← Work

Thesis

Can the difference between how a company discusses its results and how it reports them in writing help predict a stock-price crash?

My role
Author
When
May 2026
Where
Tilburg University
Tools
R, Python, WRDS
What they say and what they writeOut loud, to analystsIn writing, in the annual report2014201720202023Does this gap warn of a stock crash?Illustration, not real data

The question

Every year, the executives of a listed company explain their results twice: out loud, on a call with analysts, and in writing, in the annual report required by the US regulator (the 10-K).

I wondered whether differences in how managerial tone varies between earnings calls and annual reports could be associated with a higher risk of stock-price crashes the following year.

It mattered because, if so, any investor would have a warning sign just by reading two public documents.

How I researched it

I worked with S&P 500 companies, the 500 largest in the United States, between 2014 and 2023. In total, 2,406 cases: each case is one company in one year.

To measure tone I used a financial dictionary that classifies words as positive or negative, and calculated how much the tone swings through each document. The “mismatch” is the difference between how much the call swings and how much the report does.

I then used statistical regressions to see whether more mismatch was followed by more crashes, after accounting for other factors such as company size or profitability.

What I decided

I used the index's historical membership, so as not to study only the companies that survived, and put the result through more than twenty checks before accepting it.

I used AI tools to write and debug the code, as disclosed in my thesis. My work focused on developing the hypotheses, selecting variables and methods, reviewing the analyses and interpreting the results.

What I found

The two main hypotheses were not statistically supported. The models did not provide sufficient evidence that inconsistencies in tone variability predicted future stock-price crashes.

In a supplementary analysis including analysts' questions and answers, a statistically significant association emerged. However, it did not hold across alternative specifications, so I did not treat it as a sufficiently robust conclusion. Grade: 7.5 out of 10.

What I learned

I learned that a statistically significant result can disappear when a methodological choice changes. That's why robustness checks matter just as much as the initial finding.