Extra weight on past falls cut 5% Nikkei daily tail forecast loss by 1.3% in a delayed 2015–2025 archive test

We tested whether a fixed extra response to past price falls improves daily Nikkei 225 loss forecasts. Each input came from an ALFRED record dated before the start of its target return. The test had 2,687 returns in 2015–2025. Giving squared falls weight 1.5 and squared rises weight 0.5 reduced lower 5% quantile loss by 1.30% against equal weights. The paired 95% 21-session block interval was 0.14% to 2.53%. Five- and 63-session blocks and removal of stale inputs retained a positive interval. Removing 2020 or using more delayed inputs gave intervals that included zero. The years 2020 and 2025 supplied 81.5% of the total score gain. The gain against a raw historical quantile was also uncertain. First-receipt and later return labels were equal. This is a market transport test of a published Nasdaq rule, with common world-market years. It is not independent future evidence. A later exchange-calendar audit found one priced source row on a closed day. The final analysis removes it from every vintage and target schedule. This correction was made after the first scores. The complete first analysis is preserved. Separate calculation checks passed. A full second analysis matched seven core public files and two private ledgers byte for byte. The package preserves code, source records and hashes, all 75,236 forecasts, model states, aggregate scores, 105,000 resampling rows, figures, negative findings and repeat instructions. Restricted source-derived outcomes are retained separately. The result supports a small risk forecast gain under stated delays. It does not establish reliable return-direction prediction or trading profit.

Nikkei 225 · tail forecasts · dated data · asymmetric volatility · filtered historical simulation · chronological tests · negative findings