Forecasting Without Fine-Tuning: Zero-Shot Foundation Models vs. Traditional Approaches

This study conducts a comparative evaluation of the zero-shot forecasting model TiRex and five widely used forecasting techniques using real-world production data. The primary goal is to assess TiRex’s ability to forecast weekly SKU-level demand across multiple time series with varying lengths, data availability, and seasonal patterns. To simulate realistic forecasting conditions, a rolling-origin (walk-forward) validation framework is applied. Forecast accuracy is primarily measured using Mean Absolute Error Percentage, enabling consistent comparisons across different time series. Results show that while traditional models such as XGBoost and Prophet provide stable and reliable performance, the TiRex model achieves comparable accuracy without requiring task-specific fine-tuning. These findings highlight the potential of zero-shot forecasting for scalable, efficient deployment in dynamic and data-constrained production environments.