Paper: arXiv 2407.15532
Authors: Kamesh Korangi, Christophe Mues, Cristián Bravo
Abstract
Apart from assessing individual asset performance, investors in financial markets also need to consider how a set of firms performs collectively as a portfolio. Whereas traditional Markowitz-based mean-variance portfolios are widespread, network-based optimisation techniques offer a more flexible tool to capture complex interdependencies between asset values. However, most of the existing studies do not contain firms at risk of default and remove any firms that drop off indices over a certain time. This is the first study to also incorporate such firms in portfolio optimisation on a large scale. We propose and empirically test a novel method that leverages Graph Attention networks (GATs), a subclass of Graph Neural Networks (GNNs). GNNs, as deep learning-based models, can exploit network data to uncover nonlinear relationships. Their ability to handle high-dimensional data and accommodate customised layers for specific purposes makes them appealing for large-scale problems such as mid- and small-cap portfolio optimisation. This study utilises 30 years of data on mid-cap firms, creating graphs of firms using distance correlation and the Triangulated Maximally Filtered Graph approach. These graphs are the inputs to a GAT model incorporating weight and allocation constraints and a loss function derived from the Sharpe ratio, thus focusing on maximising portfolio risk-adjusted returns. This new model is benchmarked against a network characteristic-based portfolio, a mean variance-based portfolio, and an equal-weighted portfolio. The results show that the portfolio produced by the GAT-based model outperforms all benchmarks and is consistently superior to other strategies over a long period, while also being informative of market dynamics.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.5)
Why this score: The paper employs advanced deep learning architectures (Graph Attention Networks) with custom layers and a non-standard loss function (Sharpe ratio derived), requiring significant mathematical formulation, yet it is heavily grounded in empirical validation using 30 years of real-world mid-cap data, detailed graph construction methods (distance correlation, TMFG), and benchmarking against multiple established strategies.
Research Flowchart
flowchart TD
A["Research Goal:<br>Large-scale Time-Varying Portfolio<br>Optimisation for Mid/Small-cap Firms"] --> B["Data Processing & Graph Construction<br>30y data; Distance Correlation &<br>Triangulated Maximally Filtered Graph"]
B --> C["Graph Attention Network<br>GAT model with Sharpe Ratio loss<br>Custom weight & allocation constraints"]
C --> D["Optimisation & Model Training<br>Including defaulting/removed firms<br>Large-scale processing"]
D --> E["Output: Time-Varying Portfolio<br>Dynamic asset weights & allocations"]
E --> F["Benchmarking & Comparison<br>vs. Mean Variance, Network Characteristic,<br>& Equal-weight portfolios"]
F --> G{"Performance Evaluation"}
G --> H["Key Findings<br>GAT portfolio consistently<br>outperforms benchmarks<br>Superior risk-adjusted returns (Sharpe)<br>Informative of market dynamics"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a novel Graph Attention Network (GAT) based model for large‑scale mid‑cap portfolio optimisation.
Source sentence
We propose and empirically test a novel method that leverages Graph Attention networks (GATs), a subclass of Graph Neural Networks (GNNs).
Assumes 2
- No transaction costs are incurred when trading the equal‑weight benchmark. execution
Source sentence
For simplicity, however, we assume that there are no transaction costs1, and use this equal-weight portfolio as the market benchmark in our study.
- The classical mean‑variance model assumes normality of returns and absence of transaction costs. model
Source sentence
Assumptions about the normality of returns and absence of transaction costs
Data
mid-cap equities · equities
- 30 years of mid‑cap firm data market-prices unknown
Source sentence
This study utilises 30 years of data on mid-cap firms, creating graphs of firms using distance correlation and the Triangulated Maximally Filtered Graph approach.
Costs
Not modeled
Source sentence
For simplicity, however, we assume that there are no transaction costs1, and use this equal-weight portfolio as the market benchmark in our study.
Finds 3
Sharpe 1.08 (benchmark 0.83)
- The GAT model achieves the highest Sharpe ratio (1.082) on test data, outperforming all benchmark strategies. performance
Source sentence
On the test data, the GAT models have the highest Sharpe ratio of 1.082.
- The GAT portfolio exhibits the lowest excess turnover (4.10%) compared with mean‑variance (5.50%) and network index (4.84%) models. performance
Source sentence
With an overall excess turnover of 4.10%, the GAT model is thus associated with a relative reduction of 25% and 15%, compared to the mean-variance and network index models, respectively.
- The GAT model holds far fewer firms than the mean‑variance implementation, reflecting a regularisation mechanism. performance
Source sentence
the GAT model requires holding far fewer firms than our mean-variance implementation, which is unsurprising given that the former includes a regularisation mechanism that is lacking in the latter.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 29 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.