What does this research mean for the field?

Combining zero-pattern and temporal lag features significantly improves intermittent demand forecasting across deep learning architectures, with GRU and LSTM outperforming TCN, while revealing a fundamental trade-off between zero-demand detection and non-zero demand prediction. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To evaluate how feature engineering strategies and novel zero-detection metrics influence intermittent demand forecasting accuracy across GRU, LSTM, and TCN architectures.

August 30, 2026Journal of Intelligent ManufacturingOpen Access

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

View Full Paper

Ask AI

Bookmark

Authors

AEAhmed Othman El-meehy AEAmin K. El-Kharbotly MEMohammed M. El-Beheiry

Discussion

Member takes

Overview

Empirical study demonstrates that combining zero-pattern and lag features improves intermittent demand forecasting in deep learning models, highlighting operational inventory trade-offs.

Key Points

To evaluate how feature engineering strategies and novel zero-detection metrics influence intermittent demand forecasting accuracy across GRU, LSTM, and TCN architectures.
Benchmarked GRU, LSTM, and TCN deep learning models across nearly 1,140 forecasting scenarios constructed from 19 real-world demand datasets.
Evaluated 20 distinct feature engineering strategies combining temporal lag context and zero-pattern features.
Introduced two decomposed accuracy metrics, Z% (zero-demand accuracy) and NZ% (non-zero-demand accuracy), alongside WMAPE% to assess forecast performance across varying lumpiness levels.
Combined lag and zero-pattern features yielded the most robust forecasting improvements across all architectures, with GRU and LSTM consistently outperforming TCN.
Zero-detection accuracy (Z%) negatively correlated with forecast error (WMAPE%), with the correlation strengthening as demand intermittency grew (pooled r = −0.495 at average demand interval ≈ 2.00).
Higher Z% values corresponded to reduced inventory holding levels accompanied by increased shortage risks, whereas gains in NZ% demonstrated the opposite operational trade-off.

Cite This Study

El-meehy et al. (2026) studied this question.

synapsesocial.com/papers/6a93f1396c1a8fb52e79e015 https://doi.org/10.1007/s10845-026-02964-7

View Full Paper

Ask AI

Bookmark

August 30, 2026Journal of Intelligent ManufacturingOpen Access

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

View Full Paper

Ask AI

Bookmark

Authors

AEAhmed Othman El-meehy AEAmin K. El-Kharbotly MEMohammed M. El-Beheiry

Discussion

Member takes

Overview

Empirical study demonstrates that combining zero-pattern and lag features improves intermittent demand forecasting in deep learning models, highlighting operational inventory trade-offs.

Key Points

To evaluate how feature engineering strategies and novel zero-detection metrics influence intermittent demand forecasting accuracy across GRU, LSTM, and TCN architectures.
Benchmarked GRU, LSTM, and TCN deep learning models across nearly 1,140 forecasting scenarios constructed from 19 real-world demand datasets.
Evaluated 20 distinct feature engineering strategies combining temporal lag context and zero-pattern features.
Introduced two decomposed accuracy metrics, Z% (zero-demand accuracy) and NZ% (non-zero-demand accuracy), alongside WMAPE% to assess forecast performance across varying lumpiness levels.
Combined lag and zero-pattern features yielded the most robust forecasting improvements across all architectures, with GRU and LSTM consistently outperforming TCN.
Zero-detection accuracy (Z%) negatively correlated with forecast error (WMAPE%), with the correlation strengthening as demand intermittency grew (pooled r = −0.495 at average demand interval ≈ 2.00).
Higher Z% values corresponded to reduced inventory holding levels accompanied by increased shortage risks, whereas gains in NZ% demonstrated the opposite operational trade-off.

Cite This Study

El-meehy et al. (2026) studied this question.

synapsesocial.com/papers/6a93f1396c1a8fb52e79e015 https://doi.org/10.1007/s10845-026-02964-7

View Full Paper

Ask AI

Bookmark