# LIMITATIONS.md — living list (WP39)

## Data
- **Survivor-conditioned accounting panel.** Fundamentals and earnings-release rows exist only for the 5,451 tickers active
  on 2026-11-01. Firms that delisted during 2012–2026 (including those that deteriorated and failed) are absent from the
  firm-quarter sample, although their minute bars exist. Base rates of severe deterioration are therefore understated and
  H3/H5 effects for the most vulnerable firms are not identified. A delisted-fundamentals backfill is proposed (DATA_AUDIT §8).
- **PIT is a documented proxy.** `available_at = filing_date + 1 day` is a policy; all facts are `revision_id = 1` observed
  on 2026-11-01/02, i.e. the vendor's current (possibly restated) history. Original-as-reported values are not available.
  Inputs use prior quarters only, so restatements affect the *measurement of inputs and targets*, not the direction of time.
- **OHLCV bars, not ticks.** No quotes, no trades, no order flow. Bar-based statistics (down-bar volume share, bar
  autocorrelation) are descriptive proxies and are never interpreted as informed trading or spreads.
- Vendor merges ticker changes under the latest ticker (fully since 2022); CIK/SIC come from the SEC's *current*
  `company_tickers.json`, so historical ticker reuse before 2022 is only partially auditable (`formerNames` check).
- `release_time` is unknown for ~35 % of release events; we never use release-day bars, so this affects only the
  descriptive timing exhibit.
- The vendor `eps_estimate` has an unknown vintage and is missing for a quarter of events → exploratory only.
- 1-minute coverage is thinner before ~2005 and for illiquid names; the sample starts in 2012 and requires ≥ 100 regular
  bars per session, which excludes the most illiquid firm-days (a documented selection on liquidity).

## Design
- The event-time cohort conditions on the *existence* of a release in the next 65 trading days; a hypothetical observer
  knows the approximate schedule but not the exact date. The fixed-calendar design is reported as the robustness check.
- Prediction, not causation: a positive H1 shows that bar-level information is statistically informative; it does not
  establish who traded on private information.
- No B7/B8 (neural/text) models by owner instruction; the frequency comparison is therefore within linear and
  gradient-boosted families only.

## To be completed as results arrive
- (Phase 3) achieved sample sizes, base rates, coverage gaps by year/size.
- (Phase 5) which placebo/stress tests failed or were inconclusive.

## Completed from results (2026-11-02)
- Achieved sample: 3,796 firms, 102,522 firm-quarters (2013–2026); the sample starts in 2013 because eight prior quarters
  are needed for the seasonal expectation; 2020 lies in the validation period, so the COVID subgroup is not part of the test.
- The logistic-family confidence interval around the 1-minute increment (±0.0027) is comparable to, not narrower than,
  the pre-registered MDE (0.0025); the precise-null claim rests on the boosted family (±0.0006) and on the consistency of
  the sign across horizons, targets and subgroups.
- The announcement-window CAR coefficient on the incremental 1-minute component is positive and nominally significant
  (t = 2.9); it is one of several H5 tests, has the "wrong" sign for an information story, and is reported, not exploited.
- The placebo differential on shuffled labels is of the same order (−0.0004) as the primary LightGBM differential
  (+0.0016), so differences of this size should be read as zero.
- Delisted-firm stress tests and as-originally-reported specifications were not run (data not available from the vendor).