VIX from the prior session reduced 22-session S&P 500 variance forecast loss in a 2016–2025 test of 2026 data snapshots

We tested whether VIX from the prior session improved forecasts of S&P 500 risk over the next 22 trading sessions. We used official Cboe historical snapshots retrieved in October 2026. The test had 2,492 origins in 2016–2025. Inputs ended at the prior close. Monthly fits used only targets that had finished. The reference model combined two measures of past-return variance. Adding VIX reduced mean QLIKE loss from 0.702 to 0.493. The paired difference was -0.209, with a 95% 66-session block-bootstrap interval of [-0.390, -0.069]. A check with 114 target windows that did not overlap gave -0.151, with interval [-0.307, -0.029]. The better risk forecast did not give a clear trading gain. At an assumed five-basis-point cost, the allocation rule with VIX earned 6.18% a year, compared with 7.32% for the past-return rule and 8.15% for fixed 60% equity. These are simulated price returns with zero cash interest. A fixed ridge return forecast gave an uncertain 1.39% error reduction and no added direction accuracy. Raw VIX also gave lower full-test loss than the calibrated model with VIX. All forecasts, code, fit records, uncertainty draws, source hashes, and negative findings are preserved. Independent checks and a full repeat passed. A forward paper forecast for 26 October 2026 is saved and remains unscored. The files are current historical snapshots, not verified original historical data versions. The evidence assumes availability of the stated closes by the next session. It supports a scoped risk benchmark, not reliable stock-return direction prediction or a new profitable strategy.

variance forecasting · VIX · S&P 500 · chronological evaluation · point-in-time assumptions · trading costs · negative results · forecast provenance