How to build a demo for Cross Sell in Banking using IBM Cloud Pak for Data

Published in

IBM Data Science in Practice

4 min readJul 16, 2021

a person screaming into a telephone receiver. the telephone is lying on a surface in front of the person as are a pair of glasses — Photo by Icons8 Team on Unsplash

Imagine you are in charge of customer accounts at a large bank. You deal on a daily basis with customer complaints on how their plan doesn’t match their needs. Now imagine there was a solution to help you and your sales staff provide bespoke service to each customer. My colleagues Fawaz Siddiqi, Khalil Faraj and I developed a prototype of a solution for targeting the right products to the right customers.

This data set that we have used contains many demographic fields such as relationship status, age, number of kids, among others, as well financial information like credit score, income, and assets. Using these multiple fields, we built a prediction model with SPSS Modeler and visualized the insights to show which groups of customers are more likely to buy or show an interest in a specific bank product. To get started, make sure to create your IBM Cloud account to access our services to build the solution.

Sign up for IBM Cloud

Edit description

ibm.biz

In our use case example, the bank products and plans are:

Increase net worth
Philanthropy
Capital Acquisition
Education Planning
Retirement Planing
Estate Planning

Our goal is to target the right plans to the right customers.

mridulrb/Cross-Sell-in-Banking

This is a CP4D demo asset that focuses on Up/Cross sells in the Banking industry focusing on targeting customer for the…

github.com

Technologies Used

Watson Studio (CP4DaaS)
SPSS Modeler: to build our prediction Model
Cognos Dashboard: To visualize our prediction output and show the insights

Architecture Flow

an Architecture Diagram showing the data in the external portion of the system. The data then goes into Cloud Pak for Data’s Watson Studio, starting with SPSS modeler. The next step in the pipeline is that SPSS modeler outputs new data that is then input into the Cognos embedded dashboard. — Architecture Diagram

The dataset is uploaded to Watson Studio to use in SPSS Modeler for the prediction model (Here the data is already cleaned with Data Refinery).
SPSS Modeler targets the field Pursuit that contains the values of the bank products/plans.
SPSS Modeler generates a new file with the prediction results that is used in Cognos Dashboard to visualize the insights.

SPSS Modeler

Flow

In this flow we are using our cleaned data set and it’s partitioned so that 20% is used for testing and 80% is used for building the model.
Our target field, as discussed above, is Pursuit that contains the values of the bank products/plans.
This flow generates an output file that has the prediction result

Prediction Model

Our prediction model consists of a combination of five different algorithms:

Random Trees
XGBoost Tree
XGBoost Linear
Logistic Regression
CHAID

a table of the results from the models: random trees shows an accuracy of 98.010 percent, a build time of 1 minute, and used 36 fields. XGBoost Tree1 had an accuracy of 97.512%, a build time of 1 minute, and used 36 fields. XGBoost Linear 1 had an accuracy of 97.015%, took 1 minute to build, and used 36 fields. Logistic regression 1 had an accuracy of 96.517%, took 1 minute to build, and used 36 fields. CHAID had an accuracy of 96.020%, took 1 minute to build, and used 9 fields. — Prediction Model Results

Our most accurate model was Random Trees with an accuracy of 98.010. All of the estimators had a relatively short build time. However, CHAID notably stood out, with delivering an impressive accuracy of 96.020 with the use of only nine fields of data, unlike all of the other algorithms which required 36 fields.

Data Visualization — Cognos Dashboard

In the above dashboard, the graph on the left shows the predicted pursuit given the monthly housing cost. The graph on the right shows the market share of predicted pursuits.

In this dashboard view, the graph on the left shows shows the predicted Pursuit values with respect to the customers’ annual income and the customers’ acquisition cost. The acquisition cost is the cost to the bank of winning a customer to purchase a product. The graph on the right gives us information on what type of products these consumers are pursuing based on their profession and annual income of the household.

Creating the dashboard is subjective because when we are visualizing data we are telling stories from these insights thus each person will have his own perspective to tell a story and share the insights. We have attached a canvas in the GitHub Repository which will allow you to import the dashboard into Cognos Dashboard Embedded so that you can examine it for yourself.

Conclusion

In this prototype, we built a predictive model for a Cross Sales in Banking use case with SPSS Modeler Flow. The model helps in identifying the right product for each customer. We then visualized the data to get some insights and for better understanding of the results provided by the predictive model. Click Here to see a Demo of the results. We have an upcoming masterclass at the Artelligence Forum where we will be demonstrating this prototype. More details about the event will be shared once finalized.