A negative-return share gave an uncertain 0.75% cut in weekly Nasdaq variance forecast loss in 2016–2025

We tested whether the share of recent squared daily returns from negative returns improves next-week Nasdaq Composite variance forecasts. Each Tuesday forecast used a dated ALFRED record, past inputs and completed training targets. The main test has 505 usable weeks in 2016–2025; 16 stale archive cases are excluded. A fixed ridge model used past variance over 5, 21 and 63 sessions. Adding the negative-return share over 21 sessions reduced mean QLIKE loss from 0.51693 to 0.51304. The paired change was -0.00389, with a 95% eight-week block-bootstrap interval of -0.01375 to +0.00531. The relative loss reduction was 0.75%, with interval -1.01% to +2.63%. Other planned block and time checks also included zero. Variance mean squared error fell by 1.50%, a favorable secondary result. Return forecasts failed to beat the expanding mean; the asymmetric return model had 2.21% higher mean squared error. The evidence gives no clear gain on the planned primary score. It does not rule out a modest benefit or test high-frequency signed variation. No trading strategy was tested. Independent source and numerical checks passed. A second run matched 11 core files byte for byte. Removing future source releases left 207 early forecasts unchanged. The package preserves code, source requests and hashes, all forecasts and fits, uncertainty draws, negative results, figures and repeat instructions. Original prices retain provider rights and remain external.

variance forecast · Nasdaq Composite · negative returns · point-in-time data · chronological test · ridge regression · negative result