This repository is the implementation for the Project: Transformer versus LSTMs for electronic trading.
The Implmentation of this project is based on a few open-source repositories. I declared them here and thanks for their valueable works.
This repository is built on the code base of Autoformer.
The implementation of Autoformer, Informer, Reformer, Transformer is from:
https://github.com/thuml/Autoformer
The implementation of FEDformer is from:
https://github.com/MAZiqing/FEDformer
The implementation of DeepLOB is based on:
https://github.com/zcakhaa/DeepLOB-Deep-Convolutional-Neural-Networks-for-Limit-Order-Books
The implementation of DLSTM is inspired by:
https://github.com/cure-lab/LTSF-Linear
This repository has three branches: The main branch, branch price, and branch price_diff. The detail of each branch is described below:
main: most of the contribution of this project is in the main branch. The task completed in the main branch is to predict the Limit Order Book (LOB) Mid-Price Movement (rise, stationary, fall), which is a classification task. In this task, A new model called DLSTM is developed for this task and outperforms other models. In addition, the architecture of Transformer-based models is adapted for the classification.
price: The task in the price branch is to compare the performance of Transformer-based models and LSTM-based models in predicting the absolute LOB Mid-Price. In this task, the implementation of models is from previous works.
price_diff: The task in the price_diff is to compare the performance of Transformer-based models and LSTM-based models in predicting the LOB Mid-Price difference (i.e. predict the difference between the future mid-price in timestamp t+k and mid-price in timestamp t). In this task, the implementation of models is from previous works.
It is recommended to run the code in a virtual environment. After initializing the virtual environment, install the requirement by:
pip install -r requirements.txtIf you want to download the data for a fast run, skip reading Data Pre-Processing part and directly go to the Section of Download Dataset. If you want to use your own dataset, jump to Section Other Files Description.
The LOB data collected by kdb is saved every day in .csv files. The data needs some pre-processing before use. The raw LOB data can be download from Google Drive. Use two jupyter notebooks in folder ./lob_data_process to finish pre-processing and labelling.
data_preprocess.ipynb: The file used to compress multiple LOB data into one file. You can compute and add new features as columns such as mid-price, log mid-price, and bid-ask imbalance.
generate_label.ipynb: you can generate labels for the dataset using the smoothing label method. Generating labels for 12 days datasets will cost around 40-80 minutes (depending on the computer's speed).
The datasets can be easily downloaded from Google Drive.
After downloading the dataset, run mkdir hy-tmp, unzip the dataset and put it in the /hy-tmp directory. Note: If you want to put the dataset under your own defined path, change parameter root_path inside the script files in ./scripts directory.
If you want to save time for training, jump to the Section Testing. To reproduce the result, experiment scripts are provided under ./scripts directory. The scripts are written in shell language. Make sure to run them in Linux or use the wsl subsystem in Windows. The training process of a model will take 30 mins/a half day/a whole day, which depends on the dataset size, computer's speed and the model choice. The example of running the script is provided below:
main:
chmod +x scripts/classification_script/DLSTM.sh
scripts/classification_script/DLSTM.sh
price :
chmod +x scripts/regression/Autoformer.sh
scripts/regression/Autoformer.sh
price_diff :
chmod +x scripts/Price_diff/LSTM.sh
scripts/Price_diff/LSTM.sh
Once starting training the model, a log file will be saved under the ./logs directory.
The model checkpoints of most models are saved. You can simply load the model from checkpoints to save training time. Models can be downloaded from Google Drive.
After downloading the checkpoints, unzip the models and put them under the /hy-tmp directory.
Note: If you want to put the models under your own defined path, change the parameter checkpoints inside the script files in ./scripts directory.
The testing process is similar to the training process, which is to run the shell scripts. An example of running a test script is given below:
main:
chmod +x scripts/classification_test_script/DLSTM.sh
scripts/classification_test_script/DLSTM.sh
price :
chmod +x scripts/regression_test/LSTM.sh
scripts/regression_test/LSTM.sh
price_diff :
chmod +x scripts/Price_diff_test/LSTM.sh
scripts/Price_diff_test/LSTM.sh
Once start the testing process, a log file will be saved under the ./logs directory.
run.py: The main python file executed by the shell script; you can define different parameters in this file and directly run python run.py to start the training/validation/testing process.
data_provider/data_loader.py : Change this file if you want to customize and use your own dataset.
exp/exp_main.py : This file controls the training/validation/testing logic.
layers: layers and components of Transformer-based models.
debug.ipynb: This notebook is used for debugging and parameter tuning. The model can be trained/validated/tested using this notebook instead of the shell scripts.