# Large-scale Time-Varying Portfolio Optimisation using Graph Attention Networks

**Authors:** Kamesh Korangi, Christophe Mues, Cristián Bravo  
**arXiv:** [2407.15532](https://arxiv.org/abs/2407.15532)  
**Date:** 2024-07-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.15532v2/

## Key points

- Authors Kamesh Korangi, Christophe Mues, and Cristián Bravo study large-scale time-varying portfolio optimization using Graph Attention Networks (GATs).
- Traditional mean-variance portfolios based on Markowitz principles have limitations in capturing complex interdependencies between asset values.
- The study incorporates firms at risk of default and those dropping off indices over time into portfolio optimization on a large scale.
- Researchers propose and test a novel method utilizing GATs, a subclass of Graph Neural Networks known for uncovering nonlinear relationships in network data.
- 30 years of data on mid-cap firms are used to create graphs with distance correlation and the Triangulated Maximally Filtered Graph approach as inputs to the GAT model.
- The model includes weight and allocation constraints, along with a loss function derived from the Sharpe ratio to maximize risk-adjusted returns.
- Results show that the GAT-based portfolio consistently outperforms benchmark portfolios over an extended period, providing valuable insights into market dynamics.
- This innovative approach demonstrates the potential of using Graph Attention networks for enhancing portfolio optimization strategies in financial markets.

## Summary

In their study "Large-scale Time-Varying Portfolio Optimisation using Graph Attention Networks," authors Kamesh Korangi, Christophe Mues, and Cristián Bravo explore the collective performance of a set of firms as a portfolio in financial market analysis. They highlight the limitations of traditional mean-variance portfolios based on Markowitz principles in capturing complex interdependencies between asset values. This study breaks new ground by incorporating firms at risk of default and those that drop off indices over time into portfolio optimization on a large scale. The researchers propose and test a novel method that utilizes Graph Attention networks (GATs), a subclass of Graph Neural Networks (GNNs), known for their ability to uncover nonlinear relationships in network data. The study uses 30 years of data on mid-cap firms to create graphs using distance correlation and the Triangulated Maximally Filtered Graph approach as inputs to the GAT model. The model incorporates weight and allocation constraints, along with a loss function derived from the Sharpe ratio to maximize risk-adjusted returns. Results show that the GAT-based portfolio consistently outperforms benchmark portfolios over an extended period and provides valuable insights into market dynamics. This innovative approach demonstrates the potential of utilizing Graph Attention networks for enhancing portfolio optimization strategies in financial markets.