Does SEC-filing evidence on semiconductor-theme constituents carry information that precedes full price adjustment? This module aggregates YUCLAW's evidence stream to the theme level and tests the question as a classic event study — cumulative abnormal returns (CAR) around evidence events, with the sector factor removed by a peer-benchmark model. The result below is reported exactly as measured, including where it is adverse to the hypothesis. Scope note: this is research on the SEC-filing constituents of a published index fund's holdings list — an evidence study of companies, not analysis, promotion, or an offering of any fund or product.
Why not 100%? 17 disclosed constituents are outside the universe, including the foreign-domiciled issuers ASML, NXPI, STM, TSM — they file annual 20-F / occasional 6-K, not the 8-K + Form 4 event stream this pipeline consumes, so they produce almost no EDGAR event substrate. The same constraint puts foreign-holdings ETFs (e.g. KORU-class country funds) methodologically out of scope: their constituents do not file EDGAR at all, so there is no evidence stream to aggregate — a scope statement, not a caveat.
| Ticker | SMH weight | Signal label | Composite score | C6 | Evidence grade | Event history (type×count) | 30d |
|---|---|---|---|---|---|---|---|
| NVDA | 19.01% | RISK_FLAG (weakening) | -0.119 | -0.527 | Grade B | INSIDER_SELL×42, EXEC_CHANGE×4, OTHER_MATERIAL×2, M_AND_A_CLOSE×1 | 1 |
| AMD | 5.58% | RISK_FLAG (weakening) | -0.065 | -0.462 | Grade B | INSIDER_SELL×87, M_AND_A_ANNOUNCE×1, EXEC_CHANGE×1, OTHER_MATERIAL×1, M_AND_A_CLOSE×1 | 2 |
| MU | 5.34% | RISK_FLAG (weakening) | -0.137 | -0.445 | Grade B | INSIDER_SELL×160, OTHER_MATERIAL×4, EXEC_CHANGE×1, M_AND_A_CLOSE×1 | 68 |
| AMAT | 5.33% | RISK_FLAG (weakening) | -0.046 | -0.462 | Grade B | INSIDER_SELL×38, M_AND_A_CLOSE×1 | 3 |
| INTC | 4.88% | WATCH | +0.033 | +0.436 | Grade B | M_AND_A_CLOSE×2, EXEC_CHANGE×2, INSIDER_SELL×2, OTHER_MATERIAL×1, EARNINGS_BEAT×1 | 1 |
| LRCX | 4.68% | RISK_FLAG (weakening) | -0.178 | -0.462 | Grade B | INSIDER_SELL×51, M_AND_A_CLOSE×1 | 2 |
| MRVL | 4.27% | RISK_FLAG (weakening) | -0.112 | -0.382 | Grade B | INSIDER_SELL×20, OTHER_MATERIAL×3, EXEC_CHANGE×2, M_AND_A_CLOSE×1, DIVIDEND_CHANGE×1 | 5 |
| ARM | 1.15% | RISK_FLAG (weakening) | -0.169 | -0.462 | Grade B | INSIDER_SELL×37, M_AND_A_CLOSE×1, GUIDANCE_RAISE×1 | 0 |
Research classifications, not recommendations. C6 is the evidence-impact component of the composite score; the C6 deep-dive concluded its defensible role is a risk gate, not a near-term return signal (see methodology).
Direction-aligned pooled CAR (peer model, n=45 events) is -12.0% at τ=+20 (95% CI -19.7% to -4.3%) — significantly negative: prices moved against the evidence direction. The sample is dominated by INSIDER_SELL clusters during a sector melt-up (insiders sold the strongest names, which kept outperforming peers). This replicates the C6 deep-dive conclusion that insider-sale flow is a risk gate, not a near-term direction signal. Live-detected events since forward Day 0: n=45 directional — far too few for inference; this panel accrues with the live record.
| Event type | n events | Peer-model CAR τ=+20 | 95% CI | SPY-model CAR τ=+20 | n @ τ=+20 | Event window |
|---|---|---|---|---|---|---|
| INSIDER_SELL | 67 | +9.0% | (+1.9%, +16.0%) | +10.4% | 58 | 2026-03-26 → 2026-07-17 |
| EXEC_CHANGE | 9 | -2.5% | (-32.8%, +27.9%) | -3.3% | 6 | 2026-04-03 → 2026-07-02 |
| M_AND_A_CLOSE | 9 | +8.6% | (-11.0%, +28.2%) | +18.1% | 9 | 2026-04-08 → 2026-05-14 |
| OTHER_MATERIAL | 6 | -4.8% | (-26.1%, +16.4%) | -2.3% | 5 | 2026-04-01 → 2026-07-09 |
| M_AND_A_ANNOUNCE | 1 | -21.4% | n too small for CI | -0.8% | 1 | 2026-05-15 → 2026-05-15 |
| EARNINGS_BEAT | 1 | -3.5% | n too small for CI | +2.5% | 1 | 2026-07-23 → 2026-07-23 |
CIs assume independent events; distinct ticker-days can still share calendar days across tickers, so the intervals are optimistic — stated, not hidden. Types with n < 10 are shown for completeness and are statistically uninformative. Form 4 (insider) events are a parsed batch covering 2026-02-18 to 2026-05-15; live Form-4 ingestion is a tracked pipeline item, so the live-era sample is 8-K derived and small.
Events: accepted L1 evidence events on covered constituents, deduplicated to one observation per (ticker, event type, direction, day). Day 0 = first trading day on/after the event timestamp's date. Abnormal returns: two models, never averaged — peer model (market = equal-weight of the other covered constituents that day; removes the sector factor; the fair test) and SPY market model (classic; in a strong sector regime its "abnormal" return is dominated by the sector factor, which is what it shows). Estimation: up to 60 trading days ending 6 days pre-event, minimum 30 observations (22 events skipped for insufficient estimation history — early-window events; count disclosed). CAR window [−5, +20]; mean CAR with a normal-approximation 95% CI across events; n disclosed at every horizon. Direction alignment: AR × direction for direction ≠ 0 events, so "evidence direction predicts abnormal return" is a positive number if true.
Backfill-era events (before forward Day 0 = 2026-05-18) were extracted retrospectively and carry the same parametric look-ahead disclosure as the in-sample panel of the Validation Lab; live-era events are detected in real time. The two eras are never blended. Derived statistics only — no raw prices. Constituent weights are issuer fund disclosures as of 2026-07-03.
Everything this page derives, as files — YUCLAW-derived data only (derived statistics, counts, classifications; never raw vendor price/options data). Regenerated in the daily chain.
YUCLAW Open Index Evidence, v5.0.0, data through 2026-07-23, built 2026-07-23, commit 4e9c15cde2f1; evidence-tier names excluded from scoring universe.
pip install yuclaw then yuclaw replay-lab.
No install: tools/replay_lab.py
(stdlib only) against the published bundle.
Exit 0 = every statistic and ledger root reproduced. How to report a replication →
One real Suncor 6-K, end to end:
filing → exhibit → extracted prose → event type → grade → C6 posture.
Open the trace → ·
example evidence memo (Suncor) →
Every evidence packet ships a ready citation snippet
(version, data-through, build date, source commit).
Get the citation →
Rendered from one shared source (v3/web/useful_blocks.py) on every page that shows it,
so the copies cannot drift. Statuses are measured, not aspirational.