Archive for February 25th, 2009

To catch new Ponzi schemes like Stanford and Madoff, the SEC needs a geek squad

February 25, 2009

The recent frauds by Bernie Madoff and Alan Stanford were not caught by the SEC in time to save people’s money. In Madoff’s case, he gave up, as his pyramid was collapsing and confessed (I still wonder why he didn’t fudge the collapse). In the case of Stanford International Bank (SIB), the SEC was investigating the illegal sale of CD’s suspecting that the returns were too juicy.

If Stanford had not tried to sell his CD’s in the US, he would have not been caught, except for the keen eye of Alex Dalmady, who not only saw that there was something funny about the whole set up at SIB (Many of us did), but actually sat down and wrote about it (Which we didn’t).

The question is what can be done about it going forward? How can the SEC monitor for frauds better?

Well, off hand (and on vacation) it occurs to me that given the mathematical tools available these days, something as simple as setting up a geek squad, a bunch of mathematically-oriented whiz kids who would go and devise a bunch of tests to dig out possible Ponzi schemes that the SEC could then investigate further.

I can think of three very simple tests off hand:

1) Benford’s law

I have talked about this in connection to Venezuela’s election, which have been shown to follow the law except for the 2004 recall referendum. Very simply, when you fudge data, you ignore the fact that natural data has certain characteristics. In particular, Benford’s law says that in any list of numbers generated for example, by accountants, the distribution of the first (or second) number (from left to right) follow a distribution which is not uniform. In particular, the distribution for the first number is:


That is, the number “1” has a 30.1% probability of occurring, number “2”, 17.6% and the rest of the numbers declining from there.

Why is that?

Because real world numbers tend to be distributed logarithmically and not uniformly. Consider the following: If a company issues checks between 0 and $100,000 and you look at the first number of the amount for each check, it is likely that the number of checks near $100,000 is low and starts going up as the amount gets lower. Well, if the company prints lots of checks, then you could detect fraud if more than expected show up near $100,000 or if a particular number shows up a lot (a common occurrence). This is actually used in accounting to detect fraud. This also applies to investment returns.

Thus, the geek squad at the SEC could simply look at the daily, weekly or monthly performance data supplied by all regulated mutual and hedge funds and compare it to what Benford predicts. Of course, these tests can be done using more sophisticated algorithms, using statistical measures on the first and second digit to detect discrepancies.

Paul Kredosky has actually looked at Madoff’s data and suggested that Madoff’s data did follow Benford, but his conclusion was only visual. Falkenblog actually concluded that the data did not fit Benford’s law. I would suggest the SEC geek squad could carry out a first and second digit test and calculate statistical measures like chi^2 on the differences between the data and what is expected to see how likely the returns are. (Anyone willing to do it? The data is in Markopolos 2005 document to the SEC denouncing Madoff)

Of course, if the returns of a fund did not follow Benford, there may be an explanation, but detecting it this way would allow for a more detailed study of the funds returns by the SEC.

2) Correlations between returns and markets

One of the red flags raised by Markopolos on Madoff was the fact that Madoff’s strategy was based on the stock market, but there seemed to be little correlation between Madoff’s results and the stock market. Indeed, Madoff showed positive returns on 95.5% of the months, which was not happening in the stock market.

This could all be automated.Each fund would simply define its strategy and set a benchmark for its investments and you could calculate the Correlation Coefficient between the returns of the fund and that of the underlying markets in which it participates. This can be done in an Excel spreadsheet. Basically, it would be very difficult to obtain returns which are uncorrelated to the underlying markets. As an example, last year, it would be suspicious if a fund investing in stocks had a negative correlation with the market unless it invested in gold stocks or a sub-sector of the market that did well last year, but there were very few of those.

Finding anomalies in the correlations does not constitute proof that someone is reporting fraudulent results, but one could automate the process and much as in the case of Benford’s law, it would raise a red flag and the SEC could study it further.

3) A stress test for accountants of funds

One could build two databases, In the first one, one would include the accountants who are active part of the American Association of Certified Public Accountants. A second one would include the top tier auditing firms. Then, for each fund you would ask:

a) How much did you charge to audit the fund?

b) Did you confirm who had custody of securities?

c) Ask who managed and who had custody of securities. Is there any relation between the two?

You could then compile statistics of how much do auditors charge for each size funds. And if b) was No, you would have a big red flag, as well as if in c) the answer was yes. Similarly, if there were anomalies in Benford or the correlations and you used a less known auditor, you could research it further.

That’s it. With these three points, you would go a long way at uncovering the fraudulent funds.

Using these three questions you would have caught Madoff with a), b) and c). However, only c) would have raised flags on Stanford, because there is not month to month data on Stanford’s returns, only yearly data.

But I am sure others can add some additional forensic tools to the detection of fraud in money management. I do hope the SEC may read this and start a geek squad. It would cost very little and go a long way.

And indeed, funds could learn to fudge the data going forward to fit the criteria, but their previous record is there for the geek squad to find.