Group Anomaly Detection and Risk Control of Commodity Sales Volume Data Based on LSTM-VAE Framework

Abstract

In recent years, the increasing demand for sales-volume analytics has driven growing interest in deep-learning-based anomaly detection methods for intelligent supply-chain management and operational risk control. Using point-of-sale data from brick-and-mortar retail stores, this study proposes a group-anomaly detection approach based on an LSTM-VAE framework to identify contiguous anomalous segments that are likely associated with stock-out events or supply disruptions, thereby enabling timely early warning. The LSTM-VAE (Long Short-Term Memory-Variational Auto-Encoder) integrates LSTM networks to model temporal dependencies in time-series data with the generative learning capability of a variational auto-encoder. The model is trained exclusively on historical data representing normal operational weeks, allowing it to learn baseline sales dynamics and effectively distinguish normal patterns from anomalous behaviors. During the testing phase, time-series segments exhibiting persistently high reconstruction errors are identified as potential periods of stock-out risk or supply-chain disruption. The dataset consists of weekly sales records collected from 45 retail stores in the United States between 2010 and 2023, comprising more than 50 000 observations. Each record includes sales volume as the target variable, along with inventory levels, weekly timestamps, store identifiers, product identifiers, temperature information, macroeconomic indicators, and special-holiday dummy variables. Data preprocessing procedures include missing-value imputation, outlier handling, and feature standardization to ensure model robustness. Experimental results show that the proposed LSTM-VAE model performs effectively in group-anomaly detection by accurately capturing contiguous abnormal segments associated with inventory shortages, thus providing practical insights for optimizing replenishment strategies. Overall, the proposed approach offers a scalable and efficient solution for risk control in large-scale product-level sales monitoring scenarios.