# EMPIRICAL_DESIGN.md

Working design document — expands as stages complete. Authoritative parameter values
live in `config/`. Protocol reference: project CLAUDE.md §§7–16.

## Objects of interest
1. **Shocks:** idiosyncratic residual-return extremes (lagged conditional quantiles /
   robust SD multiples), separated by sign; volume, range, and illiquidity-proxy shocks.
2. **Propagation outcomes:** breadth, speed (censoring-aware), intensity, amplification
   (ISAR), persistence, asymmetry, at horizons h ∈ {1,2,5,10,15,30,60,120} min within session.
3. **Networks:** estimated strictly pre-shock (rolling/frozen windows): correlation,
   partial correlation, lead–lag, volatility spillover, sparse VAR/glasso; centrality roles.
4. **Evolution:** flexible year effects first; era bins only as secondary presentation.

## Identification strategy (threat: common information)
- Exclude scheduled macro windows from the idiosyncratic baseline.
- Isolation filters: low contemporaneous market/sector return, no simultaneous
  extreme in >2% of universe, leave-one-out factors.
- Placebo timing matched on stock, minute-of-day, volatility state.
- Permuted networks preserving degree/sector/size/liquidity.
- Negative controls: future network weights (falsification only), pre-shock outcomes.

## Inference
Two-way clustering (source × date) as baseline; day-block bootstrap for
distribution-level statements; randomization inference for event studies;
multiple-testing corrections within outcome families.

## Status
- [x] Stage A pilot audit (2026-07-13)
- [x] Stage B full download (2026-07-13)
- [x] Stage C full raw audit
- [x] Stage D clean panels
- [x] Stage E descriptive evidence
- [x] Stage F shocks & propagation (baseline, F2 network/cumulative/co-jumps, robustness)
- [x] Stage G paper (v1.0 2026-07-13; v1.1 revised 2026-10-17)
- [x] Revision diagnostics from stored events (scripts/17, 2026-10-17)

Implemented vs planned (as of 2026-10-17): implemented — residual shocks |z|≥6 with isolation,
placebo timing, ΔΔ, day/month/ticker bootstrap, lagged top-20 correlation neighbours, cumulative
breadth, co-jump minutes, pre-event levels, co-arrival classes, stratified estimator. Not implemented
— two-way clustering regressions, local projections with fixed effects, permuted networks, future
network weights, multiple-testing corrections, sector proxies (no security master), macro-window
propagation as a separate object, ML extension. See paper §8 and REVIEW-REPORT §8.