DeepFolio | Diffusion-Transformer (DiT) for Portfolio & Execution Optimization

DeepFolio is an OpenAI Sora-inspired Diffusion-Transformer (DiT) framework for joint portfolio optimization and best execution, designed to maximize Sharpe ratio without explicit return forecasts. It leverages:

• Transformer to capture asset dependencies and encode market conditions.
• Diffusion Models to filter market noise and generate both robust allocation weights and optimized trading trajectories.
• End-to-End Strategy Execution to reduce information loss between strategy design and execution implementation, ensuring optimal real-world performance.

---

🚀 Key Features

✅ Unified Portfolio & Execution Optimization – Bridges the gap between portfolio construction and trade execution.
✅ Diffusion-Based Portfolio Generation – Generates adaptive, robust asset allocations without relying on explicit return forecasts.
✅ Market-Aware Execution Path Modeling – Uses Diffusion Models to optimize execution trajectories, reducing slippage and market impact.
✅ Scenario-Based Adaptation – Dynamically adjusts strategies for high/low volatility regimes, liquidity shifts, and market anomalies.
✅ Transaction Cost-Aware Optimization – Integrates TCA (Transaction Cost Analysis) into optimization, minimizing execution costs.

---

📜 Architecture

DeepFolio consists of two core modules:

1️⃣ Portfolio Optimization (Transformer + Diffusion)

• Transformer Encoder extracts asset relationships, learning market structure.
• Diffusion Model generates optimal portfolio weights, ensuring robustness under different conditions.

2️⃣ Execution Optimization (Trade Path Diffusion)

• Transformer encodes market microstructure (LOB, liquidity, volatility).
• Diffusion Model optimizes execution paths to minimize market impact and transaction costs.

📌 Pipeline Overview:

Documentation

For detailed documentation, please visit our documentation site.

Contributing

We welcome contributions! Please see our contributing guidelines for more details.

License

This project is licensed under the BSD-2-Clause License- see the LICENSE file for details.

Reference

[1] Damian Kisiel, Denise Gorse (2022). Portfolio Transformer for Attention-Based Asset Allocation arXiv:2206.03246 [q-fin.PM]

Acknowledgments

• This package leverages the power of TensorFlow for efficient portfolio optimization.
• Thanks to the financial machine learning community for inspiring many of the implemented methods.