🌟 Robust QD Portfolio Ensemble Optimization 🌟

This repository provides a Streamlit app implementation for Robust Quality-Diversity Portfolio, a method co-optimizing thousands of sub-portfolios and assembles them to maximize out-of-sample robustness. The method combines cutting-edge techniques such as Generative Meta-Learning, Monte Carlo Optimization through DropBlock, and sparse portfolio construction using the SparseMax layer, scaling to thousands of assets.

Key Highlights

• Quality-Diversity Optimization:
  • Behavioral diversity is measured using return correlations across co-optimized sub-portfolios.
• Scalability:
  • Efficiently scales to thousands of assets, surpassing traditional optimization methods.
• SparseMax Layer:
  • Ensures sparse portfolio construction, allowing for reduced complexity and improved interpretability.
• Generative Meta-Learning:
  • Invented for non-convex population-based optimization in high-dimensional spaces.
  • Excluded from here for simplicity, refer to https://github.com/kayuksel/generative-opt
• Monte Carlo Optimization:
  • Uses DropBlock for enhanced robustness against noise.
• Parallelization:
  • Fully parallelized with PyTorch for large-scale optimization.

🚀 Quick Start Guide

Follow the steps below to get the Streamlit app up and running:

git clone https://github.com/your-repo/robust-qd-portfolio.git
cd robust-qd-portfolio
pip install -r requirements.txt
streamlit run app.py --server.maxUploadSize 1000

Multi-Objective Optimization

The primary optimization target is based on the multiplication of the AlphaSharpe metric and the Omega Ratio (OMG), which balances portfolio performance and robustness against extreme returns.

Secondary Objectives

• Maximum Correlation Penalty: Ensures sub-portfolios within the population are diversified by reducing inter-correlation.
• Eigen Entropy Regularization (optional): Encourages further out-of-sample robustness of the individual sub-portfolios.

Modifying the Reward Function

You can modify the primary objective (alpha_sharpe(rets) * omg) to incorporate new performance metrics. However, it is critical to retain the secondary objective, which ensures robustness and diversification.

Additional Tips and Details

• Data Preprocessing: In app.py, the log-returns are first denoised. You can comment-out that part of the code if you wish to use raw data.

• Large Asset Universe: This method is best with thousands of assets. You can reduce the number of iterations to obtain a less sparse portfolio.

• Index Relative Reward Calculation: The reward is calculated relative to a selected index. To disable this, add a fake index column with static log returns.

• Performance Over Iterations:

  • This video shows the mean-portfolio returns (blue) versus the index (red) and behavioral diversity (heatmap) over iterations.
  • The app does not include monitoring over iterations for speed, please watch the video to understand what is happening behind.

Published in GECCO 2023 🎉

📄 Read the paper: https://arxiv.org/abs/2307.07811