🚀 CryptoMamba: Leveraging State Space Models for Accurate Bitcoin Price Prediction

📣 Announcements
CryptoMamba paper has been published at IEEE International Conference on Blockchain and Cryptocurrency (ICBC) 2025! 🎉
We presented CryptoMamba at the Advances in Financial AI Workshop @ ICLR 2025.

CryptoMamba is a novel Mamba-based architecture designed for accurate and efficient time-series forecasting, with a focus on cryptocurrency price prediction. Leveraging the capabilities of SSMs, CryptoMamba excels at capturing long-range dependencies and adapting to highly volatile market conditions.

Repository Includes:

• Implementation of CryptoMamba and baseline models (LSTM, Bi-LSTM, GRU, and S-Mamba).
• Two Trading Algorithms: Vanilla and Smart, for evaluating real-world performance.
• Code for data preprocessing, model training, evaluation metrics, and trading simulations.

CryptoMamba’s robust performance and generalizability make it a promising solution for various sequential forecasting tasks, including financial markets, commodities, and other time-series applications.

📖 Table of Contents

• Requirements
• Usage
  • Data
  • Config
  • Train New Model
  • Evaluate Model
  • Predict Next Day Price
• Results
• Citation
• Where to Ask for Help

🔧 Requirements

To install the requirements, you can use:

pip install -r requirements.txt

If you have difficulties installing mamba_ssm, please follow the instructions in its GitHub repository.

💡 Usage

Data

You can find the processed data that we use in here. If you want to use another data configuration, you should change the configuration in the data config file. Note that the data_path should point to the raw data file with a similar format to this.

Config

If you want to use additional features other than Open, Close, High, Low, Timestamp, and Volume, you should specify a list called additional_features in your data and training configuration files. Note that your raw data file should have these features.
If you want to change the time resolution, you should change the date_format in your data configuration file and also set jumps to your desired resolution in seconds. Note that your raw data dates and your start and end dates in the data configuration should follow the new date format.

Train New Model

To train a model, use the following:

python3 scripts/training.py --config CONFIG_NAME

Here, CONFIG_NAME is the name of a config file in the training config folder without its extension. For example, to train CryptoMamba with volume you can run the following command:

python3 scripts/training.py --config cmamba_v

Evaluate Model

To evaluate a model, run this command:

python scripts/evaluation.py --config CONFIG_NAME --ckpt_path PATH_TO_CHECKPOINT

To run the trading simulation, you can use the following command:

python scripts/simulate_trade.py --config CONFIG_NAME --ckpt_path PATH_TO_CHECKPOINT --split SPLIT_NAME --trade_mode TRADE_ALGORITHM

Where SPLIT_NAME is train, val, or test and TRADE_ALGORITHM is smart or vanilla or smart_w_short.

Predict Next Day Price

We also have a script to predict the next day's price and trade suggestions by providing the prices of its previous days. You have to create a csv file similar to [this](data/one_day_trade.csv) and use this command:

python scripts/one_day_pred.py --config CONFIG_NAME --ckpt_path PATH_TO_CHECKPOINT --date DATE

If you don't provide a value for DATE, the code automatically predicts one day after the last day that is in the file.

📈 Results

Model	RMSE	MAPE	MAE	Parameters
LSTM	2672.7	3.609	2094.3	204K
Bi-LSTM	2325.6	3.072	1778.8	569k
GRU	1892.4	2.385	1371.2	153k
iTransformer	1826.9	2.4260	1334.3	201k
S-Mamba	1717.4	2.248	1239.9	330k
CryptoMamba	1713.0	2.171	1200.9	136k
LSTM-v	2202.1	2.896	1668.9	204K
Bi-LSTM-v	2080.2	2.738	1562.5	569k
GRU-v	1978.0	2.526	1454.3	153k
iTransformer-v	1779.9	2.427	1322.1	201k
S-Mamba-v	1651.6	2.215	1209.7	330k
CryptoMamba-v	1598.1	2.034	1120.7	136k

🎯 Citation

If you use CryptoMamba in a research paper, please cite our paper:

@article{Sepehri2025CryptoMamba,
      title={CryptoMamba: Leveraging State Space Models for Accurate Bitcoin Price Prediction}, 
      author={Mohammad Shahab Sepehri and Asal Mehradfar and Mahdi Soltanolkotabi and Salman Avestimehr},
      year={2025},
      url={https://arxiv.org/abs/2501.01010}
}

❓ Where to Ask for Help

If you have any questions, feel free to open a Discussion and ask your question. You can also email sepehri@usc.edu (Mohammad Shahab Sepehri) or mehradfa@usc.edu (Asal Mehradfar).