fraud returns

Tradenator · Post by **Tradenator** » Sat Apr 18, 2009 8:53 am

Fiction, in the literal sense. I think they just made up numbers and put them out on funds databases. Most likely, there arent't any investors even. They have several funds, and one is genuine but at least one other isn't.

meteor · Post by **meteor** » Thu Apr 30, 2009 5:20 am

Not implying anything, have you guys looked at this:

http://www.scribd.com/doc/14713748/PontaNegraFeb09

Truly impressive result: not even a down month.

iparkins · Post by **iparkins** » Thu Apr 30, 2009 9:44 am

worth reading about them at www.brontecapital.blogspot.com - looks like they belong in this thread

dehaan · Post by **dehaan** » Mon May 11, 2009 9:33 am

>Funny, RiskData got the Financial Times to trumpet their new fraud detector (aka bias >ratio) to the skies as something which will make all your whites whiter. I'll never believe >anything I read there again. If they're just reissuing marketing crap, I may as well read >the Weekly World News.

Journalists being journalists...BR is a statistical indicator and should be understood as such. That is if it is , say, 6 on a fund, it means that there could be issues with returns adjustment. But it may be as well that the fund is clean. And you would do your qualitative analysis more closely only on those funds that have big BR, that are suspicuous, that's all. This indicator becomes handy when you screen, say, a thousand of hfs.

Nonius · Post by **Nonius** » Wed Jul 29, 2009 11:17 pm

we tried Benford on buttloads of in-out samples of 1000s of HF returns. It sucks. Aaron, if you can think of a good test for null other than Chi-Squared, I'm all ears.

my method works well, but maybe there's some weird bias.

aaron · Post by **aaron** » Tue Aug 04, 2009 3:53 pm

You mean it "sucks" because all hedge funds pass? That's not really surprising. Benford would only catch people making up returns in their heads, who didn't know about the concept. My guess is most frauds would report honest returns when they were acceptable, but trim losses and maybe some big gains as well. They might sometimes report a great return to stimulate more money. But since they're starting from a real number and adjusting, the digits are likely to be random.

Remember, unless you're Madoff, you have to show people numbers that add up, and some of which match publicly available ones. So you cheat by having some "ABC Trust" or exotic derivative that you assign a high value to, then add it up with all the rest.

I think Chi-square is the good test for the null hypothesis, but you have to be sure to adjust the Benford prediction for the range of returns.

doctorwes · Post by **doctorwes** » Tue Aug 04, 2009 7:32 pm

Suppose somebody took a genuine return series (e.g. for the S&P) and added a small constant x% to every value. What kind of test would pick that up, assuming you did not have access to the original series?

aaron · Post by **aaron** » Wed Aug 05, 2009 1:06 am

None. Even with the original series, there are strategies expected to have a similar return pattern (if the number were exactly the same every day, of course, it would be suspicious).

However, when people cheat, they're more likely to reduce the variance than increase the mean; or they do both. Reducing maximum drawdown and volatility can make the fund seem much more attractive, and means you don't need a lot of fictional cash for a long period of time.

adas · Post by **adas** » Wed Aug 05, 2009 2:29 am

Nonius - where are you sampling your fund returns from: HFR, TASS, CISDM or elsewhere?

Have you considered smoothing your returns to remove serial correlation? The Benford test fails if returns are clustered. Suppose a fund reports the following returns: 8, 8, 9, 13, 9, 9, 12 - clearly Benford's test is not suitable here.

Nonius · Post by **Nonius** » Wed Aug 05, 2009 9:06 pm

You mean it "sucks" because all hedge funds pass?

no meaning sucks as in for the sample of funds that we know are fraudulent and not fraudulent, there seems to be no correlation between pass or fail and fraud and not fraud. anyway, I'd prefer a model that assigns a probability of fraud, not some metric that you'd have to trust with a digital answer assuming the Chi Squared rejection works.

adas, i'm using an in-house DB, which is actually better than those commercially available DBs. I didn't think about removing serial correlation although I did notice that with my fraud detector, it is probably giving false signals on returns with high autocorrelation. (actually the two theories behind autocorr are a) PL smoothing, which may be fraudulent and b) illiquidity of assets.)