Time-Series Forecasting for Industrial Demand: A Comparative Study

Abstract

Accurate demand forecasting is critical for industrial manufacturers to optimize production planning and resource allocation. This study evaluates state-of-the-art time series forecasting models on a real-world dataset of monthly order intake for commercial vehicles from Company X, organized hierarchically by market, product group, and power supply.

In this study, classical statistical methods (AutoETS, AutoARIMA) and gradientboosting trees (XGBoost, LightGBM) are implemented and compared with recently developed state-of-the-art deep learning architectures (Temporal Fusion Transformer, TimeXer, TiDE, N-HiTS, SOFTS). Expanding-window cross-validation is employed to generate multi-horizon forecasts up to 12 months and accuracy is evaluated using RMSE and sMAPE. Models are compared with and without the utilization of exogenous variables to highlight their robustness in processing external signals to aid forecasting accuracy.

The results indicate that tree-based and statistical models, particularly LightGBM and AutoARIMA when augmented with exogenous variables, achieve the lowest average errors, suggesting that state-of-the-art models are not competitive in this setting. However, while these models are less accurate on aggregate metrics, they tend to produce forecasts with richer temporal dynamics, capturing trends and seasonality more effectively.

Overall, the findings suggest that no single approach consistently outperforms others; model effectiveness varies depending on the forecast horizon, aggregation level, and the availability of external covariates. These results highlight the importance of selecting models based on forecasting context, particularly in data-scarce industrial environments. An interesting direction for future research could be to explore hybrid or ensemble approaches to combine accuracy with actionable temporal structure.