Chronos: Learning the Language of Time Series
Abstract:
Time series forecasting plays a vital role in decision-making across various sectors like energy, retail, and finance. Traditionally, practitioners have focused on creating task-specific forecasting models tailored to specific datasets or application domains. Inspired by the success of pretrained Large Language Models (LLMs), it becomes imperative to explore whether a similar approach can be applied to forecasting: Can we train a single model on large amounts of diverse time series data, that will generalize to new unseen time series tasks? Recent work has made some strides in this direction by leveraging LLMs for forecasting via zero-shot prompting or fine-tuning, as well as by pretraining large models with sophisiticated time-series-specific modifications. In this talk, we challenge the necessity of these time-series-specific designs and introduce Chronos, a family of pretrained forecasting models based on minimal modifications to language model architectures. Specifically, we only replace the text tokenizer with scaling and quantization, thereby tokenizing time series values into a fixed vocabulary. Chronos models, trained on a vast corpus of both real and synthetic time series data, exhibit remarkable (zero-shot) performance on unseen datasets. This positions pretrained models as a promising tool to greatly simplify forecasting pipelines. Our models are available open-source at https://github.com/amazon-science/chronos-forecasting and have garnered positive feedback from both time series researchers and practitioners.
Bio:
Abdul Fatir Ansari is an Applied Scientist at Amazon Web Services working on time series forecasting and log analytics. His research interests lie in the areas of time series analysis and generative modeling, encompassing probabilistic generative modeling, variational inference, unsupervised learning, and representation learning. He has published in and served as a reviewer in top machine learning venues such as ICML, ICLR and NeurIPS. Before joining Amazon, Abdul Fatir obtained his PhD from the National University of Singapore in 2022, where he received the Dean’s Research Excellence Award.