An asymmetric volatility model cut 5% Nasdaq daily tail forecast loss by 2.0% against a symmetric model in a 2015–2025 test with dated inputs

We tested whether separate responses to positive and negative returns improve daily downside forecasts for the Nasdaq Composite. Each forecast used ALFRED values valid before its origin. The target return began at the close of the origin session. This left that session for action. We compared symmetric GARCH with asymmetric GJR-GARCH. Both used ten-year fit windows and the same five-year residual tail calibration. In 2,766 target sessions from 2015 to 2025, GJR reduced lower 5% pinball loss from 15.038 to 14.736 score basis points. The gain was 2.01%, or 0.302 score basis points, with a paired 95% block interval of 0.049 to 0.610. The interval stayed above zero with 5- and 63-session blocks. The 1% tail gave no clear gain. Breaches still clustered. A calendar check found a false numeric close on Good Friday 2019. We removed non-session dates from all inputs and targets, repeated the analysis, and preserved the initial run and failed check. We also found 78 stale-input origins during a late-2023 source outage. The 5% gain remained after removing them and with earlier outcome labels. These checks followed score inspection. The complete corrected repeat matched 11 numerical output files. The package preserves source hashes, code, all 33,192 forecasts, fits, uncertainty draws, figures, negative findings and repeat instructions. The evidence supports a small risk-score gain for this index and period. It does not show reliable return-direction prediction or trading profit. No trading strategy or live order was tested.

Nasdaq Composite · point-in-time data · tail risk · GJR-GARCH · quantile forecast · data quality · out-of-sample test · filtered historical simulation