When people hear "SVB was flagged by a language model," a reasonable reaction is: "great, but the bank collapsed in 48 hours — what good is a signal if you can't act on it?"
That's the wrong timeline. Here's the right one.
The 14-day figure isn't "filing to collapse." It's "filing to FDIC seizure" — the endpoint of the collapse, not the start. The stock didn't crash on February 24th. It crashed on March 9th, twelve days after the filing. The window between "10-K hits EDGAR" and "market reprices" was nearly two weeks.
SVB's 10-K was 72,000 words long. The relevant disclosure about the scale of unrealized losses in the held-to-maturity portfolio was buried in the financial statements. It wasn't secret — it was just exhausting to find manually. Twelve analysts covering the stock apparently didn't weight it as a red flag.
This is the actual problem FilingDrift solves. Not "predict when banks will fail." It's: "automatically surface when a filing's language has shifted in ways that are unusual for the company and unusual across the whole corpus, before someone has to read 300 pages of legalese to notice it."
The SVB 10-K flagged because the language around portfolio losses was both more specific and more severe than what was rising across the corpus in 2022. With thousands of companies filing annual reports, you need a corpus-wide lens to notice what's different. A single human analyst covering five stocks can't hold the whole corpus in their head simultaneously. A scoring system can.
We're not claiming the signal is a reliable short-selling trigger. The forward-return backtest shows a real but noisy directional bias — most of the raw vs-SPY underperformance is the small-cap size effect. The distress evidence isn't that absolute return; it's the lift — moderate-flagged companies reach a distress outcome about 1.2× the corpus base rate — and that 58% of flag events have negative alpha at 1 year, vs. ~50% you'd expect from random flagging.
What SVB illustrates is a specific and unusually clean version of something the data shows more broadly: there is often a gap between when the language in a filing changes and when the market reprices the stock. That gap varies. It was 14 days for SVB. It was ~24 months for Bed Bath & Beyond. It was ~44 months for Nikola. The filing is a leading indicator; the market is the lagging one.
The signal tells you that language has changed in a statistically unusual way. It does not tell you when the market will notice, or whether this particular flag will resolve into a distress event or a false alarm. What it does is shift the burden of proof — "this company is fine" now requires explaining away the language, not just assuming the analysts are right.
SVB's 2021 10-K — the year before the final filing — scored 45.4 against a ceiling of 51.5. That's below the threshold: not a flag. But the multi-year arc is visible: 45.4 in 2021, then 9.5 in 2022, then 57.5 in 2023 — the final filing that crossed the line 14 days before the FDIC.
We don't claim 2021 was predictive in isolation. What it illustrates is that the language deterioration wasn't a sudden event — it was a trajectory the scoring system tracked across years. The 2021 score was background noise. The 2023 score was a clear flag.
You can see the full score history on the SVB company page.
Not investment advice. Past detections don't guarantee future results. We analyze language; the interpretation is yours.
Questions: hello@filingdrift.com