Davivienda - Recommendation Machines - An Analytical Window to Financial Inclusion
Customer Analytic Model: Juan Esteban de la Calle Lizeth Roldán Jimenez Diego Alexander Maca García Juan David Jaramillo Ricardo Orjuela Garavito
Davivienda is a regional bank leader that operates in six countries (Colombia, Panama, Costa Rica, El Salvador, Honduras, and the United States). With over 49 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 20,9 million customers through a network of 665 branches and more than 2.692 ATMs.
Data Science for Good
Value at Scale
Excellence in Research
DaviPlata was launched in February 2011 by Davivienda as a cash management service in Colombia. DaviPlata is the digital solution that democratized access to financial services by leveraging the potential of mobile phones.
DaviPlata was the first platform that enabled anyone in Colombia to:
Transfer money via social media.
Make free interbank transfers.
Have a prepaid e-card instantly.
Receive subsidies through a mobile solution.
Manage one’s money over a cellphone without any data charges,
Withdraw or deposit cash on ATMs without a card.
Use its full interface even when the user has a disability.
Most users of DaviPlata are from the low-income segment of the population, which means they don’t count on financial services. Given that they don’t count with them, they lack traceability for products that the common Davivienda user has.
Typically, with bank customers that are not exclusive DaviPlata users, from the products’ story, balance, and payment behavior, we can get variables to create classification models which will enable product offers. With DaviPlata users, we lack this information, so they have a disadvantage compared to the bank user that is not a DaviPlata user.
We had to find a way of solving this problem using a model that could replace a classification-with-many-variables schema so we could enable offers to this low-income population without being at disadvantage against other customers.
With Dataiku, we ran a model called Field-Aware Factorization Machine Recommender System.
Dataiku made it easy to combine the information from different sources into one table containing clients' incomes, credit risk score, and their actual balance of products in other banks.
This way, we obtained a dataset with the customer id, loan balance of any available kind of product the client has in any other bank, and the risk score information and calculated income.
The data preparation was done using pyspark recipes, which allowed the process to be fast and easy to code. As a result, it left a table ready to be modeled.
The next step was to use python’s library xlearn to model the problem. The input table had information on credit risk score, calculated income, and actual balance. The blank spaces for this recommendation system were products the customer did not have. The method would then fulfill these spaces when done, delivering us information on which product we could offer to the customer.
When put in perspective, Dataiku helped us divide the process of preparation, curation, deployment, and model among different people, allowing us to finish the project in a time as short as three weeks.
The Field Aware Factorization Machine is a Recommendation System that allows exogenous variables to participate in the estimation of the missing cells. This is a state-of-the-art modeling tool.
Taking only into account the offers made available with classification models, there were offers available for only 7.9 million of our customers — those who already had products in the bank.
For those customers who only have the cash management service DaviPlata, no offer was available. Still, the output of our model allowed us to calculate the next best offer given the customer characteristics and credit history in other banks — an offer they could not have in any other way.
12.1 million customers now have the possibility of being offered a new product thanks to this model, which uses all available information to generate a product offer for this low-income segment of customers.
The main characteristic of the DaviPlata client is that they are part of the base of the social pyramid in Colombia, forming part of the low-income segment. This population has access to financial services and offers for the first time. Furthermore, the offers made are realistic, according to their actual ability to pay with the help of this tool.
Part of the popularity and traction gained by DaviPlata is explained by the actions of the Colombian government. From 2020 to 2022, they used the platform to send subsidies to the vulnerable population during and after the pandemic to help the most vulnerable with the difficulties derived from it. As the possibility for these people to access financial services was low, using this tool made financial offers available.
Value Brought by Dataiku:
Dataiku added velocity and traceability to the project. With Dataiku, we could divide our work into modules.
One team worked on the code enabling the python package xlearn. The other studied and implemented the folders in Dataiku needed to work with the package. The final result could be easily curated with a team from the Data Office, with the help of Dataiku’s comments.
Today, the project runs automatically on the 5th of each month, helping us deliver value to the client less than 24 hours after the last data was made available. It runs for 12.1 million customers overall.
The Field-Aware Factorization Machine is a Recommendation System that allows exogenous variables to participate in the estimation of the missing cells in a matrix, just as streaming platforms do to recommend the next movie series to watch, only better! This is a state-of-the-art modeling tool that was enabled by the collusion between programming languages enabled by Dataiku.