Pre-restatement 10-K language shows no measurable negative-tone shift, and a simple lexicon signal performs no better than chance

Summary

Comparing the 10-K MD&A filed just before a company's own non-reliance restatement disclosure (SEC Form 8-K Item 4.02) against a comparison group of same-period filers that never restated, a simple negative/uncertainty word-count tone measure shows no statistically significant difference (Welch's t = -0.51, p = 0.61, n=44 vs n=39) and, if anything, points the wrong way: the pre-restatement group scored slightly lower on negative tone (4.7 hits/1,000 words) than the never-restated comparison group (5.3/1,000). At a flagging threshold set to the top quartile of the pooled distribution, the true-positive rate (22.7%) is actually below the false-positive rate (28.2%) — this particular measure would flag more clean filers than future restaters. The underlying literature (Larcker & Zakolyukina 2012; Humpherys et al. 2011; Lopatta et al. 2017) finds real but modest signal using richer linguistic feature sets on earnings-call Q&A or fraud-specific corpora; a blunt keyword-count tone score on MD&A narrative does not replicate that signal in this sample. Restatements themselves are rare: 438 of 8,407 distinct 10-K filers in this corpus (5.2%) had an Item 4.02 restatement filing between January 2023 and August 2026, so even a moderately accurate detector would still be wrong most of the time it fires.

Step 1 — Language measure and corpus

Measure: a 13-stem negative/uncertainty lexicon count (loss, declin-, impair-, weakness, restat-, default, litigat-, investigat-, noncompli-, uncertain, adverse, fraud, irregularit-), matched as case-insensitive substrings against the full Item 7 (Management's Discussion & Analysis) narrative text of each 10-K, normalized to hits per 1,000 words. This is a simplified proxy for a Loughran-McDonald-style negative-tone score, built because the askamerica connector's query layer does not expose a regex/token-count SQL function (regexp_count is unavailable in this deployment) or a pre-computed sentiment column; counts were computed client-side in Python from the connector's raw sec.mda_sections.paragraph_text. It is narrower than the full ~2,300-word Loughran-McDonald negative list and will both miss some genuinely negative language and pick up some benign uses (e.g. 'loss' inside 'lossless') — a limitation of this run, not of the method in principle.

Form type and window: Form 10-K, Item 7 (MD&A) only, filings dated 2022-01-01 through 2025-12-31, drawn from askamerica's sec.mda_sections table (declared coverage 2010-2026).

Step 2 — Adverse-event definition and selection rule (guarding against hindsight)

Adverse event: a Form 8-K Item 4.02 filing ("Non-Reliance on Previously Issued Financial Statements") — the SEC's own bright-line marker for a restatement serious enough that management or the auditor concluded prior financial statements can no longer be relied upon. This is a stricter, less hindsight-prone criterion than a stock-price decline or a discretionary write-down, because it is the company's own regulatory disclosure, not an outcome the analyst chose after the fact by screening for bad performance.

Selection rule (stated in advance, applied mechanically): (1) identify every CIK with at least one Item 4.02 8-K in the corpus (438 firms, 574 filings, 2023-01-20 to 2026-08-06); (2) for each, take the most recent Form 10-K MD&A filed strictly before that firm's own restatement announcement date — i.e., the last annual narrative management wrote before anyone (including the firm itself) had flagged a problem, not the restated filing itself; (3) score that filing's tone. 410 of 438 restating firms had a qualifying prior 10-K in the corpus. Because of the connector's current query-latency constraints on this large a text join (see Limitations), the treatment group actually scored is a subsample of 44 filings drawn from that 410 in two unbiased batches (first 25 in ROW_NUMBER order, next 19 by CIK order), not a cherry-picked set.

Comparison group: Form 10-K MD&A filings from the same 2022-2025 window, filed by CIKs that never appear among the 438 Item-4.02 filers in this corpus. 39 comparison filings were drawn: an initial batch ordered by CIK (biased toward long-registered, typically larger legacy filers — a real limitation, noted below) plus a second batch selected by CIK modulo 401 to spread the sample across the CIK range rather than clustering on a handful of firms.

