No LLMs. No hallucinations. Pure linguistic drift.

We flagged SVB's collapse
14 days before it happened

Every year, companies bury the bad news deep in their annual filings. We read them all — and score how the language shifts year over year, relative to every peer filing that year. When SVB filed, 23 out of 24 analysts rated it Buy or Hold. Our score was already above the distress ceiling.

Built for credit analysts, quant researchers, and risk-aware portfolio managers.

See the SVB story → View all companies
40
Companies tracked
677 days
Avg lead before event
80%
Correct non-crisis rate
44.0
95th pct. control ceiling
Case Study: SVB Financial Group

The signal was in the filing.

On February 24, 2023, SVB filed its annual 10-K. The stock was trading at $267.83. FilingDrift scored that filing 53.1 — above the control ceiling of 44.0.

14 days later, the FDIC seized the bank. The stock hit $39.49.

See full analysis →
SVB Distress Score — Year over Year
Red line = control ceiling (44.0)

Crisis Detection Results

Scores relative to the 95th percentile of healthy companies (ceiling = 44.0). Crisis companies have a verified collapse event.

Company Event Peak Score vs. Ceiling
PRTY Party City Holdco Bankruptcy 210.6 4.8×
BBBY Bed Bath & Beyond Bankruptcy 137.4 3.1×
NKLA Nikola Corporation Bankruptcy 116.3 2.6×
REVLON Revlon Inc Bankruptcy 87.3 2.0×
RITEAID Rite Aid Bankruptcy 79.0 1.8×
CHKAQ Chesapeake Energy Bankruptcy 62.6 1.4×
PCG PG&E Corporation Bankruptcy 53.8 1.2×
SVB SVB Financial Group Bank collapse 53.2 1.2×

How the score is built

📄

Read every 10-K — all 200 pages

We ingest every annual SEC filing via EDGAR — risk factors, MD&A, liquidity disclosures. The full document. No truncation. No summaries.

📊

Measure linguistic drift

What phrases appeared for the first time? What escalated? We compare against every peer filing that year — so we only flag language that is genuinely unusual for that moment in time.

🧠

Find what keywords miss

Our scoring algorithm detects language that has drifted toward distress — even when no specific red-flag word appears. The CFO's fingerprints, not just the obvious phrases.

🚨

A score, not a summary

The score is deterministic — run it twice on the same filing and you get the same number. No generative model, no prompt engineering. SVB's filing always scores the same. The FDIC arrived 14 days after that score was computable.

The signal Wall Street missed

When our algorithm flagged these companies, the analyst consensus was still bullish. The language in the filings told a different story.

SVB Financial — February 24, 2023

SVB's 10-K showed statistically abnormal language patterns — score 53.1, 1.2× the distress ceiling. The FDIC arrived 14 days later. On the day that filing was processed, analyst consensus was: 12 Buy / 11 Hold / 1 Sell. Forbes had just named SVB "#20 America's Best Banks."

53.1
FilingDrift score
12 Buy / 11 Hold / 1 Sell (consensus)
Bed Bath & Beyond — April 22, 2021

BBBY's 2021 10-K scored 137.4 — 3.1× the distress ceiling — while the stock was still up 25% in the meme rally. The score flagged the filing two full years before the April 2023 bankruptcy filing. Price action hid the signal. The language didn't.

137.4
FilingDrift score
2 years before bankruptcy

Sources: Federal Reserve OIG Material Loss Review — SVB (Sep 2023); BBBY SEC EDGAR filings.

How it compares

FilingDrift Bloomberg Terminal FactSet / CapIQ LLMs (GPT, Claude)
Peer-normalized scoring (vs. all peers, same year) Partial †
Deterministic, reproducible score Varies Varies ✗ — stochastic by design
5-year historical language arc Varies by license ✗ — context window only
Automatic new filing alerts Pro tier
Annual cost Free — $3,588/yr ~$24,000/yr ~$12,000/yr Free / $240/yr

† FactSet has quantitative sentiment tools but not cross-company phrase-level normalization in the same year. Bloomberg/FactSet pricing is approximate.

Why not just use ChatGPT? LLMs analyze one document at a time with no memory of what other companies said that year. They cannot tell you whether SVB's language is unusual relative to its peers — they have no peer data. They also produce different outputs on the same input on different runs, which makes auditability impossible. FilingDrift's score is computed once and doesn't change.

Validation results — tested on 40 companies

We tested the algorithm on 10 companies with documented collapse events and 30 healthy controls. Results are limited by corpus coverage — some crisis companies had too few pre-event filings in EDGAR to score.

80%
Specificity
Healthy companies correctly cleared
677 days
Avg lead time
Before the collapse event (detected cases)
5 / 10
Pre-event recall
5 others had insufficient filing history
45%
Precision
Above-ceiling flags that preceded real events

The 5 missed cases: PCG and REVLON had no pre-event filings in our corpus; CFC, SI, and CHKAQ had 1–2 filing pairs with insufficient pre-event data. Precision is bounded by the small crisis sample; we are expanding the corpus. Full methodology at About.

Disclaimer: FilingDrift provides automated linguistic analysis of public SEC filings for research purposes only. This is not investment advice. Past detection of distress events does not guarantee future accuracy. See Terms of Service.

Get filing alerts

Enter your email and we'll notify you when a new 10-K comes in with an above-ceiling score. No noise — only signals that clear the bar.