saisat wrote:
>
> Hello guys,
>
> i am trying to profile a customer based on his transaction activity
> history. This is done so that i can compare the profile (average
> activity) against the current activity and determine if the current
> activity falls within the limits of the profile. If it doesnt, I say
> it is suspicious.
>
> But if the customers were involved in suspicious activity in their
> past, it skews the profile in such a way that even their profile is
> suspicious. So there is no way for me to determine if their current
> activity is suspicious.
>
> So I was wondering if there is a way to determine (statistically)
> their average activity (somehow excluding the suspicious)
>
> Regarding your questions
> 1. I cannot eliminate them manually because i wouldnt know what to
> eliminate.
> 2. the sample may range from thousands to 100 thousands. sometimes
> millions. It is the transaction history over the past year(s) of a
> customer.
This looks like a case for a trimmed mean. Throw away the top 5% and
the bottom 5% and average what's left.
Or possibly a trimmed geometric mean. Financial data often fit a
lognormal model better than a normal model, which makes the geometric
mean more representative in some ways.
BUT financial data also often have the feature that one is genuinely
intersted in the sum, in which case the (arithmetic) mean is the way to
go.
-Robert Dawson
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