Step 3 — Measure by group, n and effect size

Welch's two-sample t-test (unequal variances): t = -0.506, p = 0.615 — not distinguishable from no difference. Cohen's d (pooled SD) = -0.11: negligible, and in the opposite direction from the hypothesis that pre-restatement language reads more negatively. The two group means are well within one standard deviation of each other and the distributions overlap heavily.

Step 4 — False positive rate, base rate, and practical usefulness

False positive rate: flagging any filing scoring at or above the 75th percentile of the pooled tone distribution (7.86 hits/1,000 words) as "high risk" catches 10 of 44 eventual restaters (true-positive rate 22.7%) but also flags 11 of 39 comparison-group filers that never restated (false-positive rate 28.2%). A true-positive rate below the false-positive rate means this specific measure is not merely weak — it is uninformative to actively misleading as a screen; a coin flip would do at least as well.

Base rate: 438 of 8,407 distinct 10-K filers in this corpus (5.2%) filed an Item 4.02 restatement between January 2023 and August 2026 — roughly 125/year in this ingested universe, consistent with Audit Analytics' industry-wide count of ~430-480 total restatements per year across U.S. public filers (all severities, of which Item 4.02 "Big R" restatements are the more serious subset). Even a detector with real discriminating power built on a ~5% base rate needs high specificity to avoid a flood of false alarms: at 5% prevalence, a test would need roughly 90%+ specificity just to keep positive predictive value above 1-in-3.

Is the signal practically useful? No, not this version of it. The measure tested here is statistically indistinguishable from noise and, at the threshold examined, has a false-positive rate exceeding its true-positive rate. This does not settle whether pre-restatement language differs at all — the literature (below) finds real, if modest, signal using psycholinguistic deception categories in earnings-call Q&A (Larcker & Zakolyukina: 6-16 points better than random guessing out-of-sample) and specialized fraud dictionaries on hand-labeled fraud/non-fraud corpora, both richer designs than a keyword count on MD&A prose. What this run adds is a direct, corpus-computed replication check of the blunter version of that claim, on real SEC filings selected by a hindsight-resistant rule with a genuine comparison group: it did not replicate here.

Limitations

• Small n relative to the eligible population. 410 restating firms had a qualifying prior 10-K, but the full-corpus LISTAGG join needed to assemble their MD&A text timed out repeatedly against this environment's R2-backed storage (single-filing joins succeeded; the 410-firm join did not complete even at an 8-minute timeout). The reported n=44/39 is an unbiased subsample, not the full eligible set, and is likely underpowered to detect an effect the size the literature reports (Cohen's d in the 0.1-0.3 range is plausible from Larcker-Zakolyukina's headline numbers; this design has limited power to detect d<0.3 at n=44/39).
• Comparison-group sampling is imperfect. The first comparison batch is drawn from low-CIK (long-tenured, typically larger) firms; the second batch spreads across CIK space but is still a convenience sample, not matched on size, industry or filing year to the treatment group the way a rigorous design would.
• Lexicon is a coarse proxy, not the validated Loughran-McDonald dictionary, and uses substring rather than whole-word matching (minor overcounting, e.g. 'restat' inside 'restated' is intended, but a handful of terms like 'loss' can match inside unrelated words).
• Timing: the prior 10-K is the last annual filing before the restatement was announced, not necessarily before the underlying accounting problem began — management's language in the filing immediately preceding disclosure may already postdate the misstatement period itself, understating any true signal that exists earlier in the filing history.

What This Report Does Not Answer

• Cover 'major write-down' and 'sharp decline in reported performance' as adverse-event types, not just restatement: Only the Form 8-K Item 4.02 restatement definition was tested empirically, because it is the one adverse event with an unambiguous, non-hindsight-prone SEC-filed marker in this corpus. A write-down/decline-based definition (e.g. a >50% YoY net-income drop in sec.financial_facts) was scoped but not run, given the query-latency constraints already encountered on the smaller restatement analysis; extending this design to that event type is the natural next step, not a finding that it would behave the same way.