Predictive Analytics of Defective Machinery Parts in Reverse Supply Chain: A Case Study at ASML

Abstract

Accurate forecasting in Reverse Supply Chain (RSC) management is crucial for the semiconductor industry, particularly for companies like ASML, which must efficiently manage the return flow of defective machinery parts. This study addresses key gaps by developing and evaluating time series-based forecasting models tailored to ASML's RSC. Using a modified ABC-analysis, parts were categorized based on defect frequency and economic impact, focusing on the most critical components. The research applied and optimized models including SES, ARIMA, ARIMAX, and LSTM, using five years of historical defect data. The analysis showed that LSTM models excel in high-frequency (weekly) forecasts for parts with frequent early-life defects, achieving an average mMAPE of 26.32\%. ARIMAX models performed best for lower-frequency (monthly) data, particularly in sparsely represented classifications, with mMAPE as low as 4.31\% to 10.80\%. Despite a higher mMAPE of 85.42\% in one outlier, ARIMAX emerged as the most balanced model, offering a practical trade-off between accuracy and computational efficiency. Furthermore, the study highlights the computational efficiency of ARIMAX, which, although more demanding than SES, provided a favorable balance, with ARIMA and LSTM being more resource-intensive. These findings demonstrate ARIMAX's suitability for long-term forecasting and broader trend analysis, making it the preferred model for ASML's RSC. This research provides a robust, data-driven framework that enhances inventory management and capacity planning, making it possible to predict return flows with low error rates on a monthly basis, particularly in high-tech industries where many defective parts are returned. Future research should incorporate additional dynamic variables, explore hybrid models, and refine data splitting techniques to further improve predictive accuracy and support sustainable supply chain operations.