The Paragraph They Quietly Changed: Lazy Prices and the Signal Hiding in 10-K Edits

Every year, a public company files an annual report. It is long, dense, legalistic, and almost nobody reads it cover to cover.

Here is the thing that Lauren Cohen, Christopher Malloy and Quoc Nguyen built an entire strategy on: companies do not write a new one each year. They open last year's file and edit it.

Which means the interesting question is not "what does this year's report say." It is: what did they change?

The problem: nobody reads the boring parts

Financial markets are extremely good at reacting to the headline number. Earnings come out, the stock moves within milliseconds. Guidance is revised, algorithms trade it before a human has finished the sentence.

But an annual report is a hundred-plus pages of prose. The risk factors section. The legal proceedings section. The management discussion. The footnotes about accounting policy. Reading all of it, for every company you follow, every year, and comparing it word by word against last year's version, is an enormous, tedious job that essentially nobody was doing.

Which is exactly why there might be something in there.

The key idea via analogy: the friend who suddenly changes their story

Suppose a friend tells you the same story about their job every time you meet. Word for word, the same script: things are fine, work is steady, the boss is reasonable.

Then one day the script changes. They add a sentence. They drop a phrase they always used to say. They start describing their boss differently.

You would notice. And you would be right to notice, because people do not change a settled script for no reason.

Corporate filings work the same way. A company's legal and finance teams reuse last year's language because it is safe, it has been reviewed, and rewriting it is expensive and risky. Boilerplate persists because boilerplate is the path of least resistance. So when a company does take the trouble to change the wording, someone in the building decided the old wording was no longer defensible.

Cohen, Malloy and Nguyen operationalised this. Take the complete history of quarterly and annual filings by US companies. For each company, compare this year's filing to last year's, and compute a simple measure of how much the language changed. Not what it says. Not whether it is positive or negative. Just: how different is it from last time.

Then sort companies into two groups:

• Non-changers, who essentially copy and pasted.
• Changers, who materially rewrote something.

The result is startling. A portfolio that buys the non-changers and shorts the changers earned very large abnormal returns, on the order of up to 188 basis points per month in alpha in their tests, which annualises to over twenty percent.

Read that again. The signal is not the content of the change. It is the mere fact that a change occurred. Language changes are, on average, bad news. Companies rewrite their risk factors when the risks have got worse. They rewrite their legal section when the litigation is real. They rewrite the management discussion when the story they have been telling no longer survives contact with reality.

And now the finding that turns this from a curiosity into an important paper.

There was no announcement effect. When the filing came out, the stock did not move. Nothing happened. The market simply did not react at all, and then over the following months, the changers underperformed and the non-changers outperformed, and the information slowly, grudgingly worked its way into prices.

That is not the usual underreaction pattern, where the market moves a little and then keeps drifting in the same direction. This is the market moving not at all, which means investors were not partially digesting the information. They were not reading it in the first place.

Hence the title. The prices are lazy, because the investors are.

Why it mattered

• It found alpha in the absence of a signal, rather than the presence of one. The insight is inverted from normal research. You are not looking for a stock to say something. You are looking for it to say something different. That is a genuinely novel way to think about extracting information from text.
• It documented pure investor inattention, cleanly. The complete absence of a price reaction at the filing date is about as clean evidence as you will ever get that nobody was looking. Most tests of market inattention are contaminated by the possibility that the market saw the news and disagreed. Here, the market did not see it.
• It exploits a structural, self-reinforcing gap. This information is not secret. It is public, free, filed with the regulator and available to anyone. It is just boring and laborious to extract, and that is a durable moat, because tedium does not attract competition the way secrecy does.
• It changed how quants read filings. After this, textual analysis of filings shifted from "measure the sentiment of the document" to "measure the change in the document." Difference-based textual signals are now a standard family, and this paper is why.
• It gives a very practical instruction to any investor. If you own a company, download last year's annual report and this year's, and run a diff. It takes minutes. The paper says that is one of the highest-value hours you can spend, precisely because everyone else has decided it is too dull.

The honest limitations

• The alpha number is a headline, not a promise. Returns of that magnitude are extraordinary, and they come from a long-short portfolio in a research setting. Real implementation involves shorting the changers, which means borrow costs and constraints, and the effect is likely stronger in smaller and less-covered companies where trading is expensive.
• It has been published, and it is being arbitraged. The strategy was profitable partly because nobody was doing the work. It is now a well-known paper, several data vendors sell filing-change signals, and language models have made document comparison nearly free. The mechanism that generated the alpha, that reading filings is too tedious for anyone to bother, is the exact thing technology is dismantling.
• Change is a crude proxy. The signal counts whether the language moved, not what it means. A company changing its filing because it made an acquisition, changed auditors, or responded to a new regulation is lumped in with a company quietly conceding that its business is deteriorating. The signal works on average precisely because it is crude, but it will misfire on individual names.
• The measurement is sensitive to plumbing. How you strip out formatting, tables, exhibits and boilerplate legal changes materially affects what counts as a "change." Reasonable implementation choices give different answers, and the strategy's performance depends on getting that unglamorous work right.
• It is a slow signal. The information takes months to be priced. That means a long holding period, exposure to everything else that happens in the meantime, and a strategy whose payoff arrives too gradually to feel like it is working while you hold it.

The one-line takeaway

Cohen, Malloy and Nguyen realised that companies copy and paste last year's annual report, so the act of changing the language is itself the signal, and found that changers reliably underperform while the market, having not read the filing at all, does not react on the day and takes months to catch up.