Temporal Attention augmented Bilinear Networks for Financial Time-series Data Analysis

https://ieeexplore.ieee.org/document/8476227

Short description: We proposed a bilinear structure equipped with an attention mechanism that can highlight the temporal importance of each temporal event in multivariate time-series. The proposed structure is validated in the problem of predicting mid-price movements (increasing, decreasing, stationary) using Limit Order Book information. Our empirical analysis showed that the proposed structures outperformed existing models, even LSTM, CNN while being relatively fast.

Source Code

Code is written in python 2.7.13 with the following dependencies:

• keras 2.1.2
• tensorflow 1.3

The naming convention follows our paper, examples how to use our code can be seen from the following files:

• example_bl.py (BL)
• example_tabl.py (TABL)

For more information, please contact thanh.tran@tuni.fi or viebboy@gmail.com

Citation

If you use TABL in your work, please cite the following paper:

@article{tran2018temporal,
  title={Temporal Attention-Augmented Bilinear Network for Financial Time-Series Data Analysis},
  author={Tran, Dat Thanh and Iosifidis, Alexandros and Kanniainen, Juho and Gabbouj, Moncef},
  journal={IEEE transactions on neural networks and learning systems},
  year={2018},
  publisher={IEEE}
}