Skip to content

News-impact validation — what 8-K events do and do not predict

Three rounds of honest event-study methodology on SEC 8-K filings: the price dimension, the revenue dimension, and the customer→supplier chain. Ships with the revenue_model.news_impact module and the revenue_model.form8k_adapter data layer. Not investment advice.

1. The question

Direction-3 of the roadmap asked: can press-release / filing events improve revenue forecasts? The seductive version of this idea is a screen that watches 8-K filings and nudges next-quarter revenue forecasts when a "material agreement" or "M&A completion" lands. Before building that screen, we ran the honest version of the test — three times, each round correcting the last round's methodology.

2. Round 1: a single-company p-value is a mirage

The project's first pass (2026-08, SDGR case study) found exactly what a researcher hopes for: Partnership/M&A press releases → negative signal, p = 0.033. One number, one company, one story — biotech partnerships dilute, the market knows something, headline categories carry information.

Rebuilding the analysis exposed the mirage piece by piece:

Step Finding
Rebuild the revenue series Two adapter bugs (below) had corrupted the quarter data the original study ran on
Honest revenue dimension Partnership/M&A events were followed by +29.0% YoY growth vs a +20.8% baseline — wrong direction, p = 0.11 (Welch) / 0.22 (MWU)
Price dimension (monthly, market-adjusted) Event-month abnormal return +8.1%, p = 0.039/0.015 — significant! But positive, not negative as originally recorded
Pool across 6 issuers The price effect vanishes (Agreement +1.6%, p = 0.76) — it was one company's idiosyncratic run
Check the "next-month drift" Every category looked significant — because event-heavy months overlapped a bull market (a timing confound, not a signal)

The p = 0.033 was never a lie — it was a single-sample artifact. SDGR had 28 qualifying events; with ~20 categories × horizons to scan, a p < 0.05 shows up in noise almost surely. Pooling NVDA / AMD / SDGR / REGN / GILD / GM (409 events, 2019-2026), adjusting returns for SPY, and correcting for the family of tests: no 8-K category carries robust monthly-horizon information on these large caps. Which is, on reflection, exactly what an efficient market should do to the most public disclosure in existence.

3. Round 2: the revenue dimension — a selection effect, not a shock

Prices aside, the fundamental question is whether events move revenue. A second pass pooled 18 issuers across 8 sectors (semis / software / biotech / pharma / consumer / auto / industrial / energy; ~1,500 8-Ks, ~1,000 fiscal quarters), excluded Earnings filings (mechanically contemporaneous with the disclosure itself), and tested five outcomes including the growth acceleration (Δ YoY, pre- vs post-event) — the right outcome for a "shock", since growth levels are dominated by the company's regime.

The level results looked strong: post-event YoY ran 7-9pp below baseline, direction-consistent across all eight sectors, with fy_growth / Other-RegFD clearing Bonferroni (Welch p = 0.001). Three diagnostics then exposed it as selection, not causation:

Diagnostic Result
Pre-event 4Q YoY Already low (+12.7% vs +20% baseline, p = 0.013) — slow growers file more 8-Ks before anything happens
Paired delta (post − pre, within event) ≈ 0 (+1.8%, p = 0.15) — the event does not bend the trajectory
Filing rate by growth tercile Monotone: slow third 4.3 non-Earnings 8-Ks/yr, middle 3.4, fast 2.6

So the "signal" is a reverse slow variable: non-Earnings 8-K frequency is a characteristic label of low-growth companies (reorganizations, financing, management churn cluster in slow regimes). Useful as a screening feature, meaningless as a shock. The one weak true effect — a +4.6% paired acceleration after management changes (p = 0.033, "new-broom" CEO effect) — does not survive correction but is the only direction with a story worth following up (distressed-company subsamples).

