STN-TGAT: граф-нейросеть для ранжирования акций S&P 500 и сборки портфеля

How STN-TGAT is built

The STN-TGAT model is made up of two blocks: a Transformer captures long sequential patterns in price data, while a Graph Attention Network describes dynamic relationships between stocks. On top of this works a prior graph based on normalized mutual information (NMI), which tells the network which stocks are actually related.

• Architecture — temporal Transformer + Graph Attention Network
• The prior graph is built from normalized mutual information (NMI) between stocks
• A soft-threshold sparsification mechanism filters out noisy correlations while preserving meaningful links
• Training is decision-aligned, not just aligned to price forecasting
• The portfolio is the top 5 stocks within the top 50 S&P 500 companies

The key technique is "soft-threshold sparsification." It removes random, noisy correlations from the graph, but instead of cutting useful connections abruptly, it weakens them gradually. According to the authors, this makes the graph structure more resilient to market noise.

Why this is closer to real trading

STN-TGAT is evaluated not by abstract forecast accuracy, but by rules that approximate real trading. The model selects five stocks from the top 50 largest S&P 500 companies, explicitly allocates weights among them, and factors in transaction costs — the thing that usually eats into the returns of paper strategies.

This is an important difference from many academic stock-ranking models: they often show impressive prediction accuracy but ignore the fact that every portfolio rebalancing costs money. Here, transaction costs are built directly into the portfolio-construction procedure, and stock selection is limited to liquid shares of the index's largest issuers.

Does the model beat the benchmarks?

According to the authors, STN-TGAT consistently outperforms benchmark models on two fronts at once — forecast accuracy and portfolio returns on real data. In other words, the gain is not just that the network predicts price movements more accurately, but also that a portfolio built from its signals delivers more even after costs.

"The combination of decision-aligned training with adaptive relational

modeling provides a coherent and practically effective foundation for data-driven portfolio construction," the abstract of the arXiv paper states.

The number of verifiable claims here is limited: this is a preprint (announce type: new) that has not yet undergone peer review, and the authors compare against benchmarks without disclosing exact return figures in the abstract.

What this means

STN-TGAT is part of a trend in which graph neural networks are entering quantitative finance: the market is described not as a set of independent stocks, but as a network of relationships that changes over time. Practical value will emerge once the results are confirmed outside the preprint and over a long track record, including declining markets.

FAQ

What is STN-TGAT?

STN-TGAT (Soft-Threshold NMI-prior Transformer Graph Attention Network) is a neural network model for ranking stocks and building an investment portfolio. It combines a temporal Transformer, a Graph Attention Network, and a prior graph based on mutual information between stocks.

What data was the model tested on?

The model was tested on real S&P 500 stock data: the portfolio was built from the top 5 stocks within the top 50 largest companies in the index, with weight allocation and transaction costs taken into account.