Feature engineering for intermittent demand forecasting: zero-detection and forecast performance across GRU, LSTM, and TCN architectures

Abstract

Accurate forecasting of intermittent demand remains a persistent challenge for deep learning methods, partly due to the inadequacy of conventional error metrics in capturing zero-demand forecasting performance. This study tackles this gap by introducing Z% and NZ%, two decomposed accuracy metrics that distinctly quantify a model’s ability to correctly forecast periods of zero and non-zero demands, respectively, revealing a performance trade-off between zero and non-zero demand prediction, and investigating their relation to the overall forecast accuracy across various degrees of lumpiness. Using nearly 1140 forecasting scenarios derived from 19 real-world demand datasets and 20 feature engineering strategies, three deep learning architectures are systematically evaluated: GRU, LSTM, and TCN. Results reveal that combined lag features deliver the most significant improvement in both WMAPE% and Z%, while individual zero-pattern features require temporal lag context to show reliable gains, yet fail if isolated. Combined zero and lag features emerge as the most robust configuration, resulting in significant improvements in both metrics simultaneously across all deep learning architectures, highlighting the complementary relation between zero-pattern and temporal demand information. GRU and LSTM outperform TCN, which exhibits limited ability to make use of zero-pattern information regardless of feature configuration. Correlation analysis shows a negative correlation between Z% and WMAPE%, intensifying with the degree of intermittence (pooled r = − 0.495 at ADI ≈ 2.00), confirming the increasing critical role of zero-detection accuracy as the demand sparsity grows. Further analysis demonstrates that improvements in Z% are strongly associated with reduced inventory levels but may increase shortage risk, whereas improvements in NZ% exhibit the opposite effect, highlighting a fundamental inventory–service trade-off. These findings emphasize the diagnostic value of Z% and the complementary but secondary role of NZ%, providing empirical guidance on feature engineering, model behavior, and deep learning architecture selection for intermittent demand forecasting with direct implications for operational decision-making.