Interest Rate Optimizer
SUMMARY
- Client
- Equa bank
- Tech stack
- Apache Spark, Profinit propensity-to-buy lending solution, Hadoop, Python
Setting interest rates is complicated: too low cuts profit, too high drives clients away—competition is stiff, especially online. Most banks use risk-based pricing, assigning rates by default probability. Data-science algorithms improve this approach. By giving price-sensitive clients small bonuses at the expense of less price-sensitive ones while maintaining average rates per risk class, we improve acceptance, reduce turnover, and boost profit while complying with banking regulations.
Results
Project Background
Equa bank – a fast-growing Czech challenger bank and long-term Profinit partner – asked our team to create a propensity-to-buy scoring model (hereafter “propensity model”) that would help identify clients likely to apply for future consumer loan products. The project involved compiling data from the 2-year transactional history of 400,000 Equa clients as well as analysing the socio-demographic and product information available for this specific group.
Challenge
To create a sufficiently accurate propensity model, we needed to execute a number of highly complex computations based on a huge volume of unstructured data. The task would involve processing tens of millions of transactional records and hundreds of millions of links across the client network.
To that end, we knew that employing relational databases or conventional statistical methods like regression and segmentation just wouldn’t do. The challenge required a sophisticated technical solution.
Business Needs
The solution needed to meet the following business targets:
- Improve the conversion rate of consumer-loan product offerings
- Generate accurate prediction scores for Equa’s entire client base
- Evaluate added business value for different client segments
Solution & Results
Our unique propensity-to-buy solution was based on modelling client behaviour and social similarity networks. With these insights, we were able to identify client microsegments using advanced machine-learning methods developed in close consultation with our research partners at Charles University in Prague.
To handle the huge volume and complexity of input data, we built a big data computational pipeline using specifically designed data structures on Apache Spark and the Hadoop platform. The set of clients evaluated by the Profinit team was independently verified by data analysts at Equa bank. They confirmed the high precision of our prediction model (87% AUC). Not only that, they found that our propensity-to-buy scoring was effective across almost 70% of the client population including new customers, inactive clients and those with no previous loan history.
HEAR FROM THE CLIENT
Profinit’s propensity-to-buy model delivered added business value for almost 70% of our client base, including challenging segments like new and inactive customers.
