Regent: A Differentiable Graph-Based Quant Trading Pipeline (VQ-VAE and GNNs)
Regent is a differentiable graph-based quantitative trading pipeline that combines a VQ-VAE (Vector Quantized Variational Autoencoder) and Graph Neural Networks (GNNs) to process and model financial market data for trading decisions.
Background
- Regent is an open-source quantitative trading system that combines **VQ-VAE** (Vector Quantized Variational Autoencoder) with **Graph Neural Networks**. It transforms market data into a graph where assets are nodes, then learns discrete "symbols" from price action to model market regimes and cross-asset relationships.
- **VQ-VAE** (pioneered by DeepMind) compresses continuous data into a finite codebook of discrete states. In trading, this aims to discover recurring patterns or regimes in price series.
- **GNNs** operate on a graph structure — nodes are assets (stocks, crypto), edges represent correlations. The GNN propagates information across the graph to make predictions.
- Why it matters: Most quant systems treat assets independently or use only raw time-series. Regent's differentiable pipeline captures **regime changes** and **cross-asset dependencies** in one end-to-end trained model — a frontier approach in "AI for systematic trading."
- This is an individual/small-team research project, not a commercial product. No live trading performance is provided; treat it as an experimental reference.