Paper: arXiv 2601.04602

Authors: Jack Fanshawe, Rumi Masih, Alexander Cameron

Abstract

This paper studies forward-looking stock-stock correlation forecasting for S&P 500 constituents and evaluates whether learned correlation forecasts can improve graph-based clustering used in basket trading strategies. We cast 10-day ahead correlation prediction in Fisher-z space and train a Temporal-Heterogeneous Graph Neural Network (THGNN) to predict residual deviations from a rolling historical baseline. The architecture combines a Transformer-based temporal encoder, which captures non-stationary, complex, temporal dependencies, with an edge-aware graph attention network that propagates cross-asset information over the equity network. Inputs span daily returns, technicals, sector structure, previous correlations, and macro signals, enabling regime-aware forecasts and attention-based feature and neighbor importance to provide interpretability. Out-of-sample results from 2019-2024 show that the proposed model meaningfully reduces correlation forecasting error relative to rolling-window estimates. When integrated into a graph-based clustering framework, forward-looking correlations produce adaptable and economically meaningfully baskets, particularly during periods of market stress. These findings suggest that improvements in correlation forecasts translate into meaningful gains during portfolio construction tasks.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced deep learning architectures (Transformer and GNN) with hybrid modeling and Fisher-z transformations, indicating high mathematical density. Empirical results are validated on recent S&P 500 data (2019-2024) with out-of-sample error metrics and portfolio integration, demonstrating significant backtest-ready implementation.

Research Flowchart

  flowchart TD
  A["Research Goal: Forecast Stock Correlations<br>and Evaluate Trading Applications"] --> B["Data & Inputs<br>Historical Returns, Technicals, Sectors, Macro Signals"]
  B --> C["Methodology: Hybrid THGNN<br>Transformer Temporal Encoder + Graph Attention Network"]
  C --> D["Model Training: Predict Residuals<br>in Fisher-z Space (10-Day Horizon)"]
  D --> E["Out-of-Sample Validation<br>Period: 2019-2024"]
  E --> F{"Key Outcomes"}
  F --> G["Forecasting<br>Reduced Error vs. Rolling Baseline"]
  F --> H["Application: Graph Clustering<br>Economically Meaningful Baskets<br>Strong Performance During Market Stress"]

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 Temporal-Heterogeneous Graph Neural Network (THGNN) that combines a Transformer temporal encoder with a graph‑attention module to forecast 10‑day ahead stock‑stock correlations for S&P 500 constituents.

Source sentence
We cast 10-day ahead correlation prediction in Fisher-z space and train a Temporal-Heterogeneous Graph Neural Network (THGNN) to predict residual deviations from a rolling historical baseline.

Assumes 4

  • No slippage costs are considered in the backtest. execution
    Source sentence
    Firstly, no slippage costs are considered in the backtest so realized performance may be overstated.
  • The model is trained only under one set of hyper‑parameters (e.g., Transformer sequence length, learning rate, attention heads). model
    Source sentence
    the THGNN has not been compared to alternative ML models (LSTM, CNN, alternative GNN‑based models), limiting comparative claims. Thirdly, the model has been trained only under one set of hyperparameters (such as Transformer sequence length, learning rate, attention heads), and therefore it is almost certain predictive
  • A stock is included in the daily graph only if it has at least 30 days of data within the preceding 33 trading days. data
    Source sentence
    Specifically, as the sequence length of the transformer is 30 days, if a stock has 30 days of data within 33 trading days it is included and if not, it is excluded from graph construction.
  • Forward‑looking predicted correlation matrices replace backward‑looking rolling‑window correlations in the SPONGEsym clustering and trading engine. model
    Source sentence
    we retain their trading engine but use forward‑looking predicted correlation matrices from the THGNN, instead of backward‑looking rolling‑window correlations.

Data

Backtest sample: 2019 – 2024 · S&P 500 stocks · equities · daily

  • CRSP and Compustat data · WRDS market-prices commercial
    2019-2024 (out‑of‑sample) · S&P 500 constituents
    Source sentence
    Equity data is obtained from CRSP and Computstat via WRDS, and macroeconomic variables from FRED.
  • FRED macroeconomic variables · FRED macro public
    Source sentence
    Equity data is obtained from CRSP and Computstat via WRDS, and macroeconomic variables from FRED.

Recognised sources: Yahoo Finance WRDS CRSP FRED

Costs

Partial Costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers.

Flat fee of 0.05% per trade applied in robustness checks.

Source sentence
with a cost of 0.05% per trade from 2000-2024.

Finds 4

Reported performance (THGNN forward‑looking SPONGEsym basket trading strategy): Sharpe 1.84 (benchmark 0.65) · annual return 19.2% · max drawdown 9.4%

  • The THGNN reduces MAE from 0.3071 (20‑day historical rolling window) to 0.2302 and RMSE from 0.3852 to 0.2940 over the 2019‑2024 out‑of‑sample period. predictive-accuracy
    Source sentence
    The out‑of‑sample period (2019-2024) showed the model is able to reduce MAE from 0.3071 (20‑day historical rolling window) to 0.2302 and RMSE from 0.3852 to 0.2940.
  • Pearson and Spearman correlation between predicted and actual future correlations more than double to 0.778 and 0.795 respectively. predictive-accuracy
    Source sentence
    The model also reduced bias and improves both Pearson and Spearman correlation more than double to 0.778 and 0.795 respectively.
  • Using forward‑looking correlations in SPONGEsym clustering yields an own strategy annualized return of 8.27% (Sharpe 0.86) without transaction costs, but performance drops to -0.46% (Sharpe 0.00) with a 0.05% cost. performance
    Source sentence
    Own (No TC) 8.27 0.86; Own (0.05% TC) -0.46 0.00
  • Attention diagnostics show the model consistently emphasizes macro‑financial drivers such as VIX, crude oil prices, the 10‑year U.S. Treasury yield, the trade‑weighted dollar, and sector identifiers. other
    Source sentence
    Attention diagnostics show that the model consistently emphasizes a small set of macro‑financial drivers - most notably market volatility (VIX), crude oil prices, the 10‑year U.S. Treasury yield, the trade‑weighted dollar, and sector identifiers

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 41 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.

Cited References (70)

  1. Deep Learning for Portfolio Optimization — link
  2. Inductive Representation Learning on Large Graphs — link
  3. Attention Is All You Need — link
  4. Pairs Trading: Performance of a Relative Value Arbitrage Rule — link
  5. Statistical Arbitrage in the U.S. Equities Market — link
  6. Industry costs of equity — link
  7. Correlation Scenarios and Correlation Stress Testing — link
  8. Empirical Investigation of an Equity Pairs Trading Strategy — link
and 62 more
  1. High-frequency Pairs Trading on a Small Stock Exchange — link
  2. Applying Hurst Exponent in pair trading strategies on Nasdaq 100 index — link
  3. The profitability of pairs trading strategies: distance, cointegration and copula methods — link
  4. In Search of Pairs using Firm Fundamentals: Is Pairs Trading Profitable? — link
  5. Hierarchical Structure in Financial Markets — link
  6. A tool for filtering information in complex systems — link
  7. ( PDF ) Spectral Analysis of Signed Graphs for Clustering, Prediction and Visualization — link
  8. Portfolio Selection — link
  9. Common risk factors in the returns on stocks and bonds — link
  10. Dynamic Conditional Correlation - a Simple Class of Multivariate GARCH Models — link
  11. Forecasting the Dynamic Correlation of Stock Indices Based on Deep Learning Method — link
  12. Analysis of Financial Contagion and Prediction of Dynamic Correlations During the COVID -19 Pandemic: A Combined DCC - GARCH and Deep Learning Approach — link
  13. Extreme Correlation of International Equity Markets — link
  14. Structural breaks, dynamic correlations, and hedge and safe havens for stock and foreign exchange markets in Greater China — link
  15. A Review on Graph Neural Network Methods in Financial Applications — link
  16. Graph-Based Methods for Forecasting Realized Covariances — link
  17. Stock Selection via Spatiotemporal Hypergraph Attention Network: A Learning to Rank Approach — link
  18. On the difficulty of training recurrent neural networks — link
  19. Understanding the Market Trends: A Hybrid Approach to Stock Price Prediction Using RNNs and Transformer-Based Sentiment Analysis — link
  20. Forecasting realized volatility in turbulent times using temporal fusion transformers
  21. Regime-Switching Models.
  22. New Approach to Markov-Switching GARCH Models \textbar Journal of Financial Econometrics \textbar Oxford Academic — link
  23. A New Approach to Markov-Switching GARCH Models — link
  24. Stylized Facts of Daily Return Series and the Hidden Markov Model — link
  25. Computation and analysis of multiple structural change models — link
  26. Equity Portfolio Diversification: How Many Stocks Are Enough? Evidence From India — link
  27. Examining Long and Variable Lags in Monetary Policy — link
  28. Supply Chain Disruptions: Evidence from the Great East Japan Earthquake* — link
  29. Statement of Commissioner Dan M. Berkovitz on Recent Trading in the WTI Futures Contract before the Energy and Environmental Markets Advisory Committee Meeting \textbar CFTC — link
  30. Histogram-Based Algorithm for Building Gradient Boosting Ensembles of Piecewise Linear Decision Trees — link
  31. Explaining the Success of AdaBoost and Random Forests as Interpolating Classifiers
  32. Deep Learning (2016) — link
  33. Large-Scale Machine Learning with Stochastic Gradient Descent — link
  34. Applied Logistic Regression
  35. Asset returns and volatility clustering in financial time series — link
  36. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation — link
  37. An Analysis of Variance Test for Normality (Complete Samples) — link
  38. On a Measure of Lack of Fit in Time Series Models — link
  39. Automatic Block-Length Selection for the Dependent Bootstrap — link
  40. Federal Reserve Economic Data \textbar FRED \textbar St. Louis Fed — link
  41. Wharton Research Data Services — link
  42. An introduction to variable and feature selection
  43. The Average Stock Market Return: About 10% — link
  44. Monthly markets review - October 2020 — link
  45. A Review on Graph Neural Network Methods in Financial Applications — link
  46. Signed Graph Convolutional Network — link
  47. Graph Attention Networks — link
  48. Multivariate Realized Volatility Forecasting with Graph Neural Network — link
  49. Stock Price Correlation Coefficient Prediction with ARIMA - LSTM Hybrid Model — link
  50. An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling — link
  51. Financial Time Series Forecasting using CNN and Transformer — → 2304.04912
  52. Soft Calibration Objectives for Neural Networks — link
  53. Not Just a Black Box: Learning Important Features Through Propagating Activation Differences — link
  54. A Unified Approach to Interpreting Model Predictions — link
  55. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier — link
  56. Learning Important Features Through Propagating Activation Differences — link
  57. Axiomatic Attribution for Deep Networks — link
  58. SmoothGrad : removing noise by adding noise — link
  59. Full-Gradient Representation for Neural Network Visualization — link
  60. Attention is not Explanation — link
  61. Statistical arbitrage in multi-pair trading strategy based on graph clustering algorithms in US equities market — → 2406.10695
  62. Correlation Matrix Clustering for Statistical Arbitrage Portfolios — link