Lucifer744/KiT_model

KiT Model Checkpoints

Pre-trained model checkpoints for KiT: A Foundation Model for Candlestick Time-Series Forecasting via Diffusion Transformers.

Available Checkpoints

1. step_market_284M.pt

This is the official open-source checkpoint used in the paper for all reported results and evaluations.

• Parameters: 284M
• Training data: Up to 2026-01-01 (with validation split)
• Status: Evaluated on validation set (2026-01-01 to 2026-04-10)
• Use: For reproducing paper results and benchmarking

2. step_market_284M_train260410.pt

• Parameters: 284M
• Training data: Full dataset including all data up to 2026-04-10
• Status: No metrics on test sets; provided for practical use
• Use: For inference on future dates with potentially stronger predictive capability

⚠️ Note: This model has not been evaluated on any test set. It is provided to offer a version trained on the maximum available data. In theory, this model should have stronger predictive power on future dates compared to step_market_284M.pt.

Usage

See the code repository for complete training and inference examples.

Data

Demo data is available at: kit_data_demo

Citation

@article{kit2026,
  title={KiT: A Foundation Model for Candlestick Time-Series Forecasting via Diffusion Transformers},
  author={Boyu Zhang, Haorui Li},
  journal={arXiv preprint arXiv:2609.34507},
  year={2026}
}

License

See the main repository for license information.