With the rise of digitalization in the Financial world, Analytics, be it in the form of Marketing Analytics, Business Analytics or Risk Analytics, has become an integral part of the Financial institutions across the globe. The systematic approach towards analyzing internal & external data and establishing a decision making process around the same is becoming popular in Bangladesh as well. However, in most cases, analytics activities are limited to segmentation or are descriptive in nature, whereas much more can be done with the help of easily obtainable tools & skills. ‘Predictive Analytics’ is such a branch of the advanced statistics through which someone can predict unknown future events or activities with certain levels of precision and utilize the same for achieving a certain level of operational efficiency and for identifying future Risks and Opportunities.
Predictive analytics mainly answers the question ‘what is likely to happen?’ It involves analyzing historical data, discovering the trends and the drivers of specific outcomes and finally establishing the overall relationship between the historical data and the outcome in a rule or formula format which is often referred to as ‘predictive model’ or simply ‘model’. This model is treated as the core engine of the whole Predictive Analytics approach. Once the model is developed and validated, this can be applied to recent datasets to get the likelihood of the said outcome to happen in future. Often, the rule or formula is transferred into a Scorecard format for the ease of application or implementation.
Any business function, which possesses a substantial amount of historical data on customer demographics, behaviors, external bureau or any other sources, can build a predictive model and setup the strategies around it. The frequently used Predictive Models in the Financial world are as follows.
Application Scoring: This is widely known as ‘A-Score’ and is a very popular propensity model in Consumer Credit Initiation area. Customers’ demographic information and Bureau information are used in developing the model to predict the probability of the customers’ future Default / Non Payment. Once the model is prepared and its predictive power & robustness are tested to meet the standard, all new credit application goes through this model or scorecard and the model provided score determines the ‘Approve/Reject’ decision. The cut-off score determination for Approve/Reject decision depends on the organization’s Risk Appetite. Implementation of this scoring mechanism makes the credit approval process very fast and free from any human biasness. In an advanced environment, where the bureau is truly online, the credit decision making time can be reduced to few hours, if not minutes and the whole process can run without any credit approver/ underwriter in place.
Behavior Scoring: ‘B-Score’ is another popular propensity model, which is widely used for ‘Portfolio Management’ activities from both Risk & Business side. Here also, Default/Non-Payment is predicted based on the customers behavioral data like spend, payment, limit increase request, recent bureau score etc. Since, more of behavioral data is used in development, usually the predictive power of B-Scores is significantly higher than A-score. B-scores are refreshed on a periodic basis and almost all portfolio management decisions are backed by the B-score based strategy. B-Score allows the organization to take decisions on individual customer for up-sell / down-sell, over limit tolerance of Credit Card, facility renewal etc. based on customers ‘Probability of Default’ and hence ensures good portfolio growth.
Collection Scoring: ‘C-Score’ is used for optimizing collection / recovery efforts if the organization has a large delinquent customer base. Customer’s behavioral information including collection related information i.e., Call Received, Promise to pay, Broken / Kept Promise etc. are used to develop the model which eventually predicts customer’s probability of payment / recovery. Therefore, the organization can reallocate its collection staffs toward the customers where the probability of recovering the payments is high. In absence of a proper C-score model, some organizations use the available B-Score to prioritize collection calling.
Fraud Model: Fraud models can identify potential fraud transactions in real-time and allows the organization to block the transaction, initiate the investigation or any other suitable actions. If the organization has sufficient history of Fraud transaction, such a model can be built and implemented. This type of models can also screen for high-risk online transactions and trigger appropriate actions. Fraud models are gradually becoming popular with the increased number of e-commerce transactions.
Marketing Model: This kind of model predicts customers willingness to take a particular product or service when offered as a part of Cross-Selling campaigns or a fresh acquisition campaign. For fresh acquisition models, organizations mainly depend on secondary data and hence the models are usually less powerful. But, Cross-sell models are developed based on internal customer data and hence can predict customer behavior more precisely. Using this model, organizations can achieve higher success rates even with a smaller base of target customers. Organizations use this type of model for repetitive campaigns and achieve significant uplift on sales by involving less time and resources. Some Financial organization use similar models to predict individual customer’s willingness to renew a Deposit / Loan product and re-price the facility based on the model outcome.
Anti-attrition Model: Customer retention is a big challenge in competitive markets. Retention Models or sometimes called “Anti-attrition Models” can help predicting a customer’s future attrition so that the organization can take initiatives to retain him/her. Usually, for the group of customers, who have higher attrition propensity, the organization identifies some counter offer(s) based on the individual customer’s historical & future profitability and keep the CRM system updated with the offers. Once the customer communicates the bank with an account / facility closure request, the front line staff or the call center agent can immediately pick the best offer for the customer to retain him/her. The probable offer can be some additional Reward points or waiver of some fees or charges or some cash back etc. based on customers’ individual profitability profile.
The biggest criticism against ‘Predictive Modeling’ is that it takes only historical data in to consideration and by the time the model is prepared and ready for use, the data is few months old. The critics term this as ‘driving your car by looking at the rear-view mirror only’. However, one can somewhat overcome this by establishing a strong system to periodically monitor the model performance in most recent available data points and re-defining the cutoffs or strategies as soon as there is an indication that the model’s prediction power is below standard, which can be the result of a quick change in business environment or gradual change in customer behavior.
Author: M Moniruzzaman
Moniruzzaman is the Head of Group Retail Business Intelligence
of the largest bank in the Middle East.