📈 Multi-Horizon Volatility Forecasting using ODEs and Cross-Stitch Networks

📍 EPFL – Master in Data Science, Year 1 (2024)
👥 Team: Matthias Wyss, William Jallot, Thierry Sokhn
🔗 Code Repository: GitHub
📄 Final Report: Report

---

We developed a hybrid neural architecture for predicting multi-horizon volatility in high-frequency trading using the FI-2020 dataset. The architecture combines ODE networks and Cross-Stitch networks to model complex temporal dynamics in financial data.

We implemented and compared two key models:

• ODE + Cross-Stitch Network – Leverages continuous-time modeling and multi-source learning.
• Temporal Fusion Transformer (TFT) – A state-of-the-art forecasting model handling long-range dependencies.

Evaluated using Mean Relative Absolute Error (MRAE), our ODE-based Cross-Stitch model outperformed TFT, demonstrating its effectiveness in capturing the volatility patterns of high-frequency trading data.

---

🛠 Tools & Libraries:

• Python
• TensorFlow
• Keras

🧠 Techniques:

• Time Series Forecasting
• ODE Networks
• Cross-Stitch Networks
• Temporal Fusion Transformer (TFT)
• Financial Modeling