Empirical claim #4 — Pooled regime-centre own-goal test

Generated 2026-05-23 21:58 UTC by `scripts/py/intelligence/event_study_pooled_own_goal.py`. Re-run anytime; pure-Python, no network.

What this is

Claims #1, #2, and #3 each surfaced an unplanned secondary finding: the country at the centre of each policy regime trailed ACWI over the headline horizon. Individually, each sub-finding is under-powered (claim-#1 MCHI T+60d sign p=0.099, n=30; claim-#2 MCHI T+60d sign p=0.092, n=13; claim-#3 EIDO T+120d sign p=0.002, n=10). Only one sub-finding clears p < 0.05 on its own.

Claim #4 pools the three cohorts to test the narrower meta-question: does the regime-centre ETF trail ACWI when independent regimes are combined? It is one statistic computed across three sub-cohorts that remain individually-defined as in their source scripts; it is NOT a new cohort definition.

Pre-registered cohorts (verbatim from source scripts)

Date filter: announced_date in [2018-01-01, 2026-12-31]. Total raw matched: 54. In primary window (T+60 available): 53. In secondary window (T+120 available): 50.

The claim

Across the union of the three sub-cohorts, the regime-centre ETF (MCHI for the US-BIS-target and CN-MOFCOM-issuer sub-cohorts; EIDO for the ID-hilirisasi sub-cohort) abnormal return vs. ACWI at T+60 trading days is negative with sign-test p < 0.05. The own-goal pattern that each individual claim noted as suggestive holds when independent regimes are combined.

This is a historical association across the pooled sample, not a forecast and not a causal claim about any single regime.

Results — pooled abnormal returns vs. ACWI

The T+120d row is labelled secondary because only claim #3 was pre-registered at that horizon. Extending it to claim #1 and claim #2 events is an ex-post horizon extension; the pooled headline rests on T+60d.

Results — by sub-cohort (sanity check the pool isn't one-cohort-driven)

Falsification check

• (a) pooled median at T+60d: -3.67% (negative? YES).
• (b) pooled sign-test p at T+60d: 0.027 (< 0.10? YES; < 0.05? YES).
• (c) leave-one-out check (does any single sub-cohort drive the pool?):

- without claim1_us_bis: T+60d median over remaining 23 = -4.85% - without claim2_cn_mofcom: T+60d median over remaining 40 = -2.43% - without claim3_id_hilirisasi: T+60d median over remaining 43 = -4.33%

Falsification rules NOT triggered. Pooled median is negative AND sign-test p ≤ 0.10 AND no single sub-cohort removal reverses the sign — the pooled effect is not one-cohort-driven.

Events used (in primary T+60 window)

Honest caveats

1. Pooling violates trade-independence on the MCHI sub-cohorts. Claim #1 (US BIS -> CN) and claim #2 (CN MOFCOM) BOTH use MCHI as the centre ETF. The two cohorts can temporally overlap — a documented pattern in case #1 is US action followed by Chinese response within 60 days. Pooling reuses overlapping post-event windows on the same ticker; the pooled sign-test p-value is an anti-conservative upper bound. The leave-one-out check is the first defence; a stricter remedy would be a non-overlap subset (drop any event whose T+60d window overlaps an earlier event's on the same centre ticker) — flagged as a candidate revision.

2. "Regime-centre" elides target-trails vs. issuer-own-goal. Claim #1 measures the TARGETED country (China being restricted by US BIS); claims #2 and #3 measure the ISSUING country (China and Indonesia restricting the world). Both produce centre-ETF trailing — but the economic mechanism differs (export-revenue compression for the issuer; demand-cutoff for the target). The pooled test is outcome-equivalence, not mechanism-equivalence. A reader who wants the mechanism cleanly should read each source claim's audit separately.

3. The MCHI sub-cohorts share one country. Both the claim-#1 and claim-#2 sub-cohorts centre on the Chinese equity market via MCHI. Independence across the three sub-cohorts is incomplete: this is two regimes hitting one market + one regime hitting a second market, not three regimes hitting three independent markets. The strongest interpretation is "MCHI trails after either US restriction of China or Chinese restriction of others, and EIDO also trails after Indonesian restriction." That is what the pooled effect documents, no stronger.

4. No factor-return adjustment. ACWI is the only benchmark. A factor-purged alpha (market + EM + commodity) would attenuate; this is the same caveat that applies to each source claim.

5. Cohort criteria pre-registered; sub-cohort selection is not. The choice to pool claims #1 + #2 + #3 (rather than claims #1 + #2 only, or claims #2 + #3 only, or some other combination) was made after the source claims had shipped. The criteria within each sub-cohort are unchanged from the source scripts — that part is pre-registered — but the act of forming the pool is post-hoc.

6. The T+120d secondary row is the most ex-post extension. Claims #1 and #2 were pre-registered for T+5/T+20/T+60 only; extending those events to T+120 here is the most aggressive horizon stretch. The pooled headline at T+60d does not depend on it.

7. Cohort-period confounding still applies. 2020 COVID, 2022 commodity super-cycle, 2024 nickel oversupply collapse, 2024-2025 REE-price spike, 2026 §232 — these joint macro forces are not absorbed by ACWI for emerging-markets concentrated indices (MCHI, EIDO). The pooled negative could partly reflect EM-equity trailing DM-equity in the cohort window.

8. What pooling buys vs. what it doesn't. Pooling buys sample size and lets us report ONE number instead of three each with their own small-n caveat. It does NOT buy stronger evidence for the mechanism in any single regime; the per-cohort table is still the right read for that. The pooled claim is "the pattern is broad-based across policy architectures," not "the pattern is stronger in any regime."