Skip to main content

​
Overview

Chronos 2 is the latest version of the Chronos family, featuring an improved architecture with better handling of long sequences and enhanced probabilistic forecasting capabilities. Repository: amazon/chronos-2

​
Loading the Model

from samay.model import Chronos_2_Model

repo = "amazon/chronos-2"
model = Chronos_2_Model(repo=repo)

​
Loading Dataset

from samay.dataset import Chronos_2_Dataset

train_dataset = Chronos_2_Dataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="train",
    batch_size=16
)

val_dataset = Chronos_2_Dataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="test",
    batch_size=16
)

​
Zero-Shot Forecasting

metrics = model.evaluate(val_dataset, metric_only=True)
print(metrics)
# {'mse': ..., 'mae': ..., 'mase': ..., 'rmse': ..., 'crps': ...}

​
Fine-tuning

model.finetune(
    train_dataset,
    epoch=5  # Number of training epochs
)

# Evaluate after fine-tuning
metrics = model.evaluate(val_dataset, metric_only=True)

​
Visualization

model.plot(val_dataset)

​
Key Improvements over Chronos 1

  1. Better Long Sequence Handling: Improved architecture for longer context windows
  2. Enhanced Quantile Predictions: More accurate uncertainty estimates
  3. Autoregressive Forecasting: Better handling of very long forecast horizons
  4. Improved Training: More stable and efficient training process

​
Autoregressive Forecasting

Chronos 2 can handle very long forecast horizons through autoregressive prediction: 
# The model automatically uses autoregressive forecasting
# when horizon_len exceeds max_patches * patch_size

output = model.model(
    context=input_seq,
    num_output_patches=current_horizon_patch
)
The model:
  1. Predicts up to max_patches at a time
  2. Uses median predictions as input for next iteration
  3. Concatenates predictions for full horizon

​
Architecture Details

# Model configuration details
max_patches = model.model.chronos_config.max_output_patches
patch_size = model.model.chronos_config.output_patch_size
quantiles = model.model.chronos_config.quantiles

​
Handling Different Horizon Lengths

# For horizons within max_patches * patch_size
# Direct prediction in one step
output = model.model(
    context=context,
    num_output_patches=num_patches,
    future_target=target  # For training
)

​
Probabilistic Forecasting

Chronos 2 provides comprehensive quantile forecasts: 
# During evaluation, returns full quantile distribution
metrics, trues, preds, histories, quantiles = model.evaluate(
    val_dataset,
    metric_only=False
)

# quantiles shape: (num_quantiles, batch_size, channels, horizon)

​
Available Metrics

Chronos 2 supports comprehensive evaluation:
  • MSE: Mean Squared Error
  • MAE: Mean Absolute Error
  • MASE: Mean Absolute Scaled Error
  • MAPE: Mean Absolute Percentage Error
  • RMSE: Root Mean Squared Error
  • NRMSE: Normalized RMSE
  • SMAPE: Symmetric MAPE
  • MSIS: Mean Scaled Interval Score
  • ND: Normalized Deviation
  • MWSQ: Mean Weighted Sum of Quantiles
  • CRPS: Continuous Ranked Probability Score

​
When to Use Chronos 2

  • When you need the latest improvements from Amazon Research
  • For very long forecast horizons
  • When uncertainty quantification is critical
  • For production deployments requiring robust performance
  • When you need state-of-the-art accuracy

​
Example Notebook

For a complete working example, see:
Chronos 2 represents the state-of-the-art in the Chronos model family, with significant improvements in handling diverse forecasting scenarios.

Build docs developers (and LLMs) love

Get started for freeTalk to us