Causal Toolkit: Case Studies

Treated: NTAP. Donors: WDC, MU, STX, SNDK (same storage/memory peer group as sndk_supply_glut_fear, with SNDK now a donor instead of a peer input and NTAP promoted to treated). Hand-picked, not algorithmic. Treatment date is NetApp's 2026-09-11 earnings report, an AI-driven beat that pushed the stock as much as +11% intraday before settling to a smaller close-to-close gain. This case was originally planned right after sndk_supply_glut_fear but shelved: as of that writing (a few days post-event) yfinance had zero trading days of post-treatment data, so the post-period gap computed as NaN. Revisited here with a full week of post-earnings data available. **Fails the fit-quality gate** (relative RMSE 1.10x); same outcome as sndk_supply_glut_fear (1.02x): this donor pool cannot reconstruct NTAP's own pre-period path either, so the post-period gap isn't trustworthy in either direction. Shipped as an honest negative result rather than swapped for a better-fitting pool to force a cleaner-looking answer.

Treated: MU. Donors: AMD, NVDA, INTC (AI/GPU-chip suppliers facing the same demand-destruction read). Hand-picked, not algorithmic. Treatment date is 2026-07-01: Meta announced 'Meta Compute,' a plan to resell its own excess AI infrastructure capacity, which the market read as evidence hyperscaler GPU demand may be peaking rather than perpetually scarce. Chip suppliers sold off hard the same day (MU -10.6%, AMD -6.9%, NVDA -1.25%, the Philadelphia Semiconductor Index down over 10% across the next two sessions), even as META itself rose about 9% on its own announcement: a genuine divergence between the company making the move and the suppliers reading it as bad news. This is a different mechanism from sndk_supply_glut_fear (that one was about NAND/DRAM oversupply from Samsung/SK Hynix; this one is about AI-chip demand destruction from a hyperscaler reselling capacity instead of buying more). Tests whether MU's outsized drop was company-specific or the whole chip-supplier group re-priced together on the same news.

Treated: MSFT. Donors: GOOGL, META, AMZN (fellow Mag-7 hyperscalers, same donor logic as googl_capex_guidance). Hand-picked, not algorithmic. Treatment date is MSFT's Jul 30 after-close report: revenue +18% YoY to $90B, net income +31%, and guidance for ~45% constant-currency Azure growth (vs. a 40.92% estimate); the stock closed up about 16%, adding nearly $450B in market value, the largest single-day market-cap gain on record for any company. This is the mirror case of googl_capex_guidance: that one tested whether GOOGL's post-guidance selloff eight days earlier (Jul 22) was company-specific or a shared re-pricing; this tests the same question on the opposite signal (a beat instead of a miss), in the same hyperscaler group within the same two-week window.

Treated: SNDK. Donors: WDC, MU, STX, NTAP (memory/storage peers also riding the same 2026 AI-driven DRAM/NAND shortage supercycle narrative). Hand-picked, not algorithmic. Treatment date is 2026-07-02, the day SNDK fell about 14% intraday (11% close to close per same-day reporting) on Samsung/SK Hynix new-supply and AI-capex-peak fears spreading across the whole memory complex. This was the largest of several repeated 'supply glut fear' selloffs in the group through Jul-Sep 2026 (further echoes hit on 7/13, 7/29, 8/6, 8/18, and 9/10); 7/2 was chosen as the single cleanest, largest-magnitude catalyst date rather than trying to model the whole multi-month sequence at once. Tests whether SNDK's drop was company-specific or part of a shared sector re-pricing that should also show up in WDC/MU/STX/NTAP, against a backdrop where the underlying shortage thesis (DRAM/NAND contract prices up 58-75% QoQ in Q2 2026) stayed intact through the same window.

Reuses the exact same validated event/date as fed_pivot_tech, swapping the treated unit to XLU. Utilities are the textbook bond-proxy, rate-sensitive sector: this checks whether that theory actually shows up as a real synthetic-control gap, using the other 10 sectors as donors, or whether XLU just moved with the broader market like everything else that week. (Substituted for a literal single-day 'AI-datacenter-power' event test since that theme doesn't have one clean, verifiable catalyst date; this is the more rigorously anchored version of 'is XLU moving for a distinct reason.')

Treated: GOOGL. Donors: MSFT, META, AMZN (fellow Mag-7 hyperscalers, all running their own large AI-capex programs and facing similar market scrutiny over it). Hand-picked, not algorithmic. Treatment date is GOOGL's Jul 22 after-close report: Cloud grew 63% YoY, a strong number, but $190B capex guidance crushed the FCF margin (21%→10, from 21%→9.2% per prior notes) and the stock sold off anyway. Tests whether this was a GOOGL-specific reaction to its own guidance, or part of a shared 'market re-pricing hyperscaler AI-capex risk' move that should also show up in the donor pool.

Follow-up to tsla_earnings_miss_auto_peers, which failed the fit-quality gate: pure Detroit automakers (F, GM) don't move with TSLA day-to-day at all. Treated: TSLA. Donors: RIVN, LCID (EV-specific, still auto but more growth-multiple-priced than F/GM), PLTR, COIN (retail-favorite high-beta momentum names with no automotive exposure, included on the theory that TSLA trades more on 'high-beta retail-momentum growth' factor exposure than on 'automaker' exposure most days). Same Jul 22 earnings-miss treatment date. Result: improved (relative RMSE 1.29x, down from 1.78x with pure automakers) but still fails the fit-quality gate. TSLA doesn't reconstruct cleanly from either peer group tried so far. Read as: TSLA is genuinely idiosyncratic (Musk/narrative -driven) rather than a donor-pool-selection failure specifically (a case where the honest answer is 'this method can't cleanly isolate an effect here,' not a forced conclusion either way).

Treated: TSLA. Donors: F, GM, RIVN (traditional + EV auto peers). Hand-picked, not algorithmic. Treatment date is the Jul 22 earnings report (auto gross margins 16.3% vs 18.4% expected, EPS $0.33 vs $0.55 consensus, FCF went negative on AI/robotics capex). Tests whether the post-earnings selloff is a real, company-specific fundamentals break from the auto sector, or whether TSLA was already diverging from these peers beforehand for unrelated reasons. Result: fails the fit-quality gate. See tsla_earnings_miss_growth_peers for the follow-up with a different donor pool.

Treated: BE. Donors: FCEL, PLUG (fuel-cell/clean-hydrogen peers). A hand-picked peer group, not an algorithmic match. Treatment date is the Hunterbrook Capital report (2026-07-08). Tests whether BE's decline is company-specific (a real post-treatment gap vs. peers) or just riding the same clean-energy/rate-sensitive-growth wave FCEL and PLUG were already on. Pre-period deliberately short (~4 months, not 2 years): BE's own multi-year rally is so much larger than FCEL/PLUG's that a long window makes the donor pool unusable (see README). A short window centered on BE's recent behavior is the standard fix when the treated unit has a strong idiosyncratic long-run trend.

Validation case: reproduces the original one-off script's finding. Confirms the generalized toolkit recovers the same (honest, negative) result: XLK's gap vs. its synthetic counterfactual opened up months before the June 17 FOMC meeting, so the pivot itself can't cleanly take credit for it.