A study set to appear soon in the Journal of Accounting Research asks whether deceptive chief executives and chief financial officers give themselves away by the language they use in the course of their conference calls.
David F. Larcker and Anastasia Zakolyukina, both of Stanford University, Graduate School of Business, have developed prediction models (though still rather crude ones, they acknowledge) based on the “linguistic features” of such calls, working with a comprehensive set of transcripts collected by FactSet Research Systems Inc., which they compare with restatements identified by Glass Lewis. They use subsequent restatements at variance in some way with what was said during the conference call as the mark of deception.
But of course investors want to know whether the executives are speaking deceptively well before the restatement is filed. In order to help with that goal, Larcker and Zakolyukina in the paper, “Detecting Deceptive Discussions in Conference Calls,” have analyzed the transcripts, focusing especially on CEOs and CFOs because they are “the most likely executives to have knowledge about financial statement manipulation.” They have broken down communication into what they called “instances;” that is, communications (often answers to a caller’s questions) in which a CEO or a CFO spoke for more than 150 consecutive words.
They found that certain linguistic mannerisms are “significantly associated with the likelihood of deception.” For example, references to what “everybody knows.” For both CEOs and CFOs, as references to what everybody knows increase, so does the likelihood of deception. Can we quantify how much? That depends; it turns out, it depends on what particular standard for deceptiveness one employs. The authors distinguish between deceptions that merely register as non-trivial (NT) and those that lead to an SEC Accounting and Auditing Enforcement Release (AAER). Every AAER is NT, but a lot of NT deceptions never rise to the level of AAER.
So, what is the quantitative significance of the expression “everybody knows” and synonyms? When the instances come out of the mouth of the CEO, the increase in the use of such phrases by 29 increases the likelihood that deception is involved by a factor of 1.91 for NT and by 1.98 given the AAER standard. From a CFO, the use of such phrases is a bit more predictive. An increase by only 19 increases the likelihood of deception by a factor of 2.05 for NT and by 2.47 for AAER.
Another important tag is the use of what Larcker and Zakolyukina call “non-extreme positive emotion words.” Here they began with the presumption that people who are seeking to deceive tend to ramp up positive emotions, so that “happy” becomes “ecstatic” and “good” becomes “fantastic.” Honest managers, then, can be expected to make use of more instances of non-extreme emotion words: fewer “fantastics” and more plain old “goods.” That presumption matched the data. “[B][oth executives used significantly fewer non-extreme positive emotion words … in deceptive narratives,” although this was a stronger trend for CFOs than for CEOs.
In the case of both of those verbal tags, the result seems (to a non-scholarly fellow like me) common sensical. Executives who are confidently in command of the facts and who are making a case supported by those facts wouldn’t need recourse to what “everyone knows,” since they will be making use of what they specifically know. Also, executives in that happy situation will feel good about it, but won’t have to pretend they are feeling fantastic.
Attempting to Control
But Larcker and Zakolyukina seek the whys and wherefores of their findings. Do these tags have predictive value because the executives are afraid of getting caught, and their anxiety shows up in the way they talk? The authors think not. In that case, words that convey negative emotions would be predictive, too, and the data don’t show that.
Rather, these authors come down in the side of what they call the “attempted control perspective.” The words that are used, are used because the executives involved are trying to control the nature of the discussion. Their need for control becomes greater as the deception involved becomes greater, which is why the positive correspondences are more positive for an AAER level deception than for a mere NT.
What about alpha? Is the predictive value of the kind of linguistic correlations discussed in this paper sufficiently reliable to guide portfolio decisions? One would expect to see “negative returns in the months following a ‘deceptive’ conference call if the market gradually learns about misreporting” thereafter, as these authors note.
They crunch the numbers the best they can and conclude that their results are economically significant if you make your portfolio decisions based upon what the CFO is saying, but that “alpha estimates for the portfolios based on the CEO narratives are insignificant for all cutoff-deception criteria combinations.”
Perhaps this is due to the different skills sets required for the two positions. Perhaps, to put it bluntly, CEOs are sufficiently careful about what they say that their ‘tells’ aren’t reliable, at least not in the rather “coarse” state of research into these correlations thus far achieved.