How corporate America repriced risk, one sentence at a time — BaselineWatch Research
*BaselineWatch Research — July 2026. All figures below are reproducible from our point-in-time filing archive; methodology notes at the end, including one claim we killed during review.*
When the first strikes hit on February 28, every risk desk on the street repriced oil, shipping, and defense in minutes. Corporate disclosure repriced too — but in quarters, not minutes, and in language, not basis points. We read roughly 1,400 10-Q filings across seasonally matched windows to measure exactly how.
Compare the 10-Qs filed in April–May 2026 — the first written after the war began — with the same April–May filing season in 2025. Using a strict war-term list (Iran, Hormuz, Persian Gulf, blockade, hostilities, armed conflict — deliberately excluding ambiguous stems like "escalation," which also appear in routine cost language):
One in four public companies now names the conflict, or its geography, in a quarterly report. When that share doubles in a single season, the language is spreading faster than the underlying exposure — which is exactly what makes shock-quarter language hard to trade.
The 10-Qs being filed right now — the first covering a full quarter of war, ceasefire, and blockade — show the second phase. Direct war mentions are receding from the shock peak (down about 16%). What's rising is what we call transmission language: shipping disruption, freight costs, rerouting, supply-chain interruption, insurance premiums, energy prices.
Against the matched early-season 2025 window, mean transmission-language density is up 92%, and the 90th percentile has more than doubled. And here is the detail we didn't expect: transmission language was essentially flat during the shock quarter itself (+5% versus its matched control). Companies didn't quantify the war while it was breaking — they narrated it. The quantification is happening now, one quarter later, as the costs land in operations.
Shock becomes narrative; narrative becomes line items — with a one-quarter lag. That is what risk repricing looks like in disclosure.
BaselineWatch measures each filing against the same company's own prior filings — not against the market. During the war months, the marketwide tide was unmistakable in our scoring: the share of filings landing in our elevated-risk bands went from 10.7% (April 2025) to 59.4% (April 2026), and from 3.9% to 54.0% in May, under identical scoring logic across both years. Average uncertainty-language drift in April ran at more than seven times its 2025 level.
This is precisely the environment where cross-sectional language measures struggle: when everyone's language deteriorates, screening for "companies that sound worried" returns the whole market. The question that matters is different: who is deteriorating against their own baseline, beyond the sector tide — and who is hiding in it?
Because our clients are quantitative, we ran the tests any quant would run first.
Test one: do our highest-risk display bands underperform Stable-band filings over the next 60 trading days, during the war window? They did not. The high-minus-stable spread was +0.55% with a two-way clustered t-statistic of 0.18 — statistically nothing — and the 2025 control window shows the same null. We are publishing that number on purpose. Display bands compress a multi-layer composite into a risk-communication scale; they were never the alpha construction, and a naive band-sort will not find alpha in a war, or out of one. The validated performance of disclosure drift comes from specific, pre-registered cohort constructions — the coherent-deterioration methodology published on SSRN (abstract 6444659): a composite disclosure-drift factor of +4.25% (t=9.80) across 24,209 filings, and +7.36% five-factor alpha (t=11.11) on the coherent-deterioration cohort, 2011–2024. The construction is the product. The bands are the dashboard.
Test two: sector-level war-language multiples. An early cut of this analysis showed Materials issuers using direct war language at roughly 8x their prior-year rate. That figure did not survive our own review: it rested on a seasonally mismatched control, an ambiguous term ("escalation") that doubles as inflation boilerplate, and sector cells of fewer than 25 filings. Under matched controls and the strict term list, no sector-level multiple met our publication bar, and we have withdrawn all of them. The market-wide figures above did survive — and strengthened.
We would rather show you a claim we killed than let you find the flaw yourself in week one of a trial. We have killed 52 of our own signal candidates to date under the same standard.
The adaptation quarter is still being written — filings through mid-August will complete it. The analytical question for the next two quarters is the one our platform was built for: as war language settles into standardized risk-factor boilerplate, the informative signal migrates to deviations — the company whose transmission language keeps climbing after the market's has plateaued; the filer whose uncertainty density diverges from its own history with no sector excuse. That is where self-baseline measurement earns its keep.
Institutional teams can now evaluate the dataset on Eagle Alpha and Neudata, including a structured testing period.
*Methodology notes: (1) Language cohorts, seasonally matched 10-Q windows, one filing per ticker: Apr 1–May 31 2025 control (n=398) vs Apr 1–May 31 2026 shock (n=420); Jul 1–27 2025 early-season control (n=289) vs Jul 1–27 2026 adaptation (n=284, season in progress; early-season windows are matched to hold filer composition comparable). Term densities per 10,000 words, whole document. The direct-war group excludes the stem "escalat-" after a sensitivity test showed it contaminated results with routine cost-escalation language; including it, the market-wide shock-quarter increase is +51% rather than 3.4x. (2) An earlier pre-registered trigger criterion referenced cohort medians, which are zero (most filings use no direct war terms); results are reported on means and nonzero shares, and we disclose that wording flaw rather than restate the criterion. Sector-level multiples were withdrawn in review as described above. (3) Band-mix comparisons use only filings scored under identical banding logic in both years; July 2026 band statistics are excluded due to a scoring-scale unification on July 12, 2026. (4) Return test: 60 trading-day forward returns, split-adjusted closes, entry ≥ $2 with volume, entry within 5 days of filing, returns capped at ±300% and winsorized at 1/99, demeaned against the filing universe; standard errors two-way clustered by ticker and filing half-month; of the war-window filing universe, filings without qualifying entry and exit prices (price coverage, the $2 floor, delistings) fall out of the test, and delisting attrition biases toward the null we report. Five-factor neutralization is not yet possible for 2026 (factor data publication lag). (5) This research is for informational purposes only and does not constitute investment advice. Past performance is not indicative of future results.*