Your Portfolio is a Social Network: Using AI to See the Hidden Connections
Trading & Investing / Artificial Intelligence & Machine Learning
Paper Review by Helene Sy
Your Portfolio is a Social Network
Your portfolio is a social network. You just don’t have the tools to see the hidden connections. A new machine learning model acts like a LinkedIn for the entire market, with powerful results. This research paper argues that to truly price an asset, we must look beyond its individual characteristics and model the industrial chain dynamics between companies. The study entitled “Exploring Asset Pricing Through Industrial Chain Dynamics: Insights from Graph Machine Learning” by Chunhong Ye and Bin Liu explores this concept. It shows how “momentum spillover effects” can be quantified, moving beyond the limitations of traditional factor-based models.
Fusing Fundamentals and Dynamics: How the Model Works
The authors’ core innovation is a sophisticated model that sees the market through two different lenses at once, processes the information with a purpose-built engine, and fuses the signals intelligently.
- Dual-Graph Approach: The model begins by mapping the market using two distinct but complementary graphs.
- The Industrial Chain Graph: This is a static map of fundamental business relationships, built from the expert-labeled ChinaScope industry chain dataset. It shows the physical flow of goods and services between upstream suppliers and downstream customers.
- The Industrial Chain Graph: This is a static map of fundamental business relationships, built from the expert-labeled ChinaScope industry chain dataset. It shows the physical flow of goods and services between upstream suppliers and downstream customers.
Figure 1. An example of the industrial chain graph for the New Energy Vehicle (NEV) industry.
Source: Ye & Liu (2025)
- The Temporal Relationship Graph: This dynamic map identifies companies with similar stock price patterns. It is constructed using the Dynamic Time Warping (DTW) algorithm on historical return sequences to capture market relationships not documented in the official supply chain data.
- Graph Neural Network (GNN) Engine: A Spatio-Temporal GNN processes this rich network data. Unlike traditional models that see stocks as isolated rows in a spreadsheet, a GNN is designed to understand how information and financial shocks propagate through the economic web, capturing the “momentum spillover effect”.
- Cross-Attention Fusion Mechanism: The model uses a sophisticated method to intelligently fuse the signals from both the static industrial graph and the dynamic temporal graph. This mechanism allows the model to adaptively decide which type of relationship is more important at any given time, a significant step beyond simple averaging.
Putting the Model to the Test: The Results
A creative model is only useful if it performs. The authors subjected their framework to a battery of empirical tests, with compelling results.
First, they evaluated its raw predictive power using out-of-sample R-squared (R2) against eight benchmark models, including both traditional linear models and other complex machine learning models like LSTM. While non-linear models generally performed better than linear ones, the proposed model achieved the best R2 overall, demonstrating a superior fit to the data. To ensure this outperformance wasn’t a fluke, a Diebold-Mariano (DM) test was used to confirm that the difference in forecasting accuracy was genuine and statistically significant against five of the benchmark models.
Next, the authors conducted ablation experiments to justify their complex design. By testing simpler versions of their model, they proved that every component—the industrial chain graph, the temporal graph, and the cross-attention fusion mechanism—was essential to its success.
From a practical finance perspective, the results were equally strong. An Information Coefficient (IC) analysis, which measures the reliability of predictive signals, showed the model achieved the highest IC mean and ICIR of all models tested, suggesting a greater potential to generate stable excess returns. This statistical edge was then tested in a simulated practical trading strategy. The authors constructed weekly rebalanced long-short portfolios based on each model’s predictions. Their proposed model generated a portfolio with the highest annualized return (15.18%) and the best risk-adjusted returns, evidenced by a superior Sharpe Ratio of 1.783.
Figure 2. The proposed model’s long-short portfolio achieved the best Sharpe Ratio compared to key benchmarks
Source: Adapted from Ye & Liu (2025), Table 6
Figure 3. Cumulative returns of the long-short portfolio based on the model’s predictions (Jan 2021 – Jun 2023)
Source: Ye & Liu (2025)
Finally, a portfolio monotonicity analysis confirmed the signal’s robustness. When stocks were sorted into ten groups based on the model’s predicted returns, their actual returns showed an apparent monotonic increase from the lowest group to the highest. This indicates that the model effectively differentiates performance across the entire spectrum of stocks.
Takeaway: A New Direction for Alpha Generation
The paper provides a powerful proof-of-concept that the future of asset pricing may lie in modeling the fundamental, economic connections between companies. While tested on the Chinese market, this network-based approach presents a new and promising direction for research and alpha generation in the US and other international markets. It suggests that the next edge won’t be found by discovering one more factor in the “zoo,” but by understanding the entire ecosystem.
Source: Liu, Bin and Ye, Chunhong, Exploring Asset Pricing Through Industrial Chain Dynamics: Insights from Graph Machine Learning (January 07, 2025). Available at SSRN: https://ssrn.com/abstract=5244473 or http://dx.doi.org/10.2139/ssrn.5244473
