# Find abnormal credit transactions based on historical data

I have a dataset of customer transactions (multiple customers,multiple transactions) and based on the historical data, I want to know when a new credit(+ve) transaction arrives if its unusual for that particular customer .

The transaction data comes with 4 fields

customer_no, transaction_type, date_vale,amount


some transactions_types such as salary,interest (with similar amounts ,not exactly same everytime ) occur in regular interval and other transaction_types dont have any regular behaviour for any particular customer .

for now what I am doing is calculating mean,median,standard deviation and mad for each customer from the historical data and calculating z score and robust z score for each transaction (for each customer_no ) .

Few of the caveats that I want to mention

all customers dont have same no of data and dont have data spanning everyday of the time period (you can think of this like a normal bank statement for a person , you dont have transactions everyday )

The historical data is not labelled ,that is no field pointing which transaction are unusual and which are not

so my goal is to find characteristic of unusual credit transactions from historical data and find if new transactions are unusual or not based on those characteristics

I am looking for suggestions on building this model.

• This is rather open-ended question. I think the first you must do is to do a mental exercise trying to define what is "normal" and what is "abnormal" transaction. E.g. large amount transactions, unusual dates/times can be some of the flags. Jul 27, 2016 at 23:55
• @xeon by abnormal I mean the large amount transactions which happens rarely Jul 28, 2016 at 0:06
• It depends on a user. You can have a user that barely uses his credit card, but you can have another one which makes lots of transactions. This is a good indicator but for sure not the only one and not the 100% sure one. Jul 28, 2016 at 5:50