At last year’s Edinburgh Credit Scoring and Credit Control conference, Danilo Blagojevic from Tesco Bank shared some excellent insights from developing machine learning boosted tree models for credit risk.

I liked his simple analogy of an F1 racing car compared to a Ford Focus when discussing machine learning algorithms compared to logistic regression for credit risk modelling. And staying with the car theme, in a recent FICO blog Scott Zoldi explained why driving a Ford Bronco off road is a lot like solving Data Science challenges whilst talking about “Open Source junkies”.

Continuing on the racing car vs. family runabout analogy, there are many factors that need to be considered when making your car choice for the situation and what you intend to use the car for. Much like when assessing your machine learning or traditional logistic regression modelling choices, it is not just about the horsepower and speed you would like. As part of my sponsor spotlight session, also at 2021’s Edinburgh conference, I highlighted the important considerations of operationalisation, robustness, model engineering, ease of development and explainability. All are key factors to be assessed and understood for the intended use of your predictive model.

What about the fuel mix (data sources) that works best, the tyres (decision/scoring engine) that will be needed, the information available in the cockpit/dashboard (reports and explainability), the reliability, servicing and maintenance (model monitoring and maintenance) required, as well as the driver skills required (model developers and users) and the safety (understanding all risks, whether operational, regulatory or ethical). You can see the similarity here with machine learning algorithms for automated decisions. And not forgetting whether or not we need to fully understand what is happening ‘under the bonnet’.

Here at Paragon, we continue to work with our customers and partners to determine if and how machine learning algorithms can be explored and adopted for credit risk modelling. Within our Modeller software we have introduced machine learning algorithms that sit alongside the more traditional logistic regression options which have been the bedrock of our credit risk modelling software for many years. Our view is that machine learning vs. logistic regression should not be an either/or decision. The choice is situation and problem specific, and therefore being able to explore, develop and evaluate across algorithm choices within the one development tool is both attractive and efficient.

If you are interested to learn more about Modeller please get in touch.