Exploring Chronos: How fundamental ai models are setting new standards in predictive analytics
This publication was co-authored with Rafael Guedes.
Time series forecasting has been evolving into basic models due to its success in other areas of artificial intelligence (ai). In particular, we have witnessed the success of these approaches in natural language processing (NLP). The pace of development of foundational models has accelerated over time. A new, more powerful Large Language Model (LLM) is released every month. This is not limited to NLP. We also see a similar increasing pattern in computer vision. Segmentation models such as Meta's Segment Anything Model (SAM) (1) can accurately identify and segment objects in unseen images. Multimodal models like LLaVa (2) or Qwen-VL (3) can handle text and images to answer any user question. The common feature between these models is that they can perform accurate zero-shot inference, meaning they do not need to be trained on your data to perform excellently.
Defining what a foundational model is and what differentiates it from traditional approaches is probably beneficial at this time. First, a fundamental model is large-scale (i.e., its training), which provides a broad understanding of the main patterns and important nuances that we can find in the data. Second, it is general purpose, that is, the fundamental model can perform various tasks without requiring specific training for them. Although they do not require task-specific training, they can be fine-tuned (also known as transfer learning). They adapt to relatively small data sets to perform better on that specific task.
Why is it so tempting to apply it to time series forecasting based on the above? Primarily, we design fundamental models in NLP to understand and generate text sequences. Luckily, time series data is also sequential. The above point also aligns with the fact that both problems require the model to automatically extract and learn relevant features from the data sequence (temporal dynamics in time series data). Furthermore, the general-purpose nature of the fundamental models means that we can adapt them to different forecasting tasks. This flexibility allows a single, powerful model to be applied across multiple domains and…
