TabPFN-TS

Zero-Shot Time Series Forecasting with TabPFN

📌 News

• 15-09-2026: 🚀 v1.3.0 — defaults now use TabPFN-3.5 in both LOCAL and CLIENT mode. Requires tabpfn>=9.0.0 and tabpfn-client>=0.5.3.
• 12-05-2026: 🚀 v1.1.0 — defaults now ship the TabPFN-TS-3 finetuned checkpoint from the TabPFN-3 report for SOTA forecasting out of the box. Requires tabpfn>=8.0.0.
• 27-05-2025: 📝 New paper version and v1.0.0 release! Strong GIFT-EVAL results, new AutoSeasonalFeatures, improved CalendarFeatures.
• 27-01-2025: 🚀 Ranked 1st on GIFT-EVAL benchmark^[1]!
• 10-10-2024: 🚀 TabPFN-TS paper accepted to NeurIPS 2024 TRL and TSALM workshops!

[1] Last checked on: 10/03/2025

✨ Introduction

We demonstrate that the tabular foundation model TabPFN, combined with lightweight feature engineering, enables zero-shot time series forecasting for both point and probabilistic tasks. On the GIFT-EVAL benchmark, our method achieves performance on par with top-tier models across both evaluation metrics.

As of v1.3.0 the package uses TabPFN-3.5 by default — a tabular foundation model pretrained purely on synthetic data. The finetuned TabPFN-TS-3 checkpoint from the TabPFN-3 report stays available.

📖 How does it work?

Our work proposes to frame univariate time series forecasting as a tabular regression problem.

Concretely, we:

1. Transform a time series into a table
2. Extract features and add them to the table
3. Perform regression on the table using TabPFN
4. Use regression results as time series forecasting outputs

For more details, see our paper and the TabPFN-3 report.

🧮 Covariate model — what's used vs. what's not

TabPFN-TS treats forecasting as univariate, target-history + known-future covariate regression. Concretely:

If your task relies on past_dynamic or static_columns, performance may be lower than dedicated multivariate / static-aware forecasters.

Known-covariate lags. KnownCovariateLagFeature, first in TABPFN_TS_DEFAULT_FEATURES, adds three columns per numeric or boolean known covariate: its lag 1, its lag at the season length, and its trailing mean over one season. The season comes from the timestamp frequency. The features are computed per series over context + horizon, so a horizon row only reads covariate values (never the target). String or categorical covariates get no extra columns. On the 25 fev-bench tasks with numeric known covariates this lowers SQL by 4.5% on the default v3.5 model (21/25 tasks better, worst task +1.5%), with the largest gains on temperature-driven load forecasting. To turn it off, pass temporal_features= a list without it to TabPFNTSPipeline.

👉 Why give us a try?

• Zero-shot forecasting: this method is extremely fast and requires no training, making it highly accessible for experimenting with your own problems.
• Point and probabilistic forecasting: it provides accurate point forecasts as well as probabilistic forecasts.
• Known-future covariate support: pass covariates whose future values are known at forecast time (calendar, holidays, scheduled events, demand or weather forecasts) directly alongside the target — see Covariate model for details on what's supported.

On top of that, thanks to tabpfn-client from Prior Labs, you won't even need your own GPU to run fast inference with TabPFN. 😉 We have included tabpfn-client as the default engine in our implementation.

⚙️ Installation

You can install the package via pip:

pip install tabpfn-time-series

Checkpoint

Both modes pin TabPFN-3.5. In LOCAL mode tabpfn downloads its default checkpoint automatically on first init — accept the license at ux.priorlabs.ai first. In CLIENT mode the cloud API serves its v3.5_default model.

For Developers

To install the package in editable mode with all development dependencies, run the following command in your terminal:

pip install -e ".[dev]"
# or with uv
uv pip install -e ".[dev]"

🚀 Getting Started

from tabpfn_time_series import TabPFNTSPipeline

pipeline = TabPFNTSPipeline()  # uses the cloud client by default — no GPU needed
predictions = pipeline.predict_df(context_df, prediction_length=24)

For full walkthroughs:

More examples live in the examples/ directory.

📚 Citation

If TabPFN-TS is useful in your research, please cite the paper (arXiv:2501.02945) and the TabPFN-3 report for the TabPFN-TS-3 checkpoint.

📊 Anonymous Telemetry

This project collects anonymous usage telemetry by default.

The data is used exclusively to help us understand how the library is being used and to guide future improvements.

• No personal data is collected
• No code, model inputs, or outputs are ever sent
• Data is strictly anonymous and cannot be linked to individuals

What we collect

We only collect high-level, non-identifying information such as:

• Package version
• Python version
• How often fit and inference are called, including simple metadata like the dimensionality of the input and the type of task (e.g., classification vs. regression) (⚠️ never the data itself)

See the Telemetry documentation for the full details of events and metadata.

This data is processed in compliance with the General Data Protection Regulation (GDPR) principles of data minimization and purpose limitation.

For more details, please see our Privacy Policy.

How to opt out

If you prefer not to send telemetry, you can disable it by setting the following environment variable:

export TABPFN_DISABLE_TELEMETRY=1