Nicolas Gonatas, Beam’s co-founder and Head of Data Science, talks about Beam’s data science philosophy. This post is a follow-up to our series of articles on Beam’s Data Science Philosophy’ and Machine Learning Principles.

What kind of data sources and variables do you consider the most critical for training your machine learning models. How do you ensure data quality and relevance?

The way traditional bureaus source data for their reports and scorecards is through the SACCRA Hub. Every lender in SA has to report its loans and borrower repayment data, by regulation, to the SACCRA Hub. SACCRA generates reports on whether customers are making repayments, and how much of their credit limit they have used. There are however inaccuracies or errors that creep in, called blemishes in the data, which can become problematic. Although bureaus have done a lot of work to fix these issues, the data they use for scorecard building is still not perfect. 

Beam on the other hand connects directly to an applicant’s bank account, accessing user-generated financial data. By obtaining information directly from the bank, we can protect lenders against various types of fraud or manual tampering of PDF statements, it also enables real-time data transmission to lenders with 100% transparency. We see this as important, as the bank statements represent a “truth file” of a user’s spend behaviour and financial position. The user is ultimately giving us permission to analyse their financial position in the most transparent way possible.

What challenges have you encountered in incorporating AI techniques into credit rating and how have you addressed them?

Our biggest challenge over the past six months has been accessing enough data.

The way machine learning works is by learning generalised patterns across large data sets. Algorithms learn these patterns through observing causal relationships across many input and target variables. These patterns become more refined as the dataset grows, allowing models to make accurate predictions across diverse population groups.

A major risk with limited data is model bias, which can arise from various sources including the dataset’s size, collection methods, and inherent historical biases. For instance, a model trained on data from a specific income group, or LSM, would likely develop biases towards that group’s financial behaviours. This results in poor performance when predicting financial behaviours of a lower LSM group due to incorrect associations learned between inputs and outputs.

To combat these challenges, we’ve employed several strategies. Beyond striving for larger datasets, we address data bias and overfitting through techniques such as data augmentation and transfer learning. We also diversify our data sources to ensure our models are trained on datasets that reflect a wide range of demographics and financial behaviours, thereby making our AI models more inclusive and fair.

How do you balance the need for predictive accuracy with the interpretability of AI models, especially in highly regulated Industries?

We ensure both accuracy and interpretability through a battery of tests. These tests are numerous, but are largely either model-specific or model-agnostic. Model-specific methods are tailored to specific types of models, taking advantage of their internal mechanisms and structures. These methods are designed to provide insights into how the model processes inputs to arrive at outputs. 

Model-agnostic methods on the other hand are designed to work with any machine learning model, providing more flexibility. These methods don’t rely on the internal workings of the models, making them more widely applicable and generalisable across various models we test.

This is important, since this allows us to use complex (and thus unexplainable) algorithms (like neural nets), which give us the best performance, and still understand how a final decision is made, across populations, sub-groups or even individual users. This helps steer model development but also ensures we’re not having a disparate impact on the credit decisions of various demographic sub-populations – more simply put – fighting bias.

We use these interpretability techniques, combined with a host of other tools that allow us to fight the risks that may arise from automated decision-making. We collectively call this our “Model Risk Management” framework, and we apply it to all of our machine learning pipelines, ensuring that we get accurate results and that we are following interpretability and broader compliance best practices.

What role do you envision artificial intelligence and machine learning playing in the future of credit vetting and what innovations do you anticipate in this space?

Since the 1980’s, credit scoring at scale has been machine learning driven. The existing scorecards bureaus use are technically machine learning-driven, however they are largely rule-driven and use lots of hand-crafting. Now with the advent and acceleration of big data, novel algorithms and more compute we can unlock new methodologies that weren’t possible previously – being end-to-end “learned”. 

Ultimately, innovation in the credit vetting space will only keep accelerating and the whole industry will see exponential value being added to our vetting capabilities over the next five to ten years.

How is Beam’s approach to credit vetting different to that of traditional credit vetters?

Traditional credit scoring has been somewhat exclusionary, or rather disempowering of people. Consumers have no understanding or influence of their traditional bureau credit score, including whether the data used to evaluate them was accurate or outdated. Beam empowers users to use their own financial data to create a credit profile, whereas before that was completely in the lenders, banks and bureau’s hands. Providers like us are empowering customers to take control of their own financial destinies, providing them with the tools and knowledge they need to improve their financial position in the long run. 

