Hybrid Forecasting Frameworks for Fraud Rate Estimation: Addressing Delayed and Geographically Fragmented Label Arrival in Digital Commerce Platforms

Abstract

Fraud detection in digital commerce is complicated by delayed ground-truth labels, geographically heterogeneous fraud patterns, and continuous changes in fraudulent behavior. Conventional detection models may produce biased estimates when trained on incomplete transaction labels, while differences in regional payment infrastructures and dispute-resolution processes further limit model generalizability. This paper proposes a hybrid forecasting framework that integrates Autoregressive Integrated Moving Average (ARIMA) modeling, Long Short-Term Memory (LSTM) networks, and survival analysis to estimate fraud rates under delayed and geographically fragmented label conditions. The ARIMA component captures aggregate fraud trends and seasonality, while the LSTM component learns sequential transaction-level behavioral patterns. Survival analysis estimates label-completion probabilities, enabling incomplete observations to be assigned appropriate weights during model training. The framework further incorporates market-specific time-series parameters, domain-adaptive transfer learning, and incremental learning strategies to address geographic distribution shifts, concept drift, and catastrophic forgetting. A comprehensive evaluation framework is outlined using temporal holdout validation, label-complete and operational evaluation regimes, precision, recall, false-positive rates, and probability calibration. The proposed architecture also considers real-time inference, scalable deployment, and decision explainability. By integrating these complementary methods, the framework provides a structured approach to fraud rate estimation when transaction outcomes are incompletely observed or become available only after substantial delays. Future research should validate the integrated framework empirically and develop benchmark protocols that jointly account for label delay and concept drift.