Past loss and gain shares gave no clear gain in weekly Nasdaq variation forecasts in a 2016–2025 archive test

We tested whether the sign of past daily price changes adds information to weekly Nasdaq Composite variation forecasts after total past variation is known. Each Monday forecast used an ALFRED archive dated the preceding Saturday and observations through Thursday. Weekly ridge fits used 520 completed earlier cases. The main test had 505 weeks with fresh inputs in 2016–2025. Adding the negative-return shares of squared changes over 5, 21 and 63 sessions increased mean QLIKE loss from 0.50331 to 0.51179. The loss reduction was −1.68%, with an eight-week paired block-bootstrap 95% interval of −7.90% to +5.06%. Thus this fixed sign split gave no clear gain; the interval still permits a useful improvement. Keeping 16 stale archive weeks gave the same broad result. Without 2020, the point estimate favored the split but the interval still included zero. Return forecasts also failed to beat the past-mean reference: their test R-squared values were −1.21% and −1.61%. This is a historical test with one index, daily-close risk measures and delayed inputs. It does not test intraday signed-jump models or trading profit. Three direct archive checks matched reconstructed values. Scalar checks passed, and a second run matched all 15 numerical output files byte for byte. We preserve the fixed plan, code, source requests and hashes, 573 forecast records, model states, data-age audits, 45,000 resampled estimates, figures, and negative findings. Original provider daily levels are excluded and remain in separate retained storage under provider terms.

Nasdaq Composite · point-in-time data · volatility forecasts · negative results · realized semivariance · chronological evaluation