Downside shares reduced weekly Nasdaq Composite variance forecast loss in 505 dated tests from 2016 to 2025

We tested whether the share of past squared returns from negative days helped forecast next-week Nasdaq Composite price variation. We used dated ALFRED records and froze each input vector and training label when available. A fixed ridge model added negative-day shares over 5 and 21 sessions to total variation over 5, 21 and 63 sessions. In 505 available test weeks from 2016 to 2025, the added inputs reduced mean QLIKE loss by 0.02344. The paired 95% eight-week block interval for the loss change was -0.04261 to -0.00625. Variance squared error fell by 3.91%. All ten yearly mean loss changes favored the added inputs. Sixteen weeks were omitted because current source records were absent. Return forecasts did not improve: the downside model had 2.19% more squared error than the past mean. An added check on 504 common weeks did not show lower loss than a published VIX model. The result applies to one price index and a fixed model. It does not show a trading profit. The package preserves code, source request fields and hashes, model fits, forecasts, uncertainty draws, figures, checks, and repeat instructions. Raw provider prices remain outside the open package under their source terms.

variance forecast · negative returns · Nasdaq Composite · point-in-time data · chronological test · ALFRED · negative finding