PriorLabs/tabpfn-time-series

Zero-shot Time Series Forecasting with TabPFN (work accepted at NeurIPS 2024 TRL and TSALM workshops)

What it solves

TabPFN-TS addresses the challenge of time series forecasting by enabling zero-shot predictions without the need for time-consuming model training or fine-tuning. It allows users to generate both point and probabilistic forecasts quickly, making it highly accessible for rapid experimentation.

How it works

The project frames univariate time series forecasting as a tabular regression problem. It follows a four-step process:

  1. Transforms a time series into a table format.
  2. Extracts and adds temporal features (such as calendar and seasonal features) to the table.
  3. Performs regression on the table using TabPFN, a tabular foundation model.
  4. Uses the regression results as the final time series forecasting outputs.

It supports target history and known-future covariates (like holidays or scheduled events), though it currently drops past dynamic and static covariates.

Who it’s for

It is designed for data scientists and researchers who need fast, accurate forecasting without the overhead of training a model from scratch, as well as those who want to perform probabilistic forecasting with minimal setup.

Highlights

  • Zero-shot capability: Provides immediate forecasting without requiring training.
  • Probabilistic forecasting: Supports both point and probabilistic output.
  • Known-future covariate support: Can integrate external data like holidays or weather forecasts if future values are known.
  • Cloud-ready inference: Integrates with tabpfn-client, allowing users to run fast inference without needing their own GPU.

Related

  • Project
    PriorLabs/TabPFN

    A tabular foundation model that provides high-accuracy classification and regression in a single forward pass, removing the need for iterative training and manual preprocessing.

  • Project
    google-research/tabfm

    A scikit-learn compatible tabular foundation model that enables zero-shot classification and regression on mixed-type tabular datasets using in-context learning.

  • Project
    Nixtla/nixtla

    A foundation model for time series forecasting and anomaly detection that provides zero-shot inference and fine-tuning capabilities across multiple domains.

  • Project
    DataDog/toto

    A foundation model for multivariate time series forecasting optimized for observability metrics, providing zero-shot and probabilistic predictions.

  • Project
    amazon-science/chronos-forecasting

    A family of pretrained foundation models for zero-shot time series forecasting that treats time series data as tokens for language model architectures.