Updating variance within paths gave no clear gain for five-session Nasdaq 10th percentile forecasts in 2015–2025

We tested whether variance updates inside a simulated path improve five-session Nasdaq Composite lower 10% return forecasts. The test used 537 nonoverlapping periods in 2015–2025 with dated ALFRED inputs. Sixteen planned periods failed a fixed input freshness rule. The two main systems used the same starting variance and centered, scaled residual pool. One held variance fixed. The other updated variance with fixed EWMA decay 0.94 through the unobserved bridge and target path. Both systems had the same expected target variance. Recursive paths reduced mean quantile loss by 0.0357 basis points, or 0.064%. The 95% paired eight-slot block interval was −0.2924 to +0.3349 basis points. The gain was not clear. A second simulation seed and other block lengths gave the same conclusion. The point gain became negative without 2020. Lower 5% and central 80% interval scores also gave no clear gain over the flat rule. This is a fixed historical risk forecast test, not a trading-profit test. We preserve code, both forecast runs, model states, source hashes, exclusions, resampling draws, figures and checks. A fresh replay matched eight core files byte for byte. Source prices and realized returns are excluded from the public package under provider terms.

Nasdaq Composite · five-session returns · filtered historical simulation · point-in-time data · EWMA · quantile forecasts · null result · forecast evaluation