Hybrid Machine Learning Framework for Personalized Spending Forecasting in E-Commerce Ecosystems Using the Digital Payment Method

Abstract

Online shoppers in India increasingly pay through the Unified Payments Interface (UPI), and payment apps hold detailed records of when and how people spend. Forecasting an individual's spending for the next month from such records can support budgeting tools, well-timed offers, and payment recommendations. This paper proposes a hybrid framework for personalized spending forecasting. It combines exponential smoothing, a gradient-boosted tree learner, and a neural network, where the last two model what smoothing leaves unexplained. A stacking layer then blends them using weights learned for each behavioral segment. A UPI context layer is specified for inputs such as transaction success, failure history, and reward intensity. The framework is evaluated on the public Online Retail II e-commerce log, covering 5,942 customers, using five forecast months and five random seeds. It lowers the scaled forecast error by about 25% compared with a seasonal naive benchmark (0.761 versus 1.011) and matches LightGBM, but simple exponential smoothing is slightly better overall (0.747). Learned components help most for customers with short or occasional purchase histories (0.50 versus 0.79 for new customers). Segment-specific blending gave no measurable gain, and nominal 80% prediction intervals covered only 62% of outcomes. UPI transaction data were not available, so the UPI layer is specified but not tested, and the paper makes no claim about UPI behavior.