Primary Submission Category: Machine Learning and Causal Inference
Identifying Recurring Payments In Financial Transaction Data
Authors: Nate Bradshaw, Dylan Zwick, Joshua Jensen, Robert Ball,
Presenting Author: Nate Bradshaw*
Recurring payments for goods and services are a significant and growing source of expenses for individuals and institutions. While some payments arise from habitual behavior (e.g. daily coffee purchases), others are true recurring transactions such as subscriptions, utilities, and leases. Distinguishing between these patterns is essential for financial transparency but is challenging due to irregularities in real-world transaction data.
This work investigates methods for identifying recurring payments at the transaction level. We evaluate periodicity detection techniques, including Fourier transforms and convolution-based approaches, and establish a Z-score baseline based on the variance of time gaps between transactions. While this baseline achieves high recall (94.14%) on real-world data from Weber State University, its precision (28.39%) is limited by its inability to separate recurring signals from habitual noise.
To address this limitation, we engineer 25 features capturing temporal and transactional structure and train multiple machine learning models, including XGBoost and multi-layer perceptrons, using synthetic, real, and hybrid datasets. We show that augmenting real data with synthetic examples that model irregular habitual behavior improves precision while maintaining strong recall. Our results indicate that an XGBoost model trained on combined data provides the most effective balance, enabling more accurate identification of recurring expenses.