game market already told us what it is.
SHIPS WITH THE PROJECTION ENGINEThe same optimizer that runs golf, MLB, WNBA and UFC — full-pool projections with floor/ceiling bands off the validated engine, salary-cap and exposure controls, CSV export. DST slots are anchored to the market's salary curve and say so on every row. No pre-built lineups, ever.
LIVE · LOW-STAKES UNTIL WEEK 1 (SEPT 10)Open →Week-by-week survivor strategy: future value, pool size, and when the contrarian pick is the correct one. The season sim's weekly win probabilities power an informational primer.
PRIMER LIVE · INFORMATIONALOpen →The Coaching Decision Index and the cap boards run year-round — free agency, the draft, camp cuts, the trade deadline. That's where the offseason edge lives.