Hope this is mathematical enough to qualify as a question - I'm no mathematician!
I have a set of individuals travel & entertainment credit card spend, and I'd like to highlight any outliers that are worth scrutinizing further. The type of info I have available are things like transaction date, category, dollar amount, the merchant's name, the user's description of the purchase, the person's manager, whether it was card or 'cash claim', comment text relating to the item, etc.
My initial simple idea was to look at the average and stddev for each category (eg a 500 dollar plane fare is not unusual but a 500 dollar cab fare is), and report those above x standard deviations. Or, divide them into deciles and report only the top ones.
Is there any standard practice around using one of these methods, or a totally different one to do it?
If anyone has any comments around possible other 'clever' ways to identify unusual or fraudulent transactions I'd appreciate it - eg clustering, frequency analysis, neural nets etc. I will probably ask a separate question relating to applying benford's law to this type of data also.
PS. It's probably worth mentioning I have $60-100M worth of transactions, spanning 4 years across several thousand people in total. So it should be enough records to do something interesting with!