NeuroLOB
Generative Market Intelligence using Neural Point Processes & Diffusion.
Abstract Financial microstructure data (Limit Order Books) is event-driven and irregular. NeuroLOB is a neuro-symbolic framework integrating Neural Point Processes (NPP) with Diffusion Models to generate high-fidelity synthetic market data.
System Architecture
To bring this research model to the web, I implemented a Hybrid Cloud Architecture:
- Frontend: Static portfolio hosted on GitHub Pages.
- Backend: Containerized inference engine hosted on Hugging Face Spaces.
- Model: The
DiffusionPipelineruns on a custom PyTorch backend.
Key Innovations
- Continuous Time Encoding: Implemented Rotary Positional Embeddings (RoPE) to inject continuous time information directly into the attention mechanism.
- Diffusion-Based Generation: Used a conditional DDPM to capture the multi-modal distribution of price returns.
- Physics-Informed Regularization: Enforced financial stylized facts (fat tails and volatility clustering) using a GARCH(1,1) post-processing layer.