Dunno about banks, mate, but I've built a Credit Risk system for Visa in my time, and it paid no attention whatsoever to all that crap.
Of course, a credit-card risk analysis system is nicely constrained by the fact that, well, you need a credit card number to commit fraud. Building a history off this is naturally quite simple. All you have to do then is to match it against a Model (not my capitalisation; Visa's) which tracks against generic behavioural profiles.
(In the case of credit cards, it's quite easy to build a basic Model. At one end you've got unusual, but possibly legitimate, behaviour -- buying a diamond-studded camel saddle, for example, if you live in Neasden -- and at the other end you've got typical fraudster behaviour, a simple example of which is to use the credit card to buy ten quids' worth of petrol fifteen minutes after you've ripped off £1000. There are subtleties, but you don't need to know them, and I can't tell you because of black helicopters and such.)
Crossing over to poker sites, I don't see a huge behavioural difference. Naturally, you're going to want the Model to be Bayesian (I know how much you love that stuff), and naturally you're going to want to compare profile against profile, rather than a single history against a single profile. Catching a bot ring using this sort of technique should be fairly straight-forward: in fact, in the case you quote, almost automatic and instantaneous (if there's any such thing as "instantaneous" in a Bayesian world).
The Model would, I estimate, require around 50,000 lines of Java or C# or C++, and configuration files -- to be updated as necessary -- stretching to maybe 5000 lines. Given the huge granularity of poker transactions (I mean, you have to play the whole hand, not just stick a piece of plastic in a slot), I estimate from experience that you could probably fit the whole thing on a mid-range Solaris with report-time of, say, a couple of minutes from detection. You're looking at £100,000 for hardware and a couple of thousand man-days at £300 per (contract rates, UK) ... actually, now I come to think about it, £700,000 is quite a lot of money.
Particularly if bots don't affect your bottom line directly.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
no subject
Date: 2010-07-18 04:34 pm (UTC)Of course, a credit-card risk analysis system is nicely constrained by the fact that, well, you need a credit card number to commit fraud. Building a history off this is naturally quite simple. All you have to do then is to match it against a Model (not my capitalisation; Visa's) which tracks against generic behavioural profiles.
(In the case of credit cards, it's quite easy to build a basic Model. At one end you've got unusual, but possibly legitimate, behaviour -- buying a diamond-studded camel saddle, for example, if you live in Neasden -- and at the other end you've got typical fraudster behaviour, a simple example of which is to use the credit card to buy ten quids' worth of petrol fifteen minutes after you've ripped off £1000. There are subtleties, but you don't need to know them, and I can't tell you because of black helicopters and such.)
Crossing over to poker sites, I don't see a huge behavioural difference. Naturally, you're going to want the Model to be Bayesian (I know how much you love that stuff), and naturally you're going to want to compare profile against profile, rather than a single history against a single profile. Catching a bot ring using this sort of technique should be fairly straight-forward: in fact, in the case you quote, almost automatic and instantaneous (if there's any such thing as "instantaneous" in a Bayesian world).
The Model would, I estimate, require around 50,000 lines of Java or C# or C++, and configuration files -- to be updated as necessary -- stretching to maybe 5000 lines. Given the huge granularity of poker transactions (I mean, you have to play the whole hand, not just stick a piece of plastic in a slot), I estimate from experience that you could probably fit the whole thing on a mid-range Solaris with report-time of, say, a couple of minutes from detection. You're looking at £100,000 for hardware and a couple of thousand man-days at £300 per (contract rates, UK) ... actually, now I come to think about it, £700,000 is quite a lot of money.
Particularly if bots don't affect your bottom line directly.