Detecting corporate fraud is no simple task.
But now, a pair of Canadian researchers have designed a computer program that can, with stunning accuracy, pinpoint dishonesty by detecting specific words and word configurations in the management discussion and analysis (MD&A) sections of annual and quarterly financial reports.
“It’s not just so much that these words appear, but they have a different pattern of appearing in truthful versus deceptive messages,” said Lynnette Purda, a business professor at Queen’s University in Kingston.
Prof. Purda, and David Skillicorn, of the university’s school of computing, who has a background in counterterrorism detection, built a detection algorithm that looked at all the words in a sample of 4,895 fraudulent and truthful reports submitted with the U.S. Securities and Exchange Commission between 1994 and 2006.
The program counted word frequency and patterns, but then standardized the results regardless of report length. Some MD&As are a few paragraphs, while others can run many pages. While the section provides interesting information about the company’s past, present and future, it is not audited.
The program correctly classified about 87 per cent of MD&As as either fraudulent or not. Of course, it was not foolproof. It also pumped out some false positives: truthful statements that were incorrectly classified as fraudulent. But interestingly, those incidents tended to cluster around quarters when actual fraud had occurred.
For example, the program flagged Enron Corp. three quarters before any actual fraud occurred, and zeroed in problems at Adelphia Communications Corp. four quarters in advance of fraud there. Both companies are now bankrupt.
This is Fraud Prevention Month, and the RCMP estimates that fraud-related crimes cost between $10- and $30-billion a year - as lucrative as drug offences. While the vast majority of frauds are committed by organized crime, they can also be masterminded by a few key executives.
“The people that are tasked with trying to find fraud, they struggle,” said Prof. Purda, “Regulators struggle. Auditors struggle. We still see a lot of frauds identified by whistleblowers and the media - and they are not the ones sitting down crunching the numbers and analyzing the reports.”
Traditionally, linguistics and psychology experts have generated word lists to help predict when someone is lying. A lack of personal pronouns or more action-oriented words can sometimes be indicators of fraud. But the Queen’s University researchers found that those words lists aren’t helpful in predicting truth versus fraud in the MD&As. That’s because often the people writing the report aren’t actually aware of the activity, or those that are might avoid using words generally known to be red flags.
Instead, this algorithm spits out a data set, which leads to the researchers to the problematic language.
They found words such as “acquire,” “settlement” and “legal,” were highlighting situations where fraud was likely to occur. Meanwhile, other words, such as “believe” and “significant,” which would be normally dismissed as meaningless, were actually buzz words worth noting.
However, the average investor - or corporate leader - won’t be able to scan reports and identify fraudsters using those specific words as guides. And even if this research, which will soon be presented at the computing conference, is commercialized for application in the real world, it isn’t a panacea.
“There’s not going to be a single tool or a single silver bullet that will put an end to corporate fraud in either your own corporation or other corporations that you’re dealing with,” Prof. Purda said, “What we really need is to develop a suite of tools and part of that is analyzing the numbers, part of that is protection for whistleblowers ... and language we feel is a very important piece as well.”
The research was funded the Social Sciences and Humanities Research Council of Canada as well as the CA-Queen’s School of Business Centre for Governance, a joint effort between the university and the Chartered Accountants of Ontario.