An explainable machine learning framework with trend-sensitive features for predicting high-potential merchants in mobile payment ecosystems

Mobile payment platforms face a persistent decision-support challenge: payment service providers (PSPs) must identify high-potential merchants from large merchant bases without relying solely on retrospective transaction scale. This study develops an explainable decision-support framework for merchant prioritization by combining growth-trajectory feature engineering, XGBoost classification, and SHAP-based interpretability. Using 61,280 merchant-month records from 3,939 merchants of a leading payment service provider between January 2022 and March 2025, the framework predicts merchant potential while translating model outputs into decision-usable explanations. Merchant potential is operationalized through a polar-group label: merchants in the top 20% of cumulative future gross merchandise value (GMV) over the prediction horizon are treated as high-potential and those in the bottom 20% as low-potential, so the term “high-potential” denotes this top-quantile future-GMV definition rather than current transaction size. To establish methodological rigor, the framework is benchmarked against rule-based, linear, ensemble, and stacking baselines, and a feature-group ablation is conducted to isolate the contribution of trend-sensitive variables. The proposed framework attains strong predictive performance on the held-out test set (AUC = 0.841; AP = 0.868; F1 = 0.783; Precision@100 = 0.91; Lift@40% = 1.61) while outperforming logistic regression, random forest, LightGBM, raw-GMV XGBoost, and rule-based screening on operational precision. SHAP analysis indicates that the nine-month GMV trend slope and the six-month average GMV jointly shape merchant potential, with growth trajectory providing incremental predictive value beyond static scale. The study contributes to merchant analytics by reframing explainability as a structured, auditable decision-support layer rather than as a supplementary visualization. For PSPs, the proposed framework supports transparent merchant prioritization, internal auditability, and targeted resource allocation while avoiding overdependence on opaque prediction scores or simple GMV-based heuristics. As boundary conditions, the results are drawn from a single PSP and a polar-group (top-versus-bottom-quantile) sample rather than the full merchant population, and the decision-support layer is assessed structurally through mapping consistency and rank stability rather than through user studies, formal audits, or measured business outcomes.