4. Round 3: customer events → supplier revenue — the wave detector

The NVIDIA data-center story this project was born from is not "NVIDIA's own 8-Ks predicted NVIDIA." It is the customers' capex wave (Microsoft, Meta, Oracle announcing buildouts) arriving at the supplier 1-4 quarters later. Round 3 tested exactly that chain — customer non-Earnings 8-K density → supplier's lagged revenue — over eight pairs (MSFT/META/ORCL → NVDA/AMD/AVGO, NVDA → AVGO):

Customer high-activity quarter (≥2 filings) Supplier YoY at lag vs baseline
lag +1 .. +4 quarters +59% .. +62% (n = 13) +30% (n = 348), MWU p ≤ 0.008
placebo lag −1 / −2 +39% / +29% — not significant time direction holds
dose-response (density 0 / 1 / 2 / 3+) +29% / +35% / +40% / +80% monotone

The time structure is real and matches the mechanism (customers announce → orders → revenue recognition 1-4 quarters later). But honesty demands the fine print: those 13 high-activity quarters cluster entirely in two waves (the 2020-21 post-COVID cloud buildout and the 2023+ AI capex surge), and the acceleration placebo fails (lag −1 also significant) — the wave lifts both sides of the event, so this is a regime-level event study with n = 1 wave family, not a repeatable statistical regularity.

The correct product reading: news is a wave detector, not a trading signal. Monitoring ecosystem (customer / upstream) event density marks a driver regime shift 1-4 quarters before it shows up in the supplier's reported revenue — which is precisely the "news → driver revision → forecast update" loop this project's memo demo implemented by hand, and why event→driver mapping (not event→revenue regressions) is the right interface.

5. What shipped anyway

A null result with clean methodology is still a result. The validation exercise produced four library-grade assets:

  • form8k_adapter — 8-K events from SEC's submissions API: universal coverage (any ticker), official item codes (1.01 agreement, 2.01 M&A completion, 5.02 management, 2.02 earnings, ...), no key, no scraping. Strictly better than guessing IR-site domains.
  • news_impact — the honest event-study toolkit: Welch's t and Mann-Whitney U hand-rolled in pure stdlib (verified against scipy to 8 decimals), pooled event_study() with per-category rows, automatic Bonferroni family correction, and min_n guards that report small categories without a test instead of mining them.
  • Two sec_adapter bug fixes found by the rebuild: revenue-concept switching silently dropped years (SDGR lost 2019-2023), and a YTD/discrete period collision corrupted single-quarter differencing (negative quarters).
  • fetch_fiscal_quarters — a fiscal-year-general single-quarter series (NVDA's late-January FY handled), concept-merged, incomplete trailing years excluded per the M7 honesty rule.

6. Methodological checklist (encoded in news_impact)

  1. Pool. One issuer is one draw from a noisy distribution. If the effect needs SDGR specifically, it is not an effect.
  2. Adjust. Event months overlap market regimes; subtract the benchmark before calling anything abnormal.
  3. Count the tests. k categories × h horizons is a family; report the Bonferroni threshold alongside every nominal p-value (event_study does this by construction).
  4. Respect small n. A category with 5 events gets a number and a note, not a p-value.
  5. Verify the data pipeline first. Both "findings" above sat on top of broken quarter derivation. Plot the series before you test the series.
  6. Diagnose the pre-event window. A post-event level below baseline means nothing until you show the pre-event level was at baseline — otherwise you have selection (slow growers file more 8-Ks), not a shock. The paired post−pre delta is the honest shock estimate.
  7. Run placebo lags. A real lead-lag effect must vanish at negative lags. If lag −1 is as significant as lag +2, you are dating a wave, not transmitting a signal.
  8. Ask what n means. Thirteen significant quarters inside two industry waves is one draw of "wave happens", not thirteen independent confirmations.

7. Where news actually matters

Three rounds, three answers:

  • Own-company 8-Ks → own revenue next quarter: null (efficient disclosure, plus a selection artifact that fakes a level effect).
  • Own-company 8-K frequency: a reverse slow variable — a cheap screening feature for low-growth regimes, not a forecast input.
  • Ecosystem (customer/upstream) event density → supplier revenue 1-4 quarters later: real time structure — the wave detector. Statistically an n=1 wave family, operationally the trigger for driver revision.

The original vision stands, with the interface corrected: news feeds the forecast through event → driver revision → re-run, never through event → revenue regressions. The data layer for that loop (form8k_adapter + fetch_fiscal_quarters) ships in this library; the driver-mapping layer is the remaining build.

Not investment advice.