Uncooking the books 

Written by: Nicholas Blain

One of the most interesting tasks of financial analysts is trying to understand when a company is “managing” the earnings.

You can interpret the word “managing” as diplomatically as you like, but think of it here as a deliberate attempt to show a manipulated set of financial results.

Candidates at CFA Levels I and II learn about red flags, or warning signals, when investigating the meaning of accounts. In the March/April edition of the Financial Analysts Journal of CFA Institute there is an article on “Earnings Manipulation and Expected Returns” (Beneish, Lee, and Nichols, 2013). In this the authors consider an earnings manipulation detection model (Beneish, 1999a) and the relationship between the probability of manipulation, what is called the M-score, and the returns that the company generates.

Lots of fascinating insights come from this, not least of all the high likelihood of manipulators earning lower returns after publishing such accounts.

A typical manipulator is a company that (1) has extremely high sales growth, (2) shows deteriorating asset quality, profit margins and increasing gearing, and (3) adopts aggressive accounting practices, such as higher accruals (earnings minus cash flow), lower depreciation and higher receivables relative to sales.

In summary, the results of the testing were dramatic: the authors listed companies by (i) size, (ii) book-to-market, (iii) momentum, (iv) accruals, and (v) short interest. For each of the five criteria, and within all ten deciles, companies with higher M-scores earned lower returns. The difference in earnings was significant too, nearly 1% (risk-adjusted) per month.

Curiously, the original model was able to predict a large majority of famous accounting fraud cases before these situations became public. Examples of these include Enron (flagged in 1998, “discovered” in 2001), Global Crossing, Qwest Communications and Sunbeam. No model can be entirely accurate – the Beneish model did not flag WorldCom or Tyco, classic Type I errors.