FLOA - Delivering Automated, Real-time Credit Scoring at Scale

FLOA
FLOA Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered, Frontrunner 2022 Participant Posts: 1 ✭✭✭✭

Names:

Virginie Lange, Chief Data Science Officer
Laurent Hamentien, Coordinateur Projets Data Science
Aline Mathiot, Responsable IT de Projets Data Science

Country: France

Organization: FLOA

As the French leader in split payment solutions,FLOA makes life easier for consumers through split payments, mini-loans and bank cards.FLOA is a partner of major (e-)retailers (Cdiscount, Oscaro, SFR, Vide dressing, etc.), key players in the travel industry (Selectour, Misterfly, Cdiscount Voyages, Pierre et Vacances, etc.), and fintechs (Lydia, Bankin, Joe) for whom it develops tailormade services.

FLOA’s products and services are distinguished by their ease of use for customers and their rapid integration for partners. FLOA has more than 3.5 million customers and finances more than €2.5 billion worth of goods and services each year. FLOA has been voted Customer Service Provider of the Year in 2022. Already the leader in France, and present in Spain, Belgium, Italy, and Portugal, FLOA's ambition is to become the European leader in split payment solutions. Since February 2022, FLOA has joined the BNP Paribas group to accelerate its growth and become the top European provider of split payment solutions.

Awards Categories:

  • Value at Scale
  • Most Impactful Transformation Story

Business Challenge:

The FLOA data science team is made of about fifteen experts in data science and data engineering and plays a central role in the business. With the advent of new data science tools using artificial intelligence, plus the multiplication of the data used, the added value of the work of the data science team work is real and concrete. After nearly 4 years of working with Dataiku, we’re now ready to move on to the next phase of maturation: acceleration and scaling, i.e. improving the time-to-market of our products.

The data science team is, among other things, in charge of modeling credit score. Our daily challenge is therefore to fit the model to each customer journey.

For personal loans marketed directly by FLOA, we must take into account the commercial agreements made with the comparers and UX: the ingestion of data in the real-time scoring is never the same and the "acceptance/risk" balance must be adapted to the situation.

Regarding split payments, we are in a B-to-B-to-C situation. A split payment means paying for a purchase in installments directly with your bank card without having to sign a credit agreement. There is a lot of talk these days about BNPL: Buy Now, Pay Later.

With more than 300 partners (mainly e-merchants), we face a daily agility challenge: to guarantee a level of credit acceptance and a level of risk in line with expectations, while adapting to the market context! Our goal is to be as close to the business as possible and minimize time-to-market.

The algorithms depend on the partnerships signed: each one sends different data that contribute to the scoring model. From tire sizes for car accessory sellers to airports for tourism experts: we must be able to ingest all types of data with extreme customization (format, type of data, structured or unstructured, etc.)

Each business sector carries its own risk, so we need to create models adapted to each sector and to make them live according to their own commercial seasonality.

The business challenges are numerous, from being able to support merchants with developing their business thanks to split payment to wining “lead auctions” with the comparers for direct credits.

We needed a solution that allowed us to customise our models and industrialise all the necessary adaptations/modifications in order to remain the leading provider for split payments services: this solution needed to run our score models according to numerous parameters and be able to change these parameters in a few clicks with no production delay.

Business Solution:

The main goal was to outsource real-time score configuration from the scores themselves, while leaving the API Services unchanged to avoid any regression. The same applies to the validation and deployment process, which actually means changing the configuration securely in a matter of minutes.

A multidisciplinary “score” team made of data scientists, data engineers and Python developers developed a solution named POM to achieve this. It focuses on adding value using an agile framework to iterate and thus quickly deliver releases that meets changing business needs.

This team received constant feedback from the business team. The goal was to build a viable customized tool and to get business teams to adopt this co-constructed tool.

Today, the POM solution provides a generic and simple API used by all scores. It instantaneously selects the appropriate score configuration using simple to complex criteria depending on the product area. This configuration embeds both technical items (URL, note mapping, etc.) and business items (such as acceptation thresholds). The POM solution then checks input quality before mapping, converting, and sending them to the score itself.

The scores consist of API Services. They are entirely designed, validated, deployed and monitored in Dataiku. If the POM solution leans on FLOA IT expertise to develop performant and robust software, scores benefit from Dataiku’s mature backbone during the whole lifecycle. These are designed and validated in a user-friendly environment. They are later deployed and monitored using the best industry MLOps practices.

One of Dataiku’s key advantages is its openness. We were able to easily interface POM. We enriched score inputs using lookup and SQL query endpoints. And we were able to trust the Kubernetes production infrastructure to host our API Services.

Another great advantage is that Dataiku makes the Dataiku Data Science Studio evolve and provide new functionalities in all domains: algorithms, data preparation, model drift, etc. while continuously evangelizing and consistently impressing its partners.

FLOADiapositive1.jpgFLOADiapositive2.jpgFLOADiapositive3.jpg

Impact:

FLOA uses 35 real time services with more than 2,000 different configurations. During the first half of 2022, these scores were called upon more than 15 million times to provide either risk or fraud decisions.

To illustrate how tightly-coupled business and scores are, more than 1,500 configuration changes were requested during this first half-year. These changes were actually set in a matter of minutes using the POM solution to drive scores. Data scientists thus saved around 200 boring days preparing error-prone releases. This allows FLOA to increase the speed of its business transactions, but also helps to improve the consumer experience through a faster overall response time, which translates into increased business volume.

To demonstrate the whole pipeline performance and reliability, the successful mix of FLOA custom software and Dataiku scores holds response time below 400 milliseconds, error rate below 0.02 % and availability over 99.99 %.

There is another point worth mentioning: by highlighting score configurations, and granting access to this information to people outside of the data science team, the POM solution has also become a way of getting used to scoring/data science for the whole company.

Nearly fifteen people from the Commercial and Risk Management teams are now collaborating on the development of the tool, combining business and technology through a common understanding. Among the next steps, we plan to enhance the tool with a simulation option in the event of changes in settings.

KPI's2022.png

Value Brought by Dataiku:

Agility and real-time being our two main pillars in the execution of our models, the use of AI tools is an obvious choice. It is essential for the FLOA's data science team to be able to create and customize scoring models very quickly when a business need arises.

The use of Dataiku allows us to work with any data format (structured or not), to manipulate the data with internal recipes or to code them in any language (SQL, Java, Python, R, etc.) and to challenge several models in parallel in a short time. We are still able to adapt the hyper-parameters and finally to put in production complete bundles integrating all of this processing. Different technologies are compatible (SNF connection, Blob, Kubernetes, etc.) while remaining in the same environment.

Using Dataiku and the POM tool, we are able to free up time to capitalize on substantive subjects and to:

  • Deliver scoring models twice as fast, which means gains for each model made earlier in the year,
  • Work on dedicated models for large partners, which improves the acceptance rate and/or lowers risk, that is higher average gain per score,
  • Do research and development to work on innovative subjects such as:
    • Data enrichment on B-to-B with open data, open banking, industrialization of drift analysis and machine relearning, etc.
    • DBgraphs, NLP, synthetic data generation, etc.

In the Dataiku project, newcomers were able to understand the projects, be trained and deliver results in less than a month. The security of processing is possible thanks to the control of risks linked to AI with a single platform which avoids the multiplication of controls, audits are facilitated by the traceability of processing and versions and there is no break in the processing chain.

Setup Info
    Tags
      Help me…