DEMAND PATTERN-SPECIFIC FORECASTING FOR SPARE-PART INVENTORY: ENSEMBLE MODEL EVIDENCE FROM INDONESIAN HEAVY EQUIPMENT DISTRIBUTION

Abstract

PT Cakra Harmoni Sentosa (HCS), the largest heavy equipment distributor in Indonesia, faces a supply chain performance problem at its highest-revenue plant, Plant Sungai Danau (SDU). With IDR 1.1 trillion in annual spare-part transactions, SDU records a customer service level of 75%, which is five percentage points below the 80% target, while Days of Inventory reaches 87 days against the 75-day benchmark. Both shortfalls originate from reliance on a 12-month Moving Average forecasting method applied uniformly across a portfolio where 90% of SKUs show intermittent or lumpy demand patterns. This paper evaluates 23 forecasting model variants per demand pattern through walk-forward validation. Methods tested range from classical statistical approaches including Croston, SBA, and TSB to machine learning models such as XGBoost and LightGBM, deep learning architectures including LSTM and N-Beats, and ensemble combinations. Performance is assessed using RMSE and MAE on non-zero periods alongside the Stock-Keeping-Oriented Prediction Error Cost metric with stockout-weighted parameters. Findings show that a demand-pattern-specific Ensemble of SES, SBA, LSTM, and N-Beats achieves the strongest performance for lumpy and intermittent SKUs, cutting RMSE by 43.9% and SPEC by 36.3% on lumpy items relative to the MA12 baseline.