Predictive Analysis In Fraud Risk
by Deepali Dharwadker
After spending over a couple of decades in Data and Analytics, I am often asked “What kind of Analytics are most suitable for my organization?” There are many kinds of Analytics. In this article I will focus on Predictive Analytics in the Fraud Risk domain which constitutes the major play in the Analytics world in this segment.
Industry survey states that over 70% of BFSI executives believe that Big Data can play a key role in Fraud Prevention and Detection provided they embrace the technology for Statistical and Algorithmic Techniques along with continuous transaction level data monitoring.
“BI delivers Insight, Predictive Analytics delivers Action”
These days there is a shift in the concerns for the Chief risk Officer (CRO) – in addition to concerns about regulatory risks, business continuity they now seek solutions for RADAR (Risk Assessment Data Aggregation& Reporting) i.e. –
- Having an integrated picture of risk across the enterprise
- Extending risk coverage and information dissemination to the larger business community to help strategize
- Predicting risk
Assuming we are all familiar with RADAR, lets focus on the most popularly used Predictive Analytic Techniques for Fraud Risk :-
Predictive analytics thrive on models – which is a mathematical formula or an equation that takes in data and produces a calculation, such as a score. It applies itself to data as a set of instructions to deliver a particular kind of result.
The result is a score which is a numerical value generated by the model when applied to a data set. However not all models generate scores
Predictive models are often embedded in operational processes and activated during live transactions. They analyze historical and transactional data to isolate patterns e.g. :-
- What a fraudulent transaction line entry looks like ?
- Look for identical or repetitive patterns in the transaction details e.g. location, threshold amount, business unit, dates like month end, weekend etc
- What a risky customer looks like ?
These analysis draw the relationship between hundreds of data elements to isolate each customer’s risk or potential, which guides the action on that customer. This way, the customers may be tagged GREEN (good customer), RED (bad customer potential fraud) with varying scores.
Below are some popularly used Predictive Analytics in Fraud Risk
- Neural Networks :-
In environments with heavy data traffic, huge transaction volumes and abnormal data patterns, Neural Networks provide some help. However they work the best with pre-transformed smooth data and hence potentially viable for use in an RADAR ecosystem
The hidden layer is the mathematical core of a neural net. It selects the combinations of inputs (e.g., dollar amount, transaction type) that are most predictive of the output—e.g When your credit card is used or a claim is processed for payment.
- Clustering :-
Clustering models use demographic data and other customer information in order to find groups or “clusters” of customers with similar behavior, background or interests e.g. one step might be to vary the monitoring threshold of transactions for different cluster of customers. Using this cluster you could build a subset of varying thresholds to monitor his actions closely and accordingly produce Amber Data for caution As a result, clustering can be used as a precursor to predictive modeling
- Risk Maps :-
Risk Maps are very popular with the firms. They provide a list of potential risk, the ‘probability of the risk occurring’ and the ’impact size of the risk’. This way high impact risk can be closely monitored, however though this model does not provide the co-relation of 1 risk to another and hence every risk is an isolated find. For instance if we take 2 risk ‘fraudulent values of an asset’ and ‘lender of the asset’ as separate risk items, then the identification of risk is isolated whereas ideally both these risk need to be monitored in conjunction for Fraud Risk.The best way to use fraud related risk maps is through collaborative effort of different business units across varying functions within an organization to provide the business linkages to potential fraud risk.
- Simulations :-
Simulation techniques are typically used in getting information about how something will behave without actually testing it in real life. It works on following principle :-
Input=uncertain numbers/values * Intermediate Calculations = Output(uncertain numbers/values)
So how do these uncertain values help Fraud Risk ? The concept here is to do a ‘What-if’ Simulation of data. Each model is executed thousands times with varying data input to determine the probability and respective outputs e.g. Monte Carlo Simulation. Unlike the Risk Map the Monte Carlo Simulation factors in correlations between the variables.
While predictive analytics can be very useful and provides an objective view of data related to risk along with mitigation ideas, it is fundamentally important to tie the model appropriately to the business case, business unit and the relevant metrices along with behavioral and economic data.
The views expressed herein are the views of the author and not necessarily the views of the employer.