Real-Time Retail Forecasting and Anomaly Detection Using Hybrid ARIMA and Neural Network Models

Khadija Elkattany (Hubei University of Automotive Technology)
Md Mutasim Billah (St. Francis College)
Shahin Alam (Hubei University of Automotive Technology)

Control Systems and Optimization Letters
Vol 4, No 2 (2026)

This paper presents a hybrid machine learning framework that addresses scalability and accuracy challenges in retail inventory management by integrating real-time demand forecasting with anomaly detection, evaluated using Walmart's historical sales data. Traditional approaches face a trade-off: maintaining individual models for each product category is computationally prohibitive, while generalized models often underperform for dissimilar items, resulting in stockouts or overstocking. To address this, we propose a department-level aggregation strategy that balances specificity and generalization, combined with a hybrid methodology: ARIMA for linear trend and seasonality modeling, cubic spline interpolation to capture nonlinear residual patterns, and neural networks for complex interactions. The framework dynamically adjusts predictions using real-time sales streams and applies residual-based anomaly detection with threshold triggers to identify sudden demand spikes or supply disruptions. Experiments on a filtered Walmart dataset (removing returns, canceled orders, and items with 30 days of historical data; 18 months, 15 departments, aggregated from 100,000+ SKUs) indicate an 18% reduction in mean absolute error (MAE) compared to exponential smoothing baselines (MAE: 235.1 ± 32.8 vs. 310.4 ± 28.5), while spline-enhanced neural networks achieve a 24% improvement over standalone ARIMA (MAE: 235.1 vs. 310.4; p 0.01). The anomaly detection module identifies 92% of simulated irregularities with a 7% false-positive rate and F1-score of 0.89. The proposed framework provides three principal advantages: (1) scalable department-level modeling without per-product customization, reducing training time from 2 hours per product to 12 minutes per department (90% improvement); (2) real-time adaptability to fluctuating demand through 6-hour incremental LSTM updates; and (3) cost-efficient inventory optimization through integrated anomaly alerts, validated in a 6-month pilot across 50 Walmart stores showing 19% stockout reduction and 14% overstocking reduction. This work offers a practical blueprint for retailers to enhance forecasting precision, mitigate supply chain risks, and reduce operational costs in volatile markets.