Time Series Foundation Models

Chronos (Amazon), TimesFM (Google), and MOMENT for zero-shot forecasting

Overview

Unlike NLP or Vision, Time Series lacks massive, cohesive public repositories, making Foundation Models (FMs) difficult to build. However, recent models have solved this by leveraging massive pre-training on diverse telemetry, finance, and weather datasets to enable zero-shot forecasting—predicting horizons without any task-specific fine-tuning.

Amazon Chronos

Chronos is a family of time series models by Amazon that treats time series like a language modeling problem.

  • Mechanism: It tokenizes time series values using scaling and quantization into a fixed vocabulary, then trains standard transformer architectures (based on T5) via cross-entropy loss.
  • Chronos-2: Expanded capabilities for univariate, multivariate, and covariate-informed forecasting entirely zero-shot via In-Context Learning (ICL).

Google TimesFM

TimesFM is Google’s 200M parameter zero-shot forecasting model.

  • Mechanism: Predicts future horizons via patch-based processing.
  • Features: Supports 16K context windows, quantile heads for probabilistic forecasting, and LoRA fine-tuning.

MOMENT

An open-source family of foundation models for general-purpose time series analysis, attempting to solve the massive data heterogeneity problem across diverse datasets.

TODO: Add patch-based forecasting mechanisms vs tokenization mechanisms.