Event study · panel vs single-LLM ablation

The same pipeline, run with and without the debate.

Panel (4 voices) · 1.0197 Single-LLM · 1.0000
cumulative return · indexed to 1.00
0.9941.0041.0141.0241.033IN-SAMPLEOUT-OF-SAMPLE

Strict OOS cutoff 2026-09-08 · 3d horizon · bar 0.6. Verdict PANEL_BEATS_OOS.

In-sample
ArmHit-rateavg P&LSharpe~Max DD
Panel0%0.00%0.000.00%
Single-LLM0%0.00%0.000.00%
Out-of-sample · never fit
ArmHit-rateavg P&LSharpe~Max DD
Panel50%1.00%0.570.74%
Single-LLM0%0.00%0.000.00%
Reported as-is · directional-signal study, not alpha

Out-of-sample the panel issued 2 advisory SHORT signals (hit-rate 50%, avg P&L 1.00%) while the single-LLM ablation cleared its own conviction bar 0 times (0.00%). The OOS sample is tiny and non-random, so this measures the VALUE OF THE PANEL'S CALLS AS SIGNALS - it is not executed returns and makes no alpha claim.

The panel's value is the auditable record - provenance, not a promise of alpha. Both arms are shown either way.