Davivienda - Building an Analytical Solution to Obtain Credit Risk Scores for Personal Banking Custo
Name:
Andres Hoyos
Paula Andrea Rivera
Country: Colombia
Organization: Davivienda
Davivienda is a regional bank leader that operates in six countries (Colombia, Panama, Costa Rica, El Salvador, Honduras, and the United States). With over 50 years of experience in the Colombian market, Davivienda offers a wide range of financial services to individuals, SMEs, and corporate customers. The bank currently serves over 20 million customers through a network of more than 650 branches and 2.500 ATMs.
Awards Categories:
Best Acceleration Use Case
Best Moonshot Use Case
Business Challenge:
Within the world of Davivienda bank credit products, originations are carried out in two ways: by request or credit campaigns.
When it is by request, the client requests to open a new credit product, and the bank decides whether to approve, modify or deny the credit. When it is by campaigns, the bank does internal studies where it seeks to identify those clients with good financial behavior and offer credit products that are part of the bank's portfolio.
Until December 2021, the selection of clients for credit campaigns was carried out through hard risk policies and filters designed by credit risk managers. These policies and filters were built around characteristics of the bank's clients, such as client income, age in the financial system, and an external credit score. This method for the selection of clients presented a stagnation in the number of people to whom the credit products were offered, so the need arose to optimize the campaign funnel, that is, to increase the number of people. This is through an analytical solution.
This analytical solution faces three paradigms on the way:
- In the first instance, Hard Filters and Policies, where the expert criteria of the managers that had three variables are replaced by a robust data science methodology with more than 30 features.
- In the second instance, Client Context, where the characteristics of the client's month are replaced by variables with historical behavior.
- In the third instance, Artificial Intelligence, which refers to the fact that there was some skepticism regarding using automatic learning models since they are considered a black box. It gives him the decision to approve or deny generated uncertainty, but as we can see, its interpretability is friendly.
Business Solution:
To build the analytical solution, a prediction model was trained to obtain a credit risk score for personal banking customers. To train this model, in the first place, the connection to Cloudera, where the databases are located, was made.
Subsequently, a pipeline was built to join information and with PySpark code to build the variables and the entire flow. The model was trained and left productive where the scenario is executed monthly.
Day-to-day Change:
The implementation of this model changed the day-to-day of the organization from two points of view.
In the first place, since it is a process that performs for more than 15 million records with more than 30 features, executing it locally on the PC was complex. Now that it runs on Dataiku, the process and the user interface are much more friendly.
Secondly, credit risk managers frequently request exercises and tests. Since the pipeline was configured in a flexible manner, results are given in less than a day, which positively impacts decision-making.
Business Area: Risk/Compliance/Legal/Internal Audit
Use Case Stage: In Production
Value Generated:
The deployment of this model in Dataiku is generating a monthly increase in the campaign funnel of 1.4 million people, from 2.3 million to 3.7 million people. The deployment also improves the behavior of new credit disbursements, going from a default probability of going from 4.9% to 3.8%. That is, the default in the payment of credits decreases by 1.1%
Value Brought by Dataiku:
The specific value provided by Dataiku is that it allows us to perform most of the steps of the CRISP-DM methodology in a single interface. It allows users to write code in different languages, put different models to compete in a single step, and create an easily traceable logical flow.