Timestral Tested: How Much Data AI Demand Forecasting Needs (winder.ai)

In a recent test of Mistral's newly released Timestral for AI demand forecasting, the model demonstrated promising results using minimal historical data. Evaluated against various forecasting methods, including Amazon’s Chronos-2 and Google’s TimesFM, Timestral achieved a retail forecasting error 14% lower than the traditional Excel method and 3% better than LightGBM when provided with two days of demand history. The tests underscore the model's effectiveness in a zero-shot forecasting context, where it makes predictions without training on specific datasets. However, Timestral struggled with seasonal fluctuations not present in its training history, highlighting a limitation in its predictive capabilities. This trial is significant for the AI/ML community as it emphasizes the ongoing evolution of pretrained models and their applications in real-world scenarios. The findings suggest a critical balance between leveraging historical data and understanding product-specific behavior, particularly with new or seasonal items. While Timestral excels under certain conditions, it accentuates the need for supplemental knowledge about external variables, like weather impacts on demand, to enhance predictive accuracy. The results also raise vital questions about the reliance on pretrained models in dynamic forecasting environments, ultimately advocating for hybrid approaches that incorporate traditional demand modeling.