Empirical claim #4 (sensitivity) — Non-overlapping pooled own-goal subset

Generated 2026-05-26 16:50 UTC by `scripts/py/intelligence/event_study_pooled_own_goal_nonoverlap.py`. Re-run anytime; pure-Python, no network.

What this sensitivity defends

event-study-pooled-own-goal.md caveat #1: pooling claims #1 and #2 reuses post-event windows on MCHI (both sub-cohorts use the same centre ETF), so the pooled sign-test p-value is an anti-conservative upper bound. The leave-one-out check was the first defence; this script is the direct defence — walk events in date order per centre ticker, drop any event whose T0 falls inside an earlier accepted event's T+60d window on the SAME centre ticker, re-run the pooled stats on the reduced cohort.

Filter is anchored on the headline T+60d horizon, not T+120d — secondary horizon is reported, not used for filtering.

The headline pooled audit stays published verbatim. This file is its sensitivity, not a replacement.

Cohorts before vs. after the non-overlap filter

In primary window (T+60 available): 20. In secondary window (T+120 available): 19.

Results — pooled abnormal returns (non-overlap subset)

Comparison vs. the headline pooled audit

The "headline pooled" column is reported here as the published values from event-study-pooled-own-goal.md (committed 2026-05-23 21:47 CEST) so this audit is readable independently. Re-run the headline audit to confirm those values; they do not change unless the underlying register or parquet data changes.

Results — by sub-cohort (non-overlap subset only)

Sensitivity verdict

• Non-overlap pooled T+60d: median -3.03%, sign-test p 0.263, Wilcoxon p 0.279, n=20.
• Sign preserved vs. headline (-3.67%): YES.
• Sign-test p preserved (≤ 0.10 floor): NO (headline was 0.027).

Caveat #1 partially defused. The sign of the pooled negative T+60d effect survives non-overlap filtering, but the sign-test p-value crosses the 0.10 floor — consistent with the trade-independence concern being real (some of the headline significance came from window overlap on MCHI). The directional claim survives; the strength of the claim is reduced. Equity-research framing should drop p < 0.05 language for this pooled effect.

Events dropped by the non-overlap filter

Events used in the non-overlap primary T+60 window

Honest caveats (sensitivity-specific)

1. Cohort shrinks meaningfully. The non-overlap filter drops every claim-#2 MCHI event whose T0 falls inside an earlier claim-#1 MCHI event's T+60 window (or vice versa). This is exactly the trade-independence enforcement the headline caveat #1 flagged, and it costs a substantial share of the pooled sample. The smaller n means the Wilcoxon test moves more than the sign test (Wilcoxon is more sensitive to sample size at small n).

2. The greedy first-wins acceptance rule is one defendable choice among several. "Walk in date order, accept the first event per ticker, then block until its T+60 ends" is the simplest non-overlap rule. Alternative rules — "accept the more severe event," "accept the most recent event in each block," or "thin to mutually-exclusive windows by global optimisation" — give different subsets. The first-wins rule was chosen because it is unambiguous and matches the "post-event windows on the SAME ticker overlap" framing in the headline caveat #1.

3. EIDO is filtered too, but never reduces. The claim-#3 (Indonesia) sub-cohort is the only one centred on EIDO and its events are sparse enough that the non-overlap filter leaves it unchanged at primary T+60d. The sensitivity is effectively a check on the MCHI overlap; the EIDO sub-cohort sails through the filter and behaves the same as in the headline audit. This is honest about where the defence does and doesn't bind.

4. The non-overlap subset is post-hoc. This sensitivity is a defence of a published claim, not an independent claim. The discipline is that we publish whichever result the non-overlap test produces — including a reversal — and that both versions stay published per the project file's rollback rule.

5. Cohort-period confounding still applies. All caveats from the headline audit that are not about window overlap still apply here: no factor adjustment, EM-DM divergence in 2020-2026, the "regime-centre" outcome-equivalence vs. mechanism-equivalence point.

6. The T+120d secondary row is reported even though the filter is anchored on T+60d. Filtering on the headline horizon is the conservative choice; using T+120 as the filter horizon would shrink the kept set further. The T+120d row in the table is informational — read it with the understanding that overlaps in the T+60-to-T+120 tail are not filtered.