The 2026 Draft Board
Built by a machine that studied 14 seasons of NFL history, weighs ten signals on every player, prices its every opinion against the live draft market, and retrains itself weekly. It doesn't guess 10,000 times and call that science, it shows its work once, in public, losses included.
player-seasons studied
vs the market, 10 replayed drafts
priced against live ADP
self-retraining through drafts
Showing the top 200 overall, position tabs go deeper. The CSV cheat sheet always includes the complete board.
Value badges compare our positional rank to the market's (ADP snapshot 2026-09-08). In the ALL view players follow market draft order, cross-position strategy is your league's call; the badges and tiers are ours. Projections are anchored to sportsbook season futures (snapshot 2026-09-15); a NO LINE tag means the books currently decline to post this player's season markets.
Inside the machine, and its audited record
A word about "we simulated the season 10,000 times": anyone can run arithmetic ten thousand times, simulation volume is marketing, not skill. The only question that matters is whether a process beats the market out-of-sample, and that's the only thing we publish.
Every player gets a season projection built from opportunity and efficiency treated separately: projected touches and targets (sticky) times points-per-touch (regressed), with touchdown rates re-anchored to yardage, because last year's TD luck is the market's favorite mistake, plus team-level context computed from roster arithmetic: targets and carries vacated by departures, mouths added by arrivals and drafted rookies, and each offense's projected volume. Aging curves come from 14 years of data; rookies get draft-capital priors. Then we blend our ranking with live ADP per position, weights fit on recent seasons out-of-sample.
The honest results, because we tested a full decade and publish the losses: across ten replayed draft seasons (2015–2024) the blended board went 7–3 against drafting straight off ADP, averaging +0.78% more value per season (2024, the latest graded year: +2.49% out-of-sample). Three of the ten years were losses, they're in the published backtest, parameters and all, because a record you can't audit is marketing. The board leans on the model where it has earned trust (quarterbacks and running backs, after touchdown-luck correction) and defers to the market where it hasn't (the TE board is the market, on purpose). Floor and ceiling ranges come from real backtest errors, not vibes. Full methodology on the methodology page.
Two more things we publish because they're true: the decade backtest is graded in PPR scoring, re-graded without re-tuning, the same board is roughly market-even in half-PPR and standard (numbers in the published backtest), so treat the edge claim as a PPR claim. And we also built injury-history and QB-change features, tested them the same way, and rejected them when they failed out-of-sample, the lab log of every configuration we tried ships with the receipts. This board freezes with every weekly rebuild and gets graded against actual season results after the year, the same public accountability as our game predictions. Somewhere between a spreadsheet and a precog.