CS-433 Machine Learning - Project 2 Report - Multi-Horizon Volatility Forecasting using Ordinary Differential Equation and Cross-Stitch Networks

The aim of this project is to introduce a novel architecture for predicting the volatility of Limit Order Books using the FI-2020 Dataset. Our approach combines the strengths of Ordinary Differential Equation (ODE) networks and Cross-Stitch networks to address the challenges of multi-horizon forecasting in high-frequency trading environments. To evaluate the effectiveness of our approach, we compare it against the Temporal Fusion Transformers (TFT) architecture, a state-of-the-art method for temporal forecasting tasks.

Our group is composed of :

• William Jallot 341540
• Thierry Sokhn 345880
• Matthias Wyss 329884

Installation

conda create --name <env_name> python=3.11.9
conda activate <env_name>
pip install -r requirements.txt

The requirements.txt has been generated from an Apple Intel computer.

The run.ipynb file contains the final run code for the two models we obtained.

The data should be inside the ./data folder (if it does not exist create it, using the following link: https://drive.google.com/file/d/1OFdK_Ya_wJEgCKqAaF5RxyR9V8yOX0IS/view?usp=drive_link). This data has been fetched and selected from https://etsin.fairdata.fi/dataset/73eb48d7-4dbc-4a10-a52a-da745b47a649 which are the first publicly available datasets that contain representations and annotations for a limit order book in the High Frequency universe.

Implemented Models:

• Cross-Stitch Network using Ordinary Differential Equations
• Temporal Fusion Transformer

All the models are being run inside the run.ipynb file.

Project structure

.
│   .gitignore
│   run.ipynb
│   README.md
│   requirements.txt
│
├───checkpoint
├───data
│       Test_Dst_NoAuction_DecPre_CF_7.txt
│       Test_Dst_NoAuction_DecPre_CF_8.txt
│       Test_Dst_NoAuction_DecPre_CF_9.txt
│       Train_Dst_NoAuction_DecPre_CF_7.txt
│
├───Cross_Stitch_Models
│   │   cross_stitch_network.py
│   │   inception.py
│   │   linear_cross_stitch_unit.py
│   │   loop_layer.py
│   │   odefunc.py
│   │   ode_layer.py
│   │   reshape.py
│
├───Temporal_Fusion_Transform
│   │   README.md
│   │   tft_model.py
│   │
│   ├───data_formatters
│   │   │   base.py
│   │   │   ts_dataset.py
│   │   │   utils2.py
│   │   │   volatility.py
│   │   │   __init__.py
│   │
│   └───expt_settings
│       │   configs.py
│       │   __init__.py
│
└───utils
    │   dynamic_losses.py

The project structure includes a folder Cross_Stitch_Models containing our Cross-Stitch ODE network with the different layers, while our folder Temporal Fusion Transform implements our temporal fusion transformer.

Useful functions implemented

• What was handled during the data cleaning / feature processing:

  • Midprice computation
  • Stocks splitting
  • Log returns computation of the volatility
  • Standardization (z-score).
  • Predictions upscaling (reverted when computing performance metrics)

• What has been used for evaluation metrics

  • Mean Relative Absolute Error