Forecasting continuous and intermittent demand: A segmented empirical benchmarking framework for mixed-demand portfolios

Abstract

Accurate demand forecasting is a significant challenge in retail, particularly with products exhibiting both continuous and intermittent demand patterns. Most existing studies analyze these demand types separately, limiting insights into model performance. This study presents a segmented empirical benchmarking framework that evaluates various forecasting methods—statistical, intermittent-demand, and deep learning—within a mixed-demand context, using consistent validation strategies. The framework classifies 577 products based on demand characteristics and utilizes tailored models, including Facebook Prophet (additive and multiplicative), intermittent-demand methods (ADIDA, IMAPA, TSB, optimized SES), and deep learning models (RNN and LSTM). It is validated with transaction data from an Indonesian retailer in the Bicycle Parts, Accessories, and Apparel (PAA) sector, covering 2019-2025 through sliding and expanding window techniques. Results based on Mean Absolute Error (MAE) show that aggregation-based methods such as ADIDA and IMAPA achieve lower error levels for intermittent demand than the company’s current forecasting approach. For continuous demand, differences across models are relatively small, with no single method consistently dominating. Overall, performance differences across models remain moderate, and Diebold-Mariano test results suggest that these differences are not consistently statistically significant. The findings indicate that forecasting performance depends on aligning models with demand characteristics rather than relying on a single approach. The proposed framework provides practical guidance for model selection in mixed-demand environments.