Does Beam reduce exposure to bank statement or transaction fraud?

Processing PDF bank statements opens lenders up to various degrees of fraud. In many instances, customers add fake salary lines or alter transaction data. By using Beam’s automated bank linking technology, we can drastically reduce the risk of declining credit based on fraudulent information. This is because one would physically have to swipe their card and transfer money around to create fake transactions, adding a financial and unscalable time cost to committing fraud. Since it’s now manual and not cost-free, Beam reduces risk for lenders. 

Furthermore, there’s a built-in KYC element to using our automated bank linking technology. We’ve implemented a host of best practices in fraud management, including ID Verification, Anti Money-Laundering and various other Home Affairs checks. In addition, as we collect more and more data, our machine learning models will be able to identify patterns of fraudulent transactions and start to see what an anomaly bank account will look like, ultimately protecting lenders. 

Summary

At Beam, our machine learning and data science approach focuses on analysing granular bank transaction data in real-time, providing a more precise way to assess an applicant’s current financial situation. This approach is designed to build on traditional credit scoring methods to overcome conventional credit bureaus’ reliance on lagged and sometimes inaccurate information. 

We adopt non-parametric approaches to modelling, leveraging abundant data, increased compute power and evolving architectures, particularly deep learning. We believe it’s important to provide better transparency and include user permission in accessing financial data, so that we can empower individuals in managing their financial profiles. 

To conclude, our data science philosophy and machine learning principles reflect our commitment to innovation and inclusivity in reshaping credit vetting practices, ultimately empowering consumers and mitigating risks for lenders.

What is VYC by Beam?

VYC offers a seamless, single-integration solution that enables lenders to conduct comprehensive checks on affordability, identity and credit bureau information. Our user-friendly platform simplifies the process, making it more efficient and accessible for lenders to obtain all necessary verifications through one streamlined integration.

This post is a follow-up to our previous post ‘Beam’s Data Science Philosophy’ where Nic talks about Beam’s data science philosophy. In this blog, Nic discusses Beam’s Modelling and Data principles.

What is Beam’s approach to machine learning?

Beam’s philosophy on data science and risk modelling centres on the premise that the advent of machine learning, the significant increase in data generation and computing power over the last two decades have set the stage for a fundamental shift away from traditional credit scoring methods.

Is this an industry standard way of modelling data or is your approach different?

I think Beam is different across two key dimensions, number one being modelling. The way we’re modelling the data is different and modern. If you look at the traditional credit bureaus for example (and maybe this is oversimplifying because we should definitely give them credit) what they do is they use high-level, overly simplistic features about a person to infer credit risk. What do those look like? Typically, these are demographic features, credit account history, how long you have had your various bank accounts open, where you live (using a postal code as a proxy), your education status, an estimate of how much money you make as well as a few other features. 

Further, your typical consumer credit reports and credit scores also use lagged variables, sometimes carrying errors that go unnoticed. For example, your current employment could be out of date or you could have repaid an account that is showing on the credit report as unpaid since it can take up to six months for that information to be reported to a credit bureau.

We actually also use these features since they provide great signal (and in many cases we will combine features and models), but primarily what we’re using in Beam is data coming from bank statements. We analyse bank transaction data at a much more granular level in order to get much richer insight. By doing this, we get a much more up-to-date and thorough picture about a person’s financial behaviour.

Another reason why Beam is an improvement is that we’ve built our platform to keep our users’ data up to date. We can re-score you every hour, day, week or month. Every time you connect to Beam and share your bank account data, we analyse it in real time to the very last transaction you did – even if it was 30 minutes ago, so that’s really the difference.

Is there a proprietary element to this approach where you do this in a way that another company or player wouldn’t have thought to do it or is yours an industry standard approach?

The underlying model architectures we use are completely commoditized. Similar to how GitHub works, where anyone can view, download and use open source code, there’s the equivalent with machine learning models. So for most of our Data Science workflows, the base models and their architectures are open source, so by definition non-proprietary. I’d say that the three key proprietary pieces of our stack are 1) the particular data sources we use, 2) how we use this data and 3) how we merge together our modeling stack with the rest of our product stack. 

You’ve spoken a lot about using alternative data to predict the likelihood that someone will pay back a loan. Is the fact that you’re using alternative data as well as traditional data plus applying these specific risk principles and algorithms a “secret sauce”? 

The general principle with any kind of machine learning is the more data the better. We’re not just disregarding the traditional credit scoring methods that the current bureaus are using. We’ll combine the two together because the reality is that the traditional scoring methods have actually worked pretty well up to now.

If you look at the Actuarial space, we use modelling practices that are 100 years old. At Wits we learned statistical principles that haven’t changed for 200 years. I’ve always believed that we stand on the shoulders of giants – in an academic sense – and we should respect the discoveries they’ve left us. We would be remiss to completely disregard that and think that there’s nothing to be learned from them. I think if you adopt these actuarial risk principles, combined with modern-day compute and modelling technologies – you’ll get the best results. 

Now to answer your question, the cash flow data part of what Beam does is quite important if you think of the way traditional credit scoring is done by the bureaus. Basically, you need to be “credit active”, in other words, have an existing credit product in order to be scored. As you can imagine, this poses a chicken-and-egg problem for those that are financially excluded, as they are declined credit since they can’t be scored (deemed “thin file customers”) but can’t access the credit product to start building up a repayment history.

What we are doing is we’re learning the mapping between a completely independent data source – bank statement data – and credit outcomes. This is fundamentally more inclusive since ~90% of South Africans have bank accounts. In contrast, some estimates gauge that up to a third of South Africans are “thin file customers” and thus can’t get access to credit.

This being said, all of the data we get access to is user-permissioned, meaning that customers opt-in to sharing their data with us. Most South Africans aren’t aware of the fact that their bank transaction data is actually their property and they have the power and right to share this data with third party service providers like Beam for better and more fair access to financial services. This plays on our core thesis as a business, being that Open Banking and Open Finance can be a huge enabler for the average person, but more on that in another post.

We interviewed Beam’s co-founder and Head of Data Science, Nicolas Gonatas, to unpack his philosophy around building the Beam risk engine.

Nic, tell us about your background.

After graduating from the University of the Witwatersrand with a BSc in Actuarial Science in 2019, I joined PwC as an Actuarial Analyst working across several sectors including Life, Health and Short-term Insurance, Banking and Financial Instrument Valuation. I later moved into the Data Science team since I had a coding background and was somewhat proficient in Python. This move started when I was posted to a team that was tasked with translating actuarial models into Python code in order to better scale up and handle vast amounts of data. We then started to just build our own models from the ground up, and that’s pretty much data science. I worked on that for around two years, ending up managing the team. We consulted to the large Telcos and Banks, building machine learning and risk models – essentially data science within the Actuarial space. Shortly after, I co-founded PYGIO, a software engineering company with Dimitri and Ronnie, where we joined forces with Tim and Adam on Beam. 

What is your role at Beam?

I’m a co-founder of Beam and my role is Head of Data Science, so I look at data, seeing what we require and using novel techniques and modelling practices to predict and manage risk better.

What is your approach or philosophy around data science and risk modelling?

Our core thesis at Beam is that the world has fundamentally changed and that there’s an opportunity to move away from traditional credit scoring methods that use parametric approaches. In this approach, you make assumptions that data follows a particular distribution, together with a series of simplifying assumptions to make a limited class of statistical models make predictions. This approach contrasts with non-parametric methods, which do not assume a specific form for the underlying distribution and are more flexible in adapting to the data’s shape – we subscribe to the latter methodology at Beam.

There are three secular trends driving what’s happened in the last 20 years of machine learning advances. Number one is data. As I’ve said before, there’s more data being generated today than ever before. Think of the footprint one leaves with their bank transaction (think credit card) data, whereas previously it was all cash and was thus untracked. Not to mention the data being generated by one’s internet and social media footprint!

The second piece is compute power. If you look at Moore’s Law, it refers to how exponential growth is at the heart of the rapid increase of computing capabilities. The law was first described by Gordon E. Moore, the co-founder of Intel, in 1965 and it was the observation that the number of transistors on computer chips doubles approximately every two years.