Effective forecasting of automotive spare parts is essential for improving business performance, yet practical application often lags due to the complexity of models and limited resources. This research employs machine learning techniques like Extremely Randomized Trees (ETR) and XGBoost, comparing them with classical methods such as Moving Average (MA) and Triple Exponential Smoothing (TES). To improve forecasting accuracy given the large number of items and diverse product characteristics, this study also introduces a novel demand classification using K-means clustering with the coefficient of variation (CV) and the average interval between demands (ADI). By analyzing 1600 products, the research segments them into four demand patterns, finding that machine learning generally outperforms classical models, particularly in handling high variation and intermittent demand. TES also outperformed MA in terms of mean absolute error (MAE). The findings suggest businesses should adopt machine learning for more accurate forecasting in the automotive spare parts industry.