How these projections are made

Every number on this site comes from models that predict raw stat lines — yards, touchdowns, receptions — never fantasy points directly. Points are computed from those stat lines under your scoring rules, so PPR, half-PPR and standard all come from one prediction.

The headline model is a small quantile transformer (PyTorch) that reads each player's recent game log and the matchup, and predicts a floor (p10), median (p50) and ceiling (p90) for every stat. The range bands you see on the draft board and the weekly page start from those raw quantiles. In the 2023–2025 bake-off the weekly bands were calibrated per position — widened or narrowed so measured coverage tracks the 0.80 target — and that measurement is reported under “Historical band coverage” below, for the scoring and setup it was run with. It is not a claim about any particular league’s scoring: the in-season pages score a league’s own settings with sign-coherent component bands, which are not calibrated coverage for custom scoring. Season bands on the draft board build on the weekly band through a season-long simulation (more on that below); uncertainty is the product, not an afterthought. The deployed projections average three independently seeded training runs of that model (a small ensemble), and the bake-off table below scores exactly that ensemble. Classical baselines (a naive last-four-games average and XGBoost) run through the exact same evaluation harness, and the tables below report whichever model wins, exactly as it lands.

Evaluation is walk-forward only: train on seasons up to S, test on season S+1, held out across 2023–2025. No random splits — rolling features would leak the future into the past.

The draft board is consensus-anchored within each position: at every position players are ordered by FantasyPros expert-consensus ranking, because in walk-forward tests the model's own season-long ranking lost to both expert consensus and ADP. Across positions the board is ordered by VORP — value over replacement — which is what prices positional scarcity, so a tight end can sit above a receiver with a better raw consensus rank. The model earns its place with floor/ceiling bands (historical coverage is reported below, with its scope) and that value curve — the parts ADP and ECR don't give you.

Known limitations: undrafted rookies are not projected; drafted rookies get a draft-capital prior on the board (see Rookies below), and rookies enter weekly projections only once they have NFL games; players who changed teams keep their personal history with updated matchup context; retired players may appear until rosters are pruned; season projections seed every week from end-of-prior-season form, so a short late-season role spike (an injury fill-in) can be over-extrapolated across a full year while a star's slow finish under-ranks them — expect the board to have loud recency opinions; in the historical bake-off the weekly bands were calibrated per position so roughly 80% of actual player-weeks landed inside them, measured on 2023–2025 held-out seasons under the bake-off’s PPR setup (not a measurement of any league’s own scoring); season bands come from a Monte-Carlo simulation — games played sampled from historical position availability, weeks linked by a measured week-to-week correlation — and measured season coverage is measured per position against the 0.80 target, and we're saying so plainly rather than rounding up; snap-count data has no 2012 source and is treated as missing for that season. These are design choices, documented, not surprises.

Result, 10 October 2026. Measured against same-week expert rankings, under rules fixed before this reanalysis was run, the model is behind weekly FantasyPros consensus at ranking players within position. Rank correlation with actual outcomes is 0.597 for the model and 0.634 for consensus on 2023–2025: a difference of −0.037, 95% interval −0.047 to −0.028. It is behind in all three seasons and in every leave-one-season-out estimate, and the separate 2020–2022 seasons show the same (0.556 vs 0.601). It is behind at every position, running back included, so the running-back edge described below is retracted. Scope: players matched to the consensus who recorded a stat line and had not yet played when the rankings were scraped. The interval resamples weeks within these seasons and does not model dependence between weeks of a season. This is a reanalysis of seasons we had already examined, not an untouched replication. Numbers in models/diagnostics/weekly_consensus_sameweek.json.

Why the earlier numbers were wrong. The expert comparisons below used the wrong week’s rankings. The expert feed is scraped on Fridays, so the last ranking before a Thursday kickoff is the previous week’s list, made before that week’s Sunday games and injuries. Of the 38 weeks in 2020–2025 where the bye schedule can tell weeks apart, the old comparison’s rankings contradicted the week they were scored against in 36; one was consistent with the right week and one week had no ranking. With the Friday scrape dates, that supports the explanation that the experts were scored a week behind. The tie and running-back edge below are invalidated by the corrected analysis; the original write-up stays as the record and is superseded by the result above.

Historical results: measured on past seasons by the diagnostics named below, not a claim about the current output or about any league’s own scoring.

Season-long draft ranking loses to expert consensus, as stated above. Weekly ranking is measured separately, and overall it's close: across all four positions on 2023–2025, the model's weekly within-position rank correlation with actual outcomes is a statistical tie with weekly expert consensus (Spearman 0.5940 vs 0.5958; delta −0.0018, paired 95% interval [−0.0123, +0.0083], which includes zero). That's parity, not a win, and the draft board still defers to consensus for the reason given above.

Scope of that number: 172 position-weeks and 12,182 player-weeks, covering weeks 3–17 in 2023 and 2025, and weeks 5–17 in 2024. Eleven of 54 season-weeks are missing because the expert feed carries no ranking scraped in the seven days before kickoff, and we will not reach past kickoff to fill a gap. So the tie is a tie on the middle of the season, not on a full one. Numbers in models/diagnostics/weekly_consensus.json.

One position breaks the tie. At running back, weekly within-position ranking beat weekly consensus on the 2023–2025 seasons where the effect was found (Spearman 0.7215 vs 0.6958, won 3 of 3 seasons, p=0.002). Finding an edge on the seasons used to build a model is the easy part; the test that matters is whether it holds on seasons the model never touched.

It did. Before evaluating, we pre-registered the pass/fail rule — RB ranking has to beat consensus in at least 2 of 3 seasons, with a positive pooled advantage — then ran it on 2020–2022, seasons disjoint from the ones that found the effect and untouched by any of the model folds involved. RB ranking held: Spearman 0.6658 vs 0.6415, won all three seasons (2020 +0.0298, 2021 +0.0274, 2022 +0.0170), pooled delta +0.0243, paired bootstrap 95% CI [+0.0066, +0.0437] — excluding zero.

This is one position out of four, and it should be read that way. Run the identical test on the others and QB loses to weekly consensus in both samples — −0.0631 on the 2020–2022 replication and −0.0404 on 2023–2025, the latter with an interval excluding zero. TE is flat (−0.0144 then −0.0013) and WR a shade positive (+0.0027 then +0.0088); on 2023–2025 neither is separable from zero. RB is the only position positive in both samples, and the only one we have measured an interval for in each. The 2020–2022 per-position figures other than RB are point estimates without intervals, so their signs replicate but their significance is not established. The effect concentrates on the position most driven by workload — carries, targets, snap share — which the model's usage features encode directly; that specificity is why we treat it as a real signal rather than a scoring artifact. It does not generalize past RB, and it does not change how the draft board is built. Full numbers in models/diagnostics/rb_oos_weekly.json.

Simulation is off: the 2025-season test failed (on synthetic 15-player rosters), and the per-format 2026 test reports in January 2027.

Historical context, scoped: the walk-forward test on 2023–2025 used two simulated-draft formats (Gabagool Fools and FAM FOOTBALL) only. Its 80% intervals covered 65–67% of outcomes in those formats, and its waiver drop choices beat the existing method in one format but not in the second. Nothing here is a measurement of your league.

Drafted rookies have no NFL games to model, so they get an empirical prior instead: we group every drafted QB/RB/WR/TE since 2012 by position and draft capital (top-12 picks, rest of round 1, round 2, round 3, day 3), and a 2026 rookie inherits the distribution of what players with his draft profile actually did as rookies — including the ones who never played a snap. That zero-games outcome is kept in the math, which is why rookie floors are low and rookie bands are wide.

This makes rookie ranks look conservative next to draft-hype ADP — historically, that conservatism has been the accurate call on average, and the p90 ceiling still shows the upside case. Limitations we accept and label: games played and per-game quality are treated as independent within a cohort, and undrafted rookies aren't projected at all. The walk-forward rookie backtest behind this lives in models/diagnostics/.

Every figure in this section is historical: it was measured on past seasons under the scoring and method named in its record, and it is not a claim about any current league’s projections. A result applies to today’s output only when your league’s scoring and the current model and band construction match the record exactly; otherwise the pages say “No measured evaluation for your league’s scoring and this model.”