I trained an AI on every accounting fraud since 2005

Every investor runs on beliefs they have never actually tested.

You swear a certain setup works, or that a certain signal matters. Proving it would take a research team and a month, so you trust your gut and move on.

That is the part that changed this year.

You can now run a real study on a market question just by describing what you want in plain English. No code. No model building.

So I tested one of my own hunches.

I have always believed you can see accounting fraud coming if you know where to look.

So I fed 20 years of it to an AI and asked one question:

What actually tips you off before the stock collapses?

Not the story you tell after the restatement, but the signal sitting in plain sight while the stock still works.

What I had Claude Code do

I had AI pull every accounting-fraud case the SEC has brought since 2005.

For each one, we wanted Claude Code to pin down when the fraud started and the day it became public. Pull their filings as first published, build a set of clean companies to compare against, and score every warning sign that AI can find. But only count what was knowable before the fraud broke.

With a Q&A loop, here's what AI actually did:

1. Read the SEC’s fraud cases one by one, pulling the company, the fraud years, and the reveal date.
2. Added the big restatements and the short-seller reports that later proved right.
3. Went back to the filings as first published, before any numbers were rewritten.
4. Built a matched set of clean companies for each fraud, same industry and size.
5. Scored dozens of warning signs across all of them, dated to the day.
6. Checked which signs showed up before the reveal, and how far ahead.

The AI read close to 89,000 annual reports this way. A forensic accounting desk would need a year and a team to do the same thing.

Lesson 1: read the 8-K stream, not just the ratios

The three tells that topped the list are things management does, not things the numbers say.

• A prior “our financials can’t be trusted” restatement in the last two years: 3.3 times the odds of fraud
• A late filing, meaning an NT 10-K or 10-Q: 3.3 times
• An auditor change: 2 times

A company telling you its old numbers are wrong is not a subtle signal. Neither is an auditor resigning. We just do not systematically watch for either one.

Lesson 2: who catches fraud, and how early

I had the AI reconstruct who first surfaced each fraud, using only sources published at the time.

Short sellers and journalists were first in about 62 percent of cases. Auditors and the SEC, combined, were first in 6 percent.

By the time there is an enforcement action, you are the exit liquidity.

And the signal shows up early. The median lead time on those disclosure tells was 19 to 21 months.

Take the subprime lender NovaStar ($NFI). Herb Greenberg wrote his first column questioning whether the earnings were backed by cash in February 2003.

The stock did not blow up until 2007. That is a 48-month head start, printed in public, for free.

The harder proof came a year later, when regulators found its branch network was largely unlicensed or did not exist. Still three years early.

None of it needed private data. It needed someone reading what was already public.

Why this is hard: fraud is rare

Here is the catch that makes all of this difficult.

Fraud is a 2 to 3% event in any given year. Most companies with an ugly ratio are just having a bad year, not cooking the books.

So the best screen I could build ran at about 1.3 times the base rate in its top decile. Point it at one name and you mostly get false positives.

You also only catch about a third of fraud in the first place. A high score is a place to look, not a verdict.

This is why hunting one ticker at a time is a losing game. The math is against you.

The part that actually matters for you

The edge here is not a smarter formula. It is a system that never sleeps (via Claude Code or Codex skills automations).

Point AI at the whole public event stream. Every late filing, every auditor change, every non-reliance restatement, every finance-chief exit, across all 13,500 US filers.

Let it read every filing the day it drops and keep a running flag count on every name.

Because fraud is rare, that flips the problem. Instead of guessing one name at a time, you get a ranked, self-updating shortlist of the companies stacking flags right now.

You stop hunting and start triaging.

The machine does the breadth no human can match. You do the forensic work on the twenty names at the top.

That is the real edge. Not an oracle that names the fraud, but a tireless reader that narrows 13,500 names down to a watchlist you can actually cover.

What actually changed

The lesson under all the lessons is about what just became possible.

AI read every enforcement case, every restatement and 89,000 filings, and held the whole thing in memory at once. That used to be a team of forensic analysts and a year of work.

Now it is a weekend and a clear question!

Personal

Update on my Claude Code for Investors webinar.

We are down to the last 22 seats before we hit our cap. We keep it capped so that every question actually gets answered.

We have a solid crew signed up on this cohort including folks from Millennium, Temasek, Hudson Bay Capital, and more.

If you're looking to learn how to master Claude Code for buyside, this is your shot.

Check out the link for more details and let me know if you have questions!