Business Data Miners proprietary algorithm is able to detect cards that will likely have fraudulent activities within 6 months in the future. If these cards are reissued or closely monitored, we can reduce a significant amount of the fraud loss. In a live testing of our algorithm at a top 15 bank, 19% of cards that we predicted as being compromised had fraudulent activities within 3 months after our prediction. We estimate that the annual fraud savings is above $1 million. By reissuing risky cards, we prevent fraudulent activity before it happens. An added benefit of doing this is that the impact of fraud on good customers is eliminated.
Archive for the ‘Uncategorized’ Category
As of 6/3, our bad debt model has saved tens of millions dollars over the past three years by reducing the delinquency rate of a large bank by 50%!
When there are no historically known fraud data, we have to build unsupervised models that do not require a target variable. There are a number of such models that we can try: one-class support vector machine, principal component based compression net, etc. They are detecting data points that are abnormal, not fraudulent per se. We only hope that being abnormal is highly related to being fraudulent.
Feedback from the client, one of the top 15 banks in US, was great. The volume of existing ASI19 check fraud alerts is reduced while more frauds are detected.
BDM’s algorithm is able to produce a list of likely compromised cards, 19% of which will have fraudulent activities within 3 months in the future. If we reissue them today, we can reduce significant amount of fraud loss with least amount of impact to good customers.
Delivered the optimal debit card signature transaction daily limit for a client bank. Performing the similar analysis to optimize PIN-based transaction daily limit. The core idea is to maximize the net of transaction fee plus overdraft fee minus fraud loss. The main technical challenge is to merge 100 million records in settlement, authorization, overdraft and claim fuzzily and iteratively.
Most of the debit/credit card fraud detection is based on transaction level rules or scores that target the fraud transactions happening in real time. Fraud prevention operation is struggling to stop fraud in real/near real time.
If we can predict which cards will likely have fraudulent activities within the next a few months, we can either reissue those cards today or build transaction level rules to monitor them. The long term card fraud prediction is, just like long term weather forecast, doable. At BDM, we have tested such a method and found the results are good and actionable.
There are two types of rules, “superficial” rules and “inherent” rules. Superficial rules are easier to create and they are straightforward to understand, e.g., transactions happening at certain retailers. These rules may be good for a few weeks and their performance drop quickly. Inherent rules capture more time lasting ( and usually hidden and mutli-dimensional) patterns of fraud behaviors. To create them,we may have to use more sophisticated methods including predictive modeling. The PRM SQL statements for some of these inherent rules that we created have 5,000 characters. These rules work like magic and have very low false positives.
About 25% of compromised good cards identified by BDM will have fraudulent activities within the next 3 months. (In comparison, usually about 1% of cards alerted by a popular point of compromise (POC) service from a big company have fraudulent within 12 months.) The common points of compromise detected by BDM are highly accurate. Identifying POC accurately and early and reissuing cards provide the following benefits: 1. It is better than real time fraud detection because the fraud loss can be stopped before it happens; 2. Reissuing cards is less intrusive to good customers than blocking suspicious transactions.
Based on the actual debit card fraud detected by BDM rules since June of 2010, every alert generated by BDM rules stops $32 fraud. The following are some facts about the BDM rules. (BDM stands for Business Data Miners.)
1. $32 is the total dollar amount of debit card fraudulent transaction actually stopped by blocking cards divided by the total number of alerts generated by BDM rules.
2. $32 does not include fraud transactions detected by existing non-BDM rules. For example, if both existing non-BDM and BDM rules alert on the same card and an analyst blocks it, the fraud loss prevented is excluded from $32.
3. The total additional fraudulent transaction amount prevented due to BDM rules is multiple million dollars annually.
4. There is no increase in operation cost at all. This is because: a. the volume of alerts generated by BDM rules is less than 10% of total number of alerts; b. we helped the client identify ineffective rules and retire them.
5. BDM rules are not only statistically accurate but also understandable. For example, we have BDM rules targeting on risky retailers, risky MCC and risky cities.
6. BDM rules are written in standard SQL statements that can be deployed in many fraud prevention products.
7. BDM rules are refreshed monthly based on latest transactions and fraud information.
In fraud prevention analysts’ own words, “these BDM rules are amazing”.