chronos_t5_small_46m_forecasting

TS Arena wrapper for Amazon Chronos T5-Small time series forecasting model.

Model Description

Chronos is a family of pretrained time series forecasting models based on language model architectures. It tokenizes time series values using scaling and quantization, then uses a T5 model to generate probabilistic forecasts.

Attribute	Value
Parameters	46M
Architecture	T5 Encoder-Decoder
Original Repo	amazon/chronos-t5-small
Paper	Chronos: Learning the Language of Time Series
Task	Time Series Forecasting

Usage with TS Arena

import ts_arena

# Load model
model = ts_arena.load_model("chronos-t5-small")

# Generate forecasts
import numpy as np
context = np.random.randn(96)  # 96 timesteps of history
output = model.predict(context, prediction_length=24, num_samples=20)

# Access results
print(output.predictions.shape)      # Point forecasts (median)
print(output.quantiles[0.5].shape)   # Median forecast
print(output.quantiles[0.1].shape)   # 10th percentile
print(output.quantiles[0.9].shape)   # 90th percentile

Direct Usage with Chronos

from chronos import ChronosPipeline
import torch

pipeline = ChronosPipeline.from_pretrained(
    "amazon/chronos-t5-small",
    device_map="cuda",
    torch_dtype=torch.bfloat16,
)

context = torch.randn(1, 96)  # (batch, time)
forecast = pipeline.predict(context, prediction_length=24, num_samples=20)

Evaluation Results

ETTh1 Dataset (context=96, horizon=96)

Metric	Value
MSE	4.37
MAE	1.66
RMSE	2.09

Features

• Zero-shot forecasting: No training required
• Probabilistic forecasts: Returns samples and quantiles
• Variable horizons: Supports different prediction lengths
• Multivariate support: Processes each channel independently

Limitations

• Univariate model (multivariate handled channel-by-channel)
• No exogenous variable support
• Recommended max prediction length: 64

Citation

@article{ansari2024chronos,
  title={Chronos: Learning the Language of Time Series},
  author={Ansari, Abdul Fatir and Stella, Lorenzo and Turkmen, Caner and Zhang, Xiyuan and others},
  journal={arXiv preprint arXiv:2403.07815},
  year={2024}
}

License

Apache-2.0 (following the original Chronos license)

Links

• TS Arena GitHub
• Original Chronos
• Chronos Paper