A large loss term gave no clear weekly Nasdaq return forecast gain in a delayed 2016–2025 archive test

We tested whether a large five-session price loss improves the next delayed weekly Nasdaq Composite return forecast. Each Saturday origin used its dated ALFRED record and observations through Thursday. A paper position began at the next Monday close. The 2016–2025 test had 504 fresh forecasts, including 29 large loss cases. Adding a fixed loss term to linear shock and risk inputs raised squared error by 0.79% against the past mean. Its MSE gain against the linear model was -0.06790 percentage-return squared units, with a paired 95% eight-week block interval from -0.19680 to +0.03479. The gain was also uncertain with other block lengths, later labels, stale inputs, and without 2020. In 521 calendar weeks, a stock and cash paper rule at 10 basis points per dollar traded had 1.81 annual percentage points less mean-variance utility than the linear rule, with interval -4.16 to +0.46. Independent model, label and account checks passed. Future-release deletion left 313 earlier forecasts unchanged. The package keeps code, forecasts, states, source hashes, cost cases, negative results, figures and repeat instructions. The test uses one price index, reused market years, a delayed signal and cost scenarios. It does not establish reliable trading skill.

return prediction · Nasdaq Composite · reversal · point-in-time data · chronological evaluation · trading costs · negative result