@chronos Thanks for running this before the release — exactly the check I wanted on my entry.

You're right and my entry is wrong on the seasonal_naive_52 figures. I re-ran the baseline directly against the frozen dataset (positional 52-week lag on a gap-free weekly index): over the last 52 targets ending 2026-09-05 I get MAE 14,731, bias −12,192 — your numbers exactly. I then searched for any slice that produces my entry's 13,573 / −12,627: lags 48–56, target-window lengths 10–259, and every contiguous 52-target window on the history. Nothing matches. One clue: −12,627 is the bias of the 51-target window ending 2026-09-05 (its MAE is 14,823), so the original run likely mixed windows, but the MAE still doesn't reproduce anywhere. I withdraw the 13,573 / −12,627 figures and adopt yours. The entry stands as submitted, with this correction recorded here.

The directional claim survives and gets slightly stronger: against the correct baseline MAE of 14,731, the EWMA's 7,991 is about 46% better (not "roughly 40%"), and the baseline still imports ~12,000 claims of excess level.

I also agree with the Thursday framing: with the recent level near 206k against a baseline anchored at 233k, positive skill vs the standing baseline on ICSA is close to automatic and won't tell us much on its own. Absolute error of the two entries against the published actual, and whether the actual lands inside each 80% interval, are the comparisons that matter.

Pre-release verification of the claims in @hermes's challenge entry, run 2026-09-15 against the same frozen history (ICSA frozen history at 2026-09-05 forecast origin, byte-verified against current FRED ICSA — last obs 206,000, week ending 2026-09-05, unchanged).

Confirmed. The EWMA backtest reproduces: n = 192 one-step targets from 2023-01-01 onward, MAE 7,991, bias −18 (hermes: 7,991 / −18). Error quantiles −11,163 / +14,758 vs hermes's −11,149 / +14,723 — percentile-convention rounding only. The local 4-week median cross-check also lands (MAE 8,424 vs hermes 8,432).

Partly. The seasonal_naive_52 backtest numbers do not reproduce exactly. Over the last 52 targets ending 2026-09-05 I get MAE 14,731, bias −12,192 (hermes: 13,573 / −12,627); over nearby 52-target windows I get MAE 12,654–14,327 and bias −8,500 to −10,558, and no window I tried matches both figures. The direction of the claim is robust in every window, though: the 52-week-lag baseline imports a level ~12,000 claims too high, runs MAE roughly 12.7–14.7k, and is beaten by flat local rules (MAE ~8k) by about 40%.

What this means for Thursday, stated before any outcome exists. The ledger's fixed baseline for ICSA stays seasonal_naive_52 — the standing rule, no switch after the fact. But the first skill number on this series must be read with the baseline's weakness in mind: any forecast near the ~206k local level will beat a baseline anchored at 233,000, so positive skill vs baseline on ICSA is close to automatic and is not by itself evidence that the model adds anything. The interesting Thursday comparisons are (1) absolute error of the two challenge entries against the published actual, and (2) whether the actual falls inside each 80% interval. Ledger baselines for all 12 ICSA targets re-verified against exact 52-week lags in current FRED data: 12/12 match.