Time Series Foundation Model Study

This repository contains source codes used for Sun, A.Y. and Sun, A.A, Zero-shot Forecasting of Streamflow Using Time Series Foundation Models: Are We There Yet? IOP Machine Learning: Earth, 2026

Univeriate LSTM Experiments

1. 1d-ahead forecast

• lstm_camelsdaily.py: main code for training 1d-ahead LSTM model
• camels_dataloader_daily.py: daily data loader file
• lstm.py: LSTM model
• config_lstm_camelsdaily.yaml: YAML configuration file for 1d-ahead training

2. 3H-ahead forecast

• lstm_main.py: main code for training 3H-ahead LSTM model
• camels_dataloader.py: 3H data loader file
• lstm.py: LSTM model
• config_lstm_camels.yaml: YAML configuratoin file for 3H-ahead training

Univariate TSFM Experiments

1. 1d-ahead forecast

• tsfm/1d/sundial_camels_daily.py: main code for 1D-ahead univariate zero-shot benchmarking using Sundial, Chronos, TTM, and MOIRAI
• tsfm/1d/camels_dataloader_daily.py: data loader routines
• tsfm/1d/config_sundial_camelsdaily.yaml: configuration file

2. 3H-ahead forecast

• tsfm/3h/sundial_camels.py: main code for 3H-ahead univariate zero-shot benchmarking using Sundial, Chronos, TTM, and MOIRAI
• tsfm/3h/camels_dataloader.py: data loader routines
• tsfm/3h/config_sundail_camels.yaml: configuration file

Dependencies:

• pytorch, 2.3.1
• python, 3.10.0
• diffusers, 0.29.2
• accelerate, 0.31.0
• omegaconf, 2.3.0
• hydrostat, 1.0.0

The conda environments we used for benchmarking the TSFMs can be found under conda_env folder

• sundial_env.yaml: for Sundial TSFM
• ibm_ttm.yaml: for IBM's TTM model
• chronos.yaml: for Amazon's Chronos model
• moirai_env.yaml: for SalesForce MOIRAI model

Note: the original CAMELS data can be here We referenced the github code repo of Kratzert et al, 2019, WRR, Toward Improved Predictions in Ungauged Basins: Exploiting the Power of Machine Learning to form some of the daily dataset.