Fraud Detection in Banking – Part 2

 Fraud detection in banking technologies enable merchants and banks to perform highly automated and sophisticated screenings of incoming transactions and flagging suspicious transactions. While none of the tools and technologies presented here can by itself eliminate fraud, each technique provides incremental value in terms of detection ability. As it will be discussed later, the best practice implementations often utilize several of these fraud prevention techniques, if not all of the tools discussed here. The various fraud prevention techniques are discussed below:

Manual Review 

This method consists of reviewing every transaction manually for signs of fraudulent activity and involves an exceedingly high level of human intervention. This can prove to be very costly, as well as time consuming. Moreover, manual review is unable to detect some of the more prevalent patterns of fraud, such as use of a single credit card multiple times on multiple locations (physical or web sites) in a short span.

Recent Developments in Fraud Management

The technology for detecting credit card frauds is advancing at a rapid pace – rules based systems, neural networks, chip cards and biometrics are just some of the popular techniques employed by Issuing and Acquiring banks these days. Apart from technological advances, another trend which has emerged during recent years is that fraud prevention is moving from back-office transaction processing systems to front-office authorization systems to prevent committing of potentially fraudulent transactions. However, this is a challenging trade-off between response time for processing an authorization request and extent of screening that should be carried out.

Simple rule systems

Simple rule systems involve the creation of ‘if...then’ criteria to filter incoming authorizations/transactions. Rule-based systems rely on a set of expert rules designed to identify specific types of high-risk transactions. Rules are created using the knowledge of what characterizes fraudulent transactions. For instance, a rule could look like – If the transaction amount is > $5000 and card acceptance location = Casino and Country = ‘a high-risk country’. Fraud rules enable to automate the screening processes leveraging the knowledge gained over time regarding the characteristics of both fraudulent and legitimate transactions. Typically, the effectiveness of a rule-based system will increase over time, as more rules are added to the system. It should be clear, however, that ultimately the effectiveness of the system depends on the knowledge and expertise of the person designing the rules. The disadvantage of this solution is that it can increase the probability of throwing many valid transactions as exceptions, however, there are ways by which this limitation can be overcome to some extent by prioritizing the rules and fixing limits on the number of filtered transactions.


Thus, transactions can be prioritized based on the risk score and given a limited capacity for manual review, only those with the highest score would be discussed.


Technique that is the most commonly used for fraud detection is Neural Network technique. While this technique is completely omitted and ignored by credit risk modelers and scorecard developers due to its inability to provide transparent understanding of what are the risk factors. This is also often legislative requirement, so that person who is refused credit knows why he is refused credit.

In banking fraud detection, there is no such requirement for model transparency, therefore main focus is on predictions rather than on model's transparency. And Neural Networks are known to be “universal approximates” who are able to model both, linear and non-linear relationships between inputs and the target. Other approaches are based on statistical models designed to recognize fraudulent transactions, based on a number of indicators derived from the transaction characteristics. Typically, these tools generate a numeric score indicating the likelihood of a transaction being fraudulent: the higher the score, the more suspicious the order. Risk scoring systems provide one of the most effective fraud prevention tools available. The primary advantage of risk scoring is the comprehensive evaluation of a transaction being captured by a single number. While individual fraud rules typically evaluate a few simultaneous conditions, a risk-scoring system arrives at the final score by weighting several dozens of fraud indicators, derived from the current transaction attributes as well as cardholder historical activities.

While some of the other models that can be used have strong statistical origins (regressions and general linear model), neural networks are invented by applied mathematicians, computer scientists and machine learning practitioners, whose working principles are motivated by the functions of the brain – especially pattern recognition and associative memory. The neural network recognizes similar patterns, predicting future values or events based upon the associative memory of the patterns it has learned. The advantages of advanced analytical methods such as neural networks are that they are able to learn from the past transactions in terms of which values of which variables are historically associated with normal transaction and which ones with fraudulent one.

So, first step in building fraud detection system is to focus on specific type of fraud, bring not only relevant data but also domain expertise in forms of rules that can be used to construct and derive analytical input variables. Sometimes, rules are used as separate scoring mechanism that scores transactions in sync with more complex advanced analytics model, with the premise that if both, rules and the model score transaction score high in terms of likelihood of being fraudulent – this transaction is more likely to be fraudulent. Some of the examples of domain-expertise based rules:

  • If check deposit is closely followed by cash withdrawal within say 10 hrs.
  • If transaction type is above a specified number in 48 hours.
  • If active more than one session at the same time.
  • If trying to withdraw more money than the limit in credit.
  • If trying to withdraw more money than the amount in debit.
  • If trying to log on for more than 3 times at once.
  • If any transaction is more than 80% credit limit in 48 hours (one transaction or sum or transactions in the 48 hour period).
  • Deposit activity out of the normal range for any account
  • Invalid Routing Transit numbers
  • Excessive numbers of deposited items
  • Total deposit amounts greater than average
  • Large deposited items masked by smaller deposit transactions
  • The amount exceeds the historical average deposit amount by more than a specified percentage
  • A duplicate deposit is detected
  • Deposited checks contain invalid routing or transit numbers
  • The level of risk can be managed based on the age of the account (closed account getting lot of transactions suddenly).
  • The number of deposits exceed the normal activity by the customer

These characteristics may be common knowledge as potential indicators of fraud, however, they may be many more patterns that are not known that can provide incremental improvement in model quality. And so, by incorporating advanced analytics, including champion-challenger capability, based on multiple data sources and with multiple detection requirements across the entire spectrum of fraudulent techniques enriched by existing rule base approaches is the way of pushing the limits in the fight against fraudsters.

No Comments Yet.

Leave a comment