Lazy Prices Meets Survivorship: Why a Disclosure-Text
Anomaly Inverts in a Survivor-Only Sample

In one line: I tested 14,155 filings to see if rewritten wording predicts losses. The result flipped, because failed companies were missing.

Abstract. Cohen, Malloy and Nguyen (2020) show that firms which change the language of their periodic filings subsequently underperform. Disclosure inertia is good news. I test the effect on 14,155 EDGAR 10-Q and 10-K filings for 406 current large-cap index constituents, measuring text change as one minus the year-over-year cosine similarity of a filing to the same period a year earlier. The effect does not replicate. The measure is highly specification-sensitive: a naive quarter-over-quarter version is a precise zero (t = 0.13), while a seasonally-controlled, stopword-cleaned version is significant but opposite in sign. The firms that changed the most slightly outperformed (a low-minus-high spread of −1.97% over the next quarter, Newey–West t = −2.55), and even that is well below the museum’s stressed promotion bar. I argue the inversion is a survivorship artifact. A current-constituents panel deletes the informative short leg (the firms that rewrote their disclosure on the way to trouble), leaving only the changers that survived. The anomaly cannot be cleanly tested without survivorship-free data. No signal is promoted.

1. The anomaly, and why it should be orthogonal

Most published equity signals are functions of price. The Lazy-Prices effect is not one of them. It reads the language of a filing. When a firm reproduces last year’s disclosure nearly verbatim, nothing material has changed and the stock does well; when it rewrites, something is being managed, and the stock underperforms. The mechanism is forensic rather than statistical, in the tradition of reading the narrative for what the numbers omit (Mayew, Sethuraman and Venkatachalam, 2015). An orthogonal, lightly-crowded signal of this kind is exactly what a price-exhausted research program should want to test.

2. Data and method

I harvest every 10-Q and 10-K primary document filed since 2015 by the 406 names in my universe with EDGAR coverage (14,155 filings in all) and reduce each to a stripped-text word vector, discarding inline-XBRL scaffolding and high-frequency stopwords that every filing shares. Following Cohen, Malloy and Nguyen, text change is measured year over year, same period: a 10-Q is compared to the same fiscal quarter a year earlier and a 10-K to the prior year’s 10-K, so that ordinary seasonal differences in disclosure are not counted as change. Each filing’s market-adjusted return is measured over the 63 trading days beginning the day after the filing. Everything is point-in-time; the universe is current constituents, so it is survivorship-biased by construction.

3. Results: fragile, and the wrong way round

The measure does not survive its own robustness. Compared quarter over quarter, the low-minus-high-change spread is an exact zero (t = 0.13). Compared year over year and cleaned of boilerplate, the same spread becomes significant and reverses. The least-changed filings underperform and the most-changed slightly outperform. That is the opposite of the published effect, and a signal whose sign depends on the specification is not a signal one trades.

4. The mechanism: survivorship deletes the short leg

The inversion has a clean explanation. Lazy Prices is, in large part, a short-side result. The money is in avoiding, or shorting, the firms that rewrite their disclosure because trouble is coming. Those are precisely the firms that underperform, are removed from the index, and in many cases delist. A panel of current constituents has already deleted them. What remains among the heavy changers are the survivors: firms that rewrote their filings while growing, restructuring successfully, or acquiring, and lived to stay in the index. Conditioning on survival keeps the benign changers and discards the malign ones, which is enough to flip the sign of a short-heavy anomaly. The result in Table 1 is therefore most likely a measurement of who survived, not of what disclosure change predicts.

5. Verdict

Nothing is promoted. The strongest specification reaches t = −2.55, far short of the stressed, multiplicity- deflated bar this museum requires (near a raw t of 5), in the wrong direction, and with a mechanism that points at survivorship rather than information. The honest conclusion is not that Lazy Prices is false (the original result stands on a survivorship-free universe I do not have), but that it cannot be tested here. Reading the language of filings remains a promising orthogonal lane; testing it properly requires point-in-time constituents and delisting returns, the data this program has repeatedly identified as its binding constraint.