An individual's salary is one of the most valuable data points extracted from raw open banking data. It is also one of the more difficult to assign accurately. This week sees FriendlyScore upgrade their transaction classification model to address this issue directly.
What makes open banking data so powerful is the sheer depth of useful information it contains. One tool crucial in extracting this is transaction classification, which involves assigning familiar categories, such as groceries, bills or rent, to individual transactions. A loan underwriter, for example, may then quickly identify how much an individual has spent or received in that category and factor it into their decision.
One critical category from this perspective is salary. For example, details of an individual's salary strongly infer the size of loan repayments they can comfortably make. However, this is also one of the most problematic classes of open banking data to categorise reliably.
The Problem With Salary Classification
Industry-standard transaction classification models employ a specific range of machine learning techniques to analyse bank transactions. However, open banking transactions often exhibit a wide range of characteristics which are difficult for even the most well-trained models to handle. Let's consider a simple example. The expenditure transaction below almost certainly represents a purchase of clothing:
This transaction would be pretty straightforward for a basic model to handle as there are few likely alternatives. But now consider the same transaction as an income:
The above could now plausibly represent a refund or a salary. There is nothing in the data shown that could reliably distinguish it either way, which immediately adds another layer of complexity. Transaction data often exhibits this type of ambiguity which can easily confuse a generic machine learning model and lead to salary data going undetected. This leads to suboptimal analytics which negatively impacts underwriting decisions downstream.
A New Benchmark
FriendlyScore has developed a proprietary set of modelling techniques that move beyond the industry-standard machine learning tools to solve this problem. With this technology, individual transactions are now modelled within the broader context of an individual's account data to specify transaction classes more accurately.
Using these techniques, FriendlyScore has boosted their high watermark of transaction classification accuracy past ~90%. But more importantly, by reliably identifying the all-important salary class, clients can now have even more confidence in the crucial decisions made based upon the data which FriendlyScore provides.
FriendlyScore is a London-based software house which helps clients get the most out of open banking data. We service a diverse client base, which includes banks, credit unions, financial advisors, tenant referencing agencies, mortgage brokers & advisors, debt collection agencies and gaming companies. To find out more and trial our software for free, visit www.friendlyscore.com today.