Case 1
Purpose
A leading UK retail bank wanted to gauge the credit worthiness of its customer base for efficient risk management
Approach
The said client had a pre-existing scorecard for its overdraft applications but the scorecard was based mostly on heuristic scores given to different factors that gauge default. Moreover the scorecard was able to differentiate good and bad customers with only 55% accuracy
After the problem definition phase we started collating data for the below mentioned factors that would differentiate between credit-worthy customers and defaulters:
- Customer demographics
- Customer credit score
- Investment behavior
- Loans portfolio
- Transactional behavior
- Customer relationship with the bank
We then used variable selection and filtering techniques like Principal Component Analysis and Exploratory Factor Analysis to find the most important variables amongst 4000 variables. But in the end Evolutionary Optimization techniques yielded the best derived variables
The derived variables were used in machine learning algorithms like artificial neural networks and gradient boosting to create the final scorecard model. The final model was presented to the client in an R markdown documentation framework for easy consumption
Outcome
Using proprietary learning techniques we improved the accuracy of the existing scorecard by over 17%
The improved scorecard helped the bank to pro-actively differentiate between credit-worthy customers and defaulters.
The bank was able to adjust the interest rate based on the model credit score for all its overdraft customers. This led to improvement in the asset liability management and the overall bank profits by more than 20%
Case 2
Purpose
A leading retail bank wanted to under the behavior of its customer base for cross selling and up selling its products
Approach
Using our problem understanding framework we divided the core problem into 2 parts:
- Which are the potential customers that could be targeted for a cross sell or up sell strategy
- What type of products to target those potential customers
We initially segmented the customer base based on various factors like
- Number and types of accounts in the bank
- Credit history
- Loans portfolio
- Region and demographics
- Income and spend levels
The customer segments were then profiled based on simple Exploratory Data Analysis and hypothesis testing methods to understand their behavior across various segments
Finally a response model was built using Boosting techniques for every product- customer segment combination to understand what is the response rate of customer across different types of targetting channels
Outcome
Using the model the bank was able to devise a sound multi-channnel targeting strategy for its existing customers.
The initial touch point response rate improved by more than 35% and the final conversion rates improved by 20%
The bank recorded 19% improvement in product sold per customers and households due to their improved targeting strategy