2

If I'm watching a series of accounts for transactions going in and transactions going out, I want to notice unusually large or transactions for any particular account on any particular day.

So if account A typically moves a few hundred dollars and one day moves five thousand dollars, that's a clear outlier. If account B typically moves a few million dollars in or out and one day moves 20 million dollars, that's a clear outlier.

What I'd like to do is present a measure that should highlight outliers - I was thinking number of standard deviations versus a population of the rolling last 60 days, but I'm wondering if that's correct. I'm checking to see if it's a gaussian distribution, but are there better ways to hit what I'm looking for?

I think this poses a different set of questions than Robust outlier detection in financial timeseries.

Community
  • 1
MattK
  • 121
  • 2
  • yes, you should use rolling *median* and rolling *mad* (instead of rolling mean and rolling sd) for the reasons outlined [here](http://stats.stackexchange.com/questions/15426/whether-to-delete-cases-that-are-flagged-as-outliers-by-statistical-software-whe/50780#50780) – user603 Nov 08 '13 at 15:48
  • 1
    Could you please explain how your question does differ from the "Robust outlier detection" question you reference? Based on what you have written, they appear identical. – whuber Nov 08 '13 at 16:06
  • The amount transferred in or out is independent of previous transfers, but dependent on total balance. Tick data is largely dependent on previous ticks. There do not need to be large gaps, which is part of the problem of tick data, this is a daily total. – MattK Nov 11 '13 at 16:57

1 Answers1

2

What you have to do is to develop a reasonable model that may incorporate parameters reflecting day-of-the-week , changes in day-of-the-week parameters, week-of-the-year, month-of-the-year, week-of-the-month, day-of-the-month and activity around known events like holidays. The model should detect and incorporate level shifts and local time trends while being robust to pulses i.e. unaffected. The model should detect both parameter changes and changes in the error variance and incorporate remedies. We have been doing this for banking clients (atm machines and elsewhere) since 2002 using AUTOBOX (http://www.autobox.com) a piece of software that I have helped develop. If you wish to post your data ( or a coded version of your data ) please do so and I will submit to AUTOBOX in order to analyze it and then I will post the results. If you don't wish to post your data then contact me at my email address. At a minimum you might want to look at http://www.autobox.com/cms/index.php/afs-university/intro-to-forecasting/doc_download/53-capabilities-presentation as Slide 44-55 speaks directly to your problem.

IrishStat
  • 27,906
  • 5
  • 29
  • 55