Transaction categorization (identifying the “for what” of a given booking event) is a crucial building block in a bank’s financial decision-making. At Rabobank, transaction categorization empowers teams to fight economic crime, mitigate lending risk, and support customers’ financial health with relevant, accurate insights.
In this presentation, we explore how better data processing and analytics can improve categorization for payment requests and help us serve customers better. Accurately categorizing these transactions is difficult due to their non-standardized data structure and content, representing a huge opportunity to better understand and support our customers.

We present a new parsing tool that standardizes relevant data from payment request transactions, improving our ability to categorize them using the same systems we employ for normal transactions. With better data processing, we show a significant improvement in model performance but also highlight that the information contained in a payment request is highly heterogeneous, causing problems for a rules-based categorization engine.
To tackle this language-processing problem, we evaluate a range of NLP ML/AI techniques on a subset of parsed payment-request transactions. We assess these techniques based on statistical performance, regulatory/compliance burden, technical responsiveness, and sustainability/operations cost.
Finally, we summarize the results of our analysis and highlight what we see as the best balance of performance, cost, and risk in tackling the difficult but rewarding problem of categorizing payment requests.