Methodology
Six benchmark models, two event-date anchors, three primary windows, an exhaustive placebo battery, an AI-loading cross-section, a GARCH refit, and a pre-specified power analysis. The design is built to find a Dell-specific effect if one existed; the admissible claim is strictly “no detectable effect, MDE .”
Open any method in Plain English or Academic detail
Each button opens a three-tab modal: a Plain-English explanation, a worked example bound to this study’s numbers, and the academic formula with citations. No number is hardcoded — the worked examples resolve against the canonical JSON.
Design parameters
Estimation window
trading days; -day gap
Estimation σ
daily abnormal-return volatility
Primary anchor
T0 =
co-equal sensitivity anchor T0 =
Primary windows
wider windows confound-flagged
Multiple testing & power
The mechanism
The AI-loading cross-section regresses each peer’s May-5 CAR on its AI-beta. If the apparent May-5 significance were a Dell-specific signal, Dell would sit far off the line. Instead it sits on it. The full panel (N = ) gives slope , R² , p = ; screening Arista (an earnings confound, screen #2) gives N = , slope , R² , p = — the gradient strengthens. Preliminary at N = 7/8, pending the ≥19-donor export. The interactive scatter and per-firm methodology are on the Stress tests page.