Equal and online mixtures gave no clear 5% Nasdaq tail forecast gain over a selected model in 2018–2025

We tested combinations of six daily lower 5% return forecasts for the Nasdaq Composite. Dated ALFRED inputs and released outcome labels set the information limits. We selected one reference model and one online update rate from 754 eligible labels in 2015–2017. The test used 2011 sessions in 2018–2025. The equal mean had 2.98% more quantile loss than the selected GJR model. The mean loss reduction was -0.4788 score basis points, with a paired 95% interval from -1.1892 to +0.0439 from 21-session blocks. The selected online mean had 1.87% more loss and no clear gain. Breach rates were 5.02% for the equal mean and 5.62% for the selected model. A breach rate near 5% did not establish better forecasts. Longer blocks, later labels, and removal of 2020 gave no clear combination gain. A source date anomaly was contained by observation-date limits and was preserved in the audit. The study reuses a known historical sample. Availability claims depend on the source validity dates and stated date limits. It does not show trading profit. The package has code, source hashes, all forecasts and weights, all calibration scores, uncertainty draws, a fixed local plan, checks, and exact replay records. Source prices and direct outcomes remain in a separate retained item under provider rights.

return forecasts · quantile loss · forecast combination · online weights · dated data · Nasdaq Composite · negative result