AlFayyad_Forecast_Accuracy_and_Stock_Decisions.pdf
More accurate demand forecasts do not automatically produce better stock decisions. This paper examines that distinction using the results of an MSc dissertation on predictive analytics in retail supply chains. The study used a public, synthetic Kaggle dataset containing 73,100 records and 15 attributes, with daily observations described in the results chapter as covering January 2022 to January 2024. It compared SARIMAX, Random Forest, XGBoost and a Long Short-Term Memory network. Random Forest achieved the lowest reported test root mean squared error, 8.58, followed by XGBoost at 9.75; the LSTM recorded 108.26. A separate inventory-impact proxy compared XGBoost with a seven-day moving average. XGBoost reduced mean absolute error from 192.3 to 106.9, yet its under-forecast rate increased from 42.37% to 49.59%. The corresponding service-level proxy declined. These results distinguish the magnitude of forecast error from its direction and support integrating forecasts with safety-stock buffers and replenishment rules. They do not establish realised stockout reductions or an optimal buffer. The paper closes with a limited implication for essential-medicines planning: forecast accuracy should inform stock decisions alongside shortage protection, with any application requiring evaluation on medicine-specific data.
Authors
- Hussam Al-Fayyad (ORCID: https://orcid.org/0009-0002-5303-7272)
Institutions
- University of Arizona (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24227484
- Primary Topic
- Forecasting Techniques and Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00