A downside response gave no clear gain in Nasdaq variance forecasts from dated data for 2015–2025

We tested whether a stronger response to negative returns improved Nasdaq Composite variance forecasts after return size was known. Two standard models used ALFRED records available by the day before each forecast origin. Monthly fits used 1,260 past returns. The target was the single price return from the origin close to the next session close. The test had 2,766 origins in 2015–2025. The downside model reduced mean QLIKE from 1.5831 to 1.5597. Its paired gain was 0.02340, with a 95% 20-session block interval of -0.01239 to 0.05255. Five-session and 60-session intervals also included zero. The model gained in ten of eleven years, but lost strongly in 2020. A planned check without 2020 gave a gain of 0.04425, with an interval of 0.02670 to 0.06377. This does not justify removing crises from the main test. Removing 77 stale-source origins changed little. A post-plan check found that 78 of 132 downside fits gave positive returns almost no direct variance weight. A few large forecast losses in 2020 offset many smaller gains. All 528 optimizer starts succeeded. Independent scalar checks passed, and ten core files matched byte for byte in a full repeat. The package preserves code, model forecasts, fits, results, uncertainty draws, figures, source hashes, the local plan and repeat instructions. Raw source observations remain under provider rights and are retained separately. The test is a historical reconstruction from day-level release records. It is not a live trial. The evidence does not establish reliable return direction forecasts or a profitable trading rule.

variance forecasts · GARCH · GJR-GARCH · Nasdaq Composite · point-in-time data · chronological evaluation · negative result · crisis sensitivity