Skip to content
Red Zone Rank
Connect your leagueFree email alerts

The 2026 season report monthly dry run — season not over

Every source's out-of-time accuracy by position, the market's weight through the season, what the closing-line replay predicted against what happened, the weather multipliers refit on the season, the chat's calibration and the engine's record against its own baseline. Every number reconciles to the accuracy page. Built Oct 1, 2026, 6:00 AM ET.

Headline

Forecast weeks0none · back-filled: 2, 3 (not counted)
Board vs baseline0–0board-weeks won–lost on MAE, forecast weeks, PPR · back-filled 4–7 (not counted)
Start/sit skill—no scored calls yet
ReconciledYes2 weeks against app.accuracy_page · season tiles match

Every source, by position

No forecast week yet — the table fills from the first board published before its week's kickoff, scored the Tuesday after its week is final.
PPR · back-filled weeks — shown, not counted · the board's own score, then every source re-scored on the same rows from the forecast ledger (MAE, ρ below). Green = the source beat the board, brown = it did not.
PositionWeeksRowsMAEρTop-12CoverageBaselineSeason EBUsageOpp-adjMarketvs baseline
QB back-filled2486.140.3610 / 2461%7.65ρ 0.127.50ρ 0.277.35ρ 0.217.26ρ 0.277.11ρ 0.251–0
RB back-filled2724.780.5511 / 2476%5.28ρ 0.474.79ρ 0.574.95ρ 0.525.28ρ 0.505.25ρ 0.47 · 56% own2–0
WR back-filled2727.960.4413 / 2460%6.51ρ 0.375.98ρ 0.466.23ρ 0.436.79ρ 0.386.71ρ 0.35 · 51% own0–2
TE back-filled2485.680.2213 / 2458%5.57ρ 0.104.96ρ 0.395.33ρ 0.205.45ρ 0.185.58ρ 0.06 · 52% own1–1
K back-filled2484.97-0.066 / 2444%4.30ρ -0.103.77ρ -0.04——3.89ρ -0.13 · 54% own0–2
DST back-filled2484.640.3517 / 2458%4.43ρ 0.204.63ρ 0.01———0–2
Week 2 · back-filledWeek 3 · back-filled
the ledger stores the box-score sources (baseline, season_eb, usage, opp_adj) AFTER the wind multipliers; there is no separate un-weathered arm — the weather section refits the multipliers instead

The market's weight, Tuesday by Tuesday

The fitted weight of the market source (implied team totals from closing lines) per position — the mean over the components it prices. app.projection_versions keeps only the current pin; `pinned` marks today's pinned fit, earlier Tuesdays' pins are the fits in date order
FitTrained onQBRBWRTEKDST
Sun 2026-09-27 fit-2026w03-mkt2025 W5–2026 W20.2150.1800.1780.1680.352—
Tue 2026-09-29 fit-2026w03-mkt-20260929 pinned2025 W5–2026 W30.2140.1810.1780.1640.352—

What the replay predicted, and what happened

Out of time — weeks after the fit's training span · component MAE on the published boards ÷ the replay's calibrated MAE, median over the position's components (1.00 = as predicted)
PositionComponentsRowsRealised ÷ predictedBeat baseline (realised)Beat baseline (replay)
RB8384×1.235 / 80 / 8
WR7336×1.683 / 72 / 7
TE7224×1.254 / 74 / 7
K9288×1.028 / 96 / 9
DST7224×0.922 / 75 / 7
In sample — weeks inside the fit's training span (a check of the code path, not a forecast) · component MAE on the published boards ÷ the replay's calibrated MAE, median over the position's components (1.00 = as predicted)
PositionComponentsRowsRealised ÷ predictedBeat baseline (realised)Beat baseline (replay)
QB8256×0.883 / 86 / 8
RB8384×1.227 / 80 / 8
WR7336×1.053 / 72 / 7
TE7224×1.214 / 74 / 7
K9288×1.025 / 96 / 9
DST7224×0.884 / 75 / 7

Weather multipliers, refit

Team passing yards (also receiving yards; pass TDs take half the effect) · multiplier relative to the 0–9 mph bucket (n in each bucket)
WindCurrent fileRefit 2019–20262026 observed wind
0–9 mph1.000n 14681.000n 15101.000n 42
10–14 mph0.961n 5240.963n 5441.000n 20
15–19 mph0.911n 2120.911n 2120.997n 0
20+ mph0.838n 600.838n 600.997n 0
Field goals from 50+ yards · multiplier relative to the 0–9 mph bucket (n in each bucket)
WindCurrent fileRefit 2019–20262026 observed wind
0–9 mph1.000n 5451.000n 5711.000n 26
10–14 mph0.980n 1830.977n 1910.922n 8
15–19 mph0.980n 730.977n 730.922n 0
20+ mph0.980n 130.977n 130.922n 0
no non-fixture NWS forecast is stored for a final game of the season yet — the forecast refit waits for real fetches Nothing was written: the shipped parameters change only after the final report.

What the chat said

Medians shown49 numbers · MAE 8.750% of the ranges shown held the actual (target 60%).
Start/sit callsNo calls scored yetSkill = 1 − Brier / 0.25 (a coin flip scores 0).
Review loop · 4 weeks13 answers · 3 up · 0 downGate: 0 sentences checked, 0 repaired, 0 dropped · 0 chip clicks

Into next season's calibration

Pool 2019–202645,241 player-weeks · 3,838 DST-weeksplus 4052 published 20–80 ranges with actuals this season
Spread multipliers nowQB 0.713 · RB 0.86 · WR 1 · TE 1.098 · K 1 · DST 1fitted on 2019–2025; one scalar per position, 20–80 coverage target 60%
What the refit usesitem 9 refits scoring/dist_params.json spread_mult (one scalar per position; 20–80 PIT coverage target 0.60, band 58–62 %) on the pooled replay population — scoring.coverage.population: REG player-weeks (QB/RB/WR/TE/K) and DST-weeks with their replayed baseline means and actual points; the counts here are mart.player_week / mart.dst_week rows, an upper bound, since population() keeps active-roster, non-bye weeks with a replayable mean — plus, for the season, the ledger's published board intervals (kind='set', component='points', p20/p80 with actuals; one row per board a player appears on) as the live check

Reconciliation

Definitions

weatherthe ledger stores the box-score sources (baseline, season_eb, usage, opp_adj) AFTER the wind multipliers; there is no separate un-weathered arm — the weather section refits the multipliers instead
ensemblethe board's proj_median re-scored on the arms' rows (what was shown)
by sourceper format × position × source over the scored weeks, forecast and back-filled apart: MAE and coverage weighted by scored rows, Spearman averaged over weeks (the accuracy page's aggregation)
reconciliationevery week's numbers recomputed through rankings.accuracy_page.week_payload == the stored app.accuracy_page payload; forecast aggregate == the page's season tiles
source coverageshare of scored rows where the source had its own value (otherwise it is scored at the baseline's)