# Real-data validation record — 2026-09-29

Actual adjusted market history and historical macro vintages through **2026-08-31**. Example portfolio: SPY 40%, AAPL 20%, MSFT 20%, JPM 20%; monthly rebalancing; 10 bps per one-way turnover. Common history: 2000-06-30 through 2026-08-31 (315 months). This fixed portfolio was used to exercise the pipeline, not selected as a recommended strategy.

The advanced ensemble has **not established an improvement** over the historical benchmark. At 12 months it is worse on CRPS, with both exploratory paired block intervals above zero. At 1 and 6 months both intervals contain zero. Do not interpret these results as demonstrated real-market predictive superiority.

| Horizon | Origins | Historical CRPS | Ensemble CRPS | Ensemble raw 80% coverage | Difference, 95% interval (4-origin blocks) |
|---|---:|---:|---:|---:|---|
| 1 months | 84 | 0.03053 | 0.02990 | 72.6% | -0.00063 [-0.00196, 0.00061] |
| 6 months | 83 | 0.07950 | 0.08012 | 75.9% | 0.00062 [-0.00447, 0.00579] |
| 12 months | 81 | 0.11033 | 0.12234 | 79.0% | 0.01200 [0.00073, 0.02354] |

Lower CRPS is better. Differences are ensemble minus historical bootstrap. Quarterly evaluation dates have overlapping multi-month outcomes. Candidate fits and adaptive mixture weights use only earlier available data and matured targets. Future-poisoning tests verify prefix invariance, including vintage selection and horizon-specific blend weights. The 300-path backtest distributions have Monte Carlo error; repeating searches on this evaluation period would invalidate naive selection claims.

Automated checks cover return adjustment/cutoffs, unavailable/invalid inputs, macro future-vintage exclusion, GJR stationarity and positivity, exact factor-holding identity, chronological prefix invariance, full-distribution scoring, portfolio costs, synthetic fallback rejection, and server-side authentication. Synthetic fixtures remain appropriate for deterministic unit tests; they are never served as live market histories.

The live website recomputes this evaluation for each requested portfolio and horizon. Downloads include each origin's scores, the second block-length sensitivity interval, interval-adjustment coverage, model weights, macro audit and source/input hashes. See [methods](LIVE_METHODS.md) for assumptions and limitations. These results are retrospective fixed-holdings evaluations, not point-in-time stock-selection backtests or an untouched holdout.