Description • Dependencies • Setup • Content • Results

Usually, when one deals with Limit Order Book (LOB) data, the common way is to treat it as time series or tabular data. This project investigates another possible approach, which is treatment of LOB data as long event sequences with two possible events (price going up or down), which allows one to use a rich mathematical apparatus developed for temporal point processes on LOB data.

We are going to consider three main models - Neural Hawkes Process (baseline), UNIPoint, and Transformer Hawkes Process. The aforementioned models will be implemented from scratch, adapted towards usage on LOB data, tuned and tested on our self-collected limit order book dataset consisting of five tokens - Ethereum (ETH), Litecoin (LTC), EOSIO (EOS), Ripple (XRP), Binance coin (BNB). We intentionally skip the most popular crypto asset - Bitcoin (BTC), because we are interested in robustness of the models when dealing with data coming from less liquid markets. In addition to that, we also perform an out-of-sample test on Stellar coin (XLM) to see the generalization capability of the models and how they would react to an unknown coin being introduced, i.e. whether there are general features in the LOB event sequences that could exploited, opening access for e.g. transfer learning in the future.

• Python 3
• Python data stack
• PyTorch
• tqdm

Clone GitHub repository:

git clone https://github.com/rodrigorivera/mds20_deepfolio

Make sure that all dependencies are installed and run the respective notebooks, containing the experiments for each model.

This repository contains the codebase for the project, its structure is the following:

• datasets folder contains the code that was used to download and transform raw data into LOB, as well as all the required preprocessing of the data;
• models folder contains codebase for the models used in this project, as well as detailed description of the models and implementation details;
• implementations folder contains existing implementations of the used models that we used as a reference / inspiration for our own versions, as well as detailed implementation descriptions and differences with our own versions;
• images folder contains images used in the readmes.

Models were trained on the combined dataset, which was composed of sequences of all cryptocurrencies (ETH, EOS, LTC, BNB, XRP) with lengths 3000, and then tested separatly on each cryptocurrency.

Results for NHP model (Log-Likelihood loss and time / event prediction from probability density):

Results for NHP+ model (Two linear layers for event prediction and Log-Likelihood loss + time loss + event loss):

Results for UNIPoint model:

Results for THP model:

Mean scores of all models on all crypto datasets:

Results of out-of-sample testing, performed on Stellar (XLM) coin, which was not at all present among the training samples