Comparison of Statistical and Machine Learning Methods for Improving Demand Forecasting of Spare Parts in Simulators and Weapon Ranges

According to the 2015 U.S. Government Accountability Office (GAO) report, Department of Defense (DoD) supply chain management has been classified as a high-risk area since 1990, due in part to weaknesses in accurately forecasting the demand for spare parts. The inability to accurately perform inventory planning of spare parts in simulators and weapon ranges contributes to the accumulation of 3.6 billion dollars of excess inventory for the Army. Selection of an accurate predictive model is necessary for the Army to improve their forecasting capabilities in support of inventory planning for simulators and weapon ranges. Traditional statistical methods have been used historically to project intermittent demand of spare parts leading to inaccurate forecasts resulting in slow-moving inventory and increased storage costs. This praxis extends the existing body of knowledge through the identification of the most accurate model type for inventory demand forecasting of spare parts in simulators and weapon ranges operating in non-manufacturing environments. Results show that Support Vector Regression (SVR) is the best performing method when forecasting repairable and consumable part demand for simulators classified as Non-Mission Capable (NMC), Partially Mission Capable (PMC), Fully Mission Capable (FMC), as well as scheduled and non-scheduled maintenance actions over a 52-week forecast horizon. K-Nearest Neighbor Regression (KNNR) tied the performance of SVR as the best performing method when forecasting material demand for non-scheduled maintenance actions that have classified a system as Partially Mission Capable (PMC). Results show that the machine learning models used in this praxis were more accurate than statistical methods in forecasting repairable part demand for scheduled and non-scheduled maintenance actions that covered 2 and 4 years. However, the Teunter, Syntetos, and Babai (TSB) statistical forecasting method performs better than some machine learning methods when estimating repairable part demand for scheduled and non-scheduled maintenance actions that covered 6 years. The identification of the most accurate forecasting model is substantiated by determining if there is a statistically significant difference in the forecasting accuracy of statistical and machine learning methods using real-life data. Engineering and logistics managers can utilize this praxis as a reference in the selection of the best performing forecasting method, to support inventory planning.

Author

• Hendricks, Corey T

Language

• en

Keyword

• intermittent demand
• Machine Learning Forecasting
• Forecasting accuracy
• SKU demand
• Time series Forecasting
• Statistical Forecasting

Date created

• 2022

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Engineering Management

Degree

• D.Engr.

Advisor

• Yassan, Rebecca
• Blackford, J.P.

Committee Member(s)

• Etemadi, Amir

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/j3860773c