Time Series Forecasting

TabPFN extends beyond tabular and regression tasks—it can perform zero-shot time series forecasting by reframing forecasting as a tabular regression problem. The TabPFN-TS workflow combines the TabPFN model with lightweight feature engineering to deliver both point and probabilistic forecasts.

Getting Started

The following example shows how to forecast a time series dataset (with no training or GPU required) by loading data, automatically extracting features, and generating both point and probabilistic forecasts.

# Import all required components for data handling, feature engineering, and prediction
from datasets import load_dataset
from tabpfn_time_series import TimeSeriesDataFrame, FeatureTransformer, TabPFNTimeSeriesPredictor, TabPFNMode
from tabpfn_time_series.data_preparation import to_gluonts_univariate, generate_test_X
from tabpfn_time_series.features import RunningIndexFeature, CalendarFeature, AutoSeasonalFeature

# Choose dataset and define forecast horizon
ds, pred_len = "monash_tourism_monthly", 24

# Load dataset from Hugging Face and convert to a standard TimeSeriesDataFrame
raw = load_dataset("autogluon/chronos_datasets", ds)
tsdf = TimeSeriesDataFrame(to_gluonts_univariate(raw["train"]))

# Keep only the first 2 time series to keep the example lightweight
tsdf = tsdf[tsdf.index.get_level_values("item_id").isin(tsdf.item_ids[:2])]

# Split into training and test portions based on forecast horizon
train, _ = tsdf.train_test_split(prediction_length=pred_len)
test = generate_test_X(train, pred_len)

# Add useful temporal and seasonal features
features = [RunningIndexFeature(), CalendarFeature(), AutoSeasonalFeature()]
train, test = FeatureTransformer(features).transform(train, test)

# Initialize the predictor (client mode uses pretrained TabPFN weights) and generate forecasts
predictor = TabPFNTimeSeriesPredictor(tabpfn_mode=TabPFNMode.CLIENT)
pred = predictor.predict(train, test)

Key Features

Zero-Shot Forecasting: Generate forecasts such as mean and median values with no task-specific training required when using the pre-trained weights.
Probabilistic Forecasting: Natively produces quantile forecasts and confidence intervals for uncertainty quantification.
Advanced Data Handling: Integrates exogenous variables (external regressors), multiple time series, and missing data, while handling temporal patterns like trends and seasonality.
Lightweight & Fast: Uses a compact <20M parameter model (about 40× smaller than some foundation models), enabling short inference times on a local CPU or via the tabpfn-client with no GPU required.

You can find the TabPFN Time Series source code and latest updates on GitHub.

Google Colab Example

Check out our Google Colab for a detailed tutorial on using TabPFN for time-series forecasting.

Getting Started

API

Capabilities

Agentic

Integrations

Troubleshooting

Use Cases

Changelog

Time Series Forecasting

Getting Started

Key Features

Google Colab Example

​Getting Started

​Key Features

Google Colab Example

Getting Started

Key Features