Introduction

We - the data science team at Lendable - are involved in projects that touch upon all aspects of the business including:

• The creation and deployment of machine learning models that allow us to provide loan pricing based on an applicant's scored risk.
• Development of reporting tools used both internally and by our investors to report on the performance of our loans
• Streamlining our customer’s journeys

Model development

Our underwriting models decide whether an individual is accepted or rejected at the credit checking point in an application (other checks are also undertaken).

Our models are trained on a mixture of both internal and externally sourced data. An example data provider might be Experian who are one of the main credit reference agencies in the UK.

We will probably undertake initial data analysis in Jupyter notebooks to ensure there is an understanding of the type of data being utilised. Multiple model algorithms will be considered based on the types of problem being addressed. For classification problems we have historically had some success with both neural networks and gradient boosted decision trees. In particular the CatBoost library which more readily deals with some categorical features that commonly appear in applicant credit files.

Once trained, validated and meeting our performance expectations we will transition from notebooks to a centrally accessible model build repository allowing anyone within the team to reproduce any models utilised in production.

Model deployment

We typically own model deployment in the form of a web API that the Lendable platform makes requests of. FastAPI has been used in several projects to create a relatively simple web API that is hosted on AWS . Because we own model deployment, we can quickly make changes to both our models and other decisioning frameworks where credit policy requires.

Sharing insights

To manage data views that are regularly requested by the business there is a hosted dashboard containing a mixture of summary tables and charts, as well as the ability for users to run bespoke short reports as needed, for example the performance of a specific investors portfolio based on a chosen metric.

This dashboard uses React, Python and D3 to pull together the information requested by the wider business.

We also have automated scripts hosted on Jenkins that push reports out to investors on a scheduled basis, either via email or straight to SFTP servers.

Streamlining lending

We try to make our lending decisions as rapidly as possible, therefore as many additional customer checks as can be automated to reduce any wait time for customers. We are able to do this by speaking with our customer service agents and identifying current pain points where there is merit in developing a solution that may speed processes up.

A team of generalists rather than specialists

Each of us brings a different set of skills to the data science team. We therefore prefer not to be pigeon holed into working on one specific type of problem and instead, try to ensure that we are each exposed to the full spectrum of work seen in our team.

Operating in this way ensures the team comprises a group of generalists who each have a broad knowledge of Lendable and can pick up and add value across the business whatever the project.