University of Southern California logo

TEMPO

Model directory
Melady/TEMPO

80M params | Apache-2.0 | Upstream checkpoint

TEMPO is a GPT-based time-series foundation model from USC that takes an unusual route to zero-shot forecasting: instead of learning a bespoke time-series backbone, it adapts a pre-trained language model to the task. The result is an 80M-parameter forecaster that leans on the pattern-matching strength a GPT-2 backbone already carries.

Architecturally it pairs that GPT-2 backbone with trend-seasonal decomposition prompting. Input series are decomposed into trend, seasonal, and residual components, which lets the backbone reason about each structural piece separately before recombining them into a forecast. The official release is centered on the TEMPO-80M checkpoint and ships benchmark assets for ETT, electricity, traffic, and weather experiments on top of the GPT-2 base model, and it is released under Apache-2.0.

Where to run this model

Compute availability, hardware requirements, and pricing depend on the execution provider.

Hugging Face · Weights & deployment

Get model weights, deployment instructions, and current provider listings.

Hugging Face provider mappings checked 2026-10-06. No Inference Provider mapping was listed for this checkpoint.

Model Classification

Family

TEMPO

Type

time series foundation model

Pretrained time-series model for zero-shot or few-shot forecasting workloads.

Training Data

Official release is centered on the TEMPO-80M checkpoint and ships benchmark assets for ETT, electricity, traffic, and weather experiments on top of a GPT-2 base model.

Recommended For

  • • Zero-shot forecasting leveraging pre-trained language model representations
  • • Series with clear trend-seasonal structure where decomposition improves predictions

Strengths

  • • Decomposition-aware prompting lets the GPT backbone reason about trend, season, and residual separately
  • • Inherits the pattern-matching strength of the pre-trained GPT-2 backbone

Limitations

  • • GPT-2 backbone is larger than purpose-built time-series architectures at equivalent accuracy
  • • Decomposition approach assumes the input series has meaningful trend-seasonal structure
  • • Native forecast horizon is 96 steps; longer horizons require autoregressive rollout which may degrade quality

Capabilities

forecastingzero-shot

Tags

usctempogpt-baseddecompositionzero-shot

Specifications

Parameters
80M
Architecture
GPT-2 backbone with trend-seasonal decomposition prompting
Context length
n/a
Max context
n/a
Minimum history
n/a
Recommended history
n/a
Input step
n/a
Required target series
1
Temperature
Ignored
Top P
Ignored
Max output
n/a
License
Apache-2.0