This website uses cookies for analytics purposes. Learn more
Machine Learning Fraud Model
- Client: Unsecured Lender
- Project Length: 2 Months
- Personnel: Consultant
Background
Our client had been going through a period of increasing commercial success; consistently breaking their own application records month-on-month
However, this success came with operational challenges, as the growing volume of applications put underwriters under significant additional workload
The fast-growing lender was also keen to leverage the most insight from the third-party services they were receiving; moving away from ‘factory settings’ to a more bespoke approach
To help the lender with both of these issues, Vestigo were engaged as part of a longer-term partnership to build a machine learning (ML) model to predict suspected fraud
The project encompassed the entire model build process, from data sourcing and analysis through to model implementation and strategy
Data Preparation
Vestigo had access to the client’s full database, allowing the team to access a wealth of data sources, from application form information to credit bureau feeds, providing an extensive foundation to build from
Data was accessed and extracted using a variety of techniques to ensure maximum coverage and utilisation. Specialist 3rd party fraud data sources naturally proved to be strong predictors
Data cleansing is often the lengthiest element of any model build, however taking the time to ensure data is accurate and appropriate is vital to avoid poorly cleansed data eroding any additional benefit of machine learning algorithms
Modelling Approach
Vestigo’s ML modelling capabilities stretch from the simplest to the most complex available methods. This allows a variety of modelling approaches to be considered during the project
A trade off always exists between model complexity and potential predictability:
- Basic models can be implemented easily and are conceptually simple; however, can be low in predictive power
- Complex models are harder to implement and more difficult to understand. They may not necessarily be more predictive than simpler models
Random forest and XGBoost (boosted random forest) models offer a good compromise of complexity and ease of implementation
Optimisation & Strategy Implementation
Unlike traditional credit model builds, which have more generally accepted optimisation methods, Vestigo used a bespoke optimisation technique to design an implementation strategy for the newly built model
This allowed the team to determine cut offs that ensure the operational load on the underwriters is reduced, and value per application is maximised
The robust fraud model and the associated strategy allows the ambitious lender to leverage the full value of multiple new data sources and the latest modelling techniques. As well as demonstrating the business’ continued innovation in their acquisition processes, the new model and strategy will ensure the lender can continue to grow rapidly without placing their underwriting and fraud teams under undue stress
See how we can help
Contact the
Vestigo team
Experienced risk and analytics consultancy
Building analytical solutions that are focused on your real-world objectives
Get the latest news from Vestigo