Paper: arXiv 2306.15526
Authors: Yang Qiao, Yiping Xia, Xiang Li, Zheng Li, Yan Ge
Abstract
Stock selection is important for investors to construct profitable portfolios. Graph neural networks (GNNs) are increasingly attracting researchers for stock prediction due to their strong ability of relation modelling and generalisation. However, the existing GNN methods only focus on simple pairwise stock relation and do not capture complex higher-order structures modelling relations more than two nodes. In addition, they only consider factors of technical analysis and overlook factors of fundamental analysis that can affect the stock trend significantly. Motivated by them, we propose higher-order graph attention network with joint analysis (H-GAT). H-GAT is able to capture higher-order structures and jointly incorporate factors of fundamental analysis with factors of technical analysis. Specifically, the sequential layer of H-GAT take both types of factors as the input of a long-short term memory model. The relation embedding layer of H-GAT constructs a higher-order graph and learn node embedding with GAT. We then predict the ranks of stock return. Extensive experiments demonstrate the superiority of our H-GAT method on the profitability test and Sharp ratio over both NSDAQ and NYSE datasets
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)
Why this score: The paper employs advanced graph neural network architectures (higher-order GAT with LSTM) with dense mathematical notation and definitions, while presenting backtest-ready experiments with concrete profitability metrics (Sharpe ratio, IRR) on multiple real-world datasets (NYSE, NASDAQ).
Research Flowchart
flowchart TD A["Research Goal: Develop H-GAT for stock selection<br>incorporating higher-order structures and joint analysis"] --> B["Data Preparation"] B --> B1["NASDAQ & NYSE Datasets"] B --> B2["Inputs: Fundamental & Technical Factors"] B --> B3["Graph Construction"] B3 --> C["Methodology: H-GAT Model Architecture"] C --> C1["Sequential Layer<br>LSTM for Temporal Features"] C --> C2["Relation Embedding Layer<br>Higher-order Graph + GAT"] C1 & C2 --> D["Computational Process<br>Learn Stock Embeddings & Predict Returns"] D --> E["Key Findings / Outcomes"] E --> E1["Superior Profitability"] E --> E2["Higher Sharpe Ratio"] E --> E3["Effectiveness of Higher-order & Joint Analysis"]
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 Higher-order Graph Attention Network (H‑GAT) that jointly incorporates technical and fundamental analysis and learns higher‑order stock relations for ranking stock returns.
Source sentence
Motivated by them, we propose higher-order graph attention network with joint analysis (H‑GAT). H‑GAT is able to capture higher-order structures and jointly incorporate factors of fundamental analysis with factors of technical analysis.
Assumes 3
- The stock market has high liquidity, enabling immediate settlement for all purchases and sales. market
Source sentence
We assume that the stock market has high liquidity, enabling immediate settlement for all purchases and sales.
- There are no commission charges for trading. execution
Source sentence
We assume that there are no commission charges, despite engaging in high-frequency stock trading.
- Profits are not reinvested; the investment principal remains fixed. model
Source sentence
We consider profits earned by investors as not reinvested, and the investment principal remains fixed.
Data
- Comprehensive stock dataset market-prices unknown
Source sentence
We collect all stocks traded on NYSE and NASDAQ, the two largest stock markets by market capitalization in the world, with 3,404 and 5,717 stocks, respectively.
- Quarterly financial statements fundamentals unknown
Source sentence
We collected quarterly financial statements for over 9,000 stocks on the NYSE and Nsadaq over the past decade.
Costs
Not modeled
Source sentence
We assume that there are no commission charges, despite engaging in high-frequency stock trading.
Finds 4
- H‑GAT achieves the best IRR on both NASDAQ and NYSE, outperforming the second‑best method RSR by 25.00% and 5.20% respectively. performance
Source sentence
Our H‑GAT has the best IRR in both datasets, which indicates that it achieves the highest profits. Specifically, our H‑GAT is significantly increased by 25.00% and 5.20% over the second best method RSR on both datasets.
- H‑GAT beats baseline models in mean reciprocal rank (MRR) and IRR, with only a slightly lower Sharpe Ratio than RSR on NASDAQ. performance
Source sentence
H‑GAT beats baseline models in MRR and IRR evaluation metrics, and only had a slightly lower Sharpe Ratio than RSR on Nasdaq.
- Over a long‑term backtest, H‑GAT’s annualized IRR is 9–10 times that of the Dow Jones Industrial Average and S&P 500. performance
Source sentence
Our model outperforms the Dow Jones Industrial Average (DJI) and the Standard and Poor’s 500 (S&P 500). It indicates that H‑GAT is worthy for long-term investment, with an annualized IRR 9-10 times that of the stock index.
- The Sharpe Ratio of H‑GAT (Top 1) is 1.70 on NASDAQ and 1.59 on NYSE, modestly higher than the indices (0.98 and 0.88) but lower than its IRR advantage. performance
Source sentence
The Sharpe ratio of the Top 1 on NSADAQ and NYSE is 1.70 and 1.59. The sharp ratio of two stock indexes was 0.98 (S&P 500) and 0.88 (DJI).
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.
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.