quantmodel — Design held-out correlation-improvement experiment

Objective

Design held-out correlation-improvement experiment.

Notes: Selected Black Box Defense First from the three stored corr<0.95 replicas and preregistered one fixed 63-session downside-deviation experiment, 2021-04+ locked validation, six acceptance gates, and a one-attempt stop rule in held-out-bbdf-t247.md.

Approach

Reuses existing report research/findings/held-out-bbdf-t247.md — tables/metrics preserved verbatim via source_report pointer. Body below summarizes grounded evidence from that report and adjacent artifacts.

Files / code / data changed

No commit SHA confidently associated with t247 in sampled log [reconstructed — uncertain].
Source report: research/findings/held-out-bbdf-t247.md
Task-history entry: research/findings/task-history.json#t247

Results

Key table from source report (preserved):

rank	strategy	latest stored monthly correlation	decision
1	Optimum3	0.913	Do not select: t231 found that larger SCZ near-tie allowances and lookbacks did not robustly close the known residual.
2	Black Box Defense First	0.9266	Select: weakest candidate with a remaining continuous weighting ambiguity that can be tested without changing the inferred asset-selection rule.
3	Aspect RMM	0.9496	Do not select: already near the 0.95 weak threshold and its recent near-tie change improved the stored result.

Conclusions and metrics are in research/findings/held-out-bbdf-t247.md; see detail page for preserved tables/metrics.

Conclusions

Selected Black Box Defense First from the three stored corr<0.95 replicas and preregistered one fixed 63-session downside-deviation experiment, 2021-04+ locked validation, six acceptance gates, and a one-attempt stop rule in held-out-bbdf-t247.md.

Problems / follow-ups

See source report or PR discussion for follow-ups. No unsupported follow-ups fabricated.