iAmTime-base-forecast

iAmTime-base-forecast is the forecasting checkpoint of iAmTime, a foundation model for instruction-conditioned in-context time-series tasks. It produces probabilistic forecasts from historical observations, with optional covariates and input-output demonstrations that provide additional context at inference time, without updating model parameters.

iAmTime combines a Hierarchical Multi-Scope Transformer Encoder with a Task-Conditioned Patch Decoder. Structured semantic tokens gather information from designated time-series regions, exchange information across demonstrations, and condition the query representation for prediction.

Code and Usage · Paper

Model Details

Property	Description
Model variant	iAmTime-base-forecast
Task	Time-series forecasting
Framework	PyTorch
Authors	Anish Saha and Konstantin Shmakov
License	Apache 2.0

This repository provides the forecasting checkpoint. The paper describes the broader iAmTime framework and its applications to additional time-series tasks.

Usage

See the iAmTime GitHub repository for installation instructions, model loading, input preparation, and inference examples.

Use the model implementation and preprocessing provided by the repository with this checkpoint's config.json and model.safetensors files.

Model Inputs

A forecasting query consists of:

• Historical target observations: Past values of the time series to forecast, with one or more target variables.
• Forecast horizon: The number of future time steps to predict.
• Optional historical covariates: Additional variables observed alongside the target history that help explain its behavior.
• Optional future-known covariates: Variables available in advance over the forecast horizon, such as calendar features or planned events. Future target values are not required for the query.
• Optional in-context demonstrations: Example input-output pairs containing historical observations and their corresponding forecast-period target values, with covariates where applicable. These demonstrations provide evidence about the relationship to apply to the query at inference time.

Covariates should align with the corresponding historical or future time steps. Follow the GitHub examples for the exact input format, batching, and missing-value handling.

Model Outputs

The model predicts target values over the requested forecast horizon. Outputs are quantile forecasts: multiple predicted values at each future time step and for each target variable, corresponding to the quantile levels configured for the checkpoint.

The median quantile provides a central point forecast, while lower and upper quantiles describe predictive uncertainty and can be used to construct prediction intervals. These intervals reflect the model's estimated uncertainty; their empirical coverage can vary across datasets and forecasting conditions.

See the GitHub usage examples for the returned output structure and how to extract point forecasts and quantiles.

Intended Use and Limitations

This checkpoint is intended for time-series forecasting research and downstream forecasting applications. Evaluate accuracy and uncertainty estimates on representative data before deployment, particularly when the target domain differs from the model's training distribution.

Citation

If you use iAmTime in your work, please cite:

@article{Saha2026-vb,
  title         = {A foundation model for instruction-conditioned in-context time series tasks},
  author        = {Saha, Anish and Shmakov, Konstantin},
  journal       = {arXiv preprint arXiv:2603.22586},
  year          = {2026},
  month         = mar,
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  eprint        = {2603.22586},
  url           = {https://arxiv.org/abs/2603.22586}
}

License

This model is released under the Apache License 2.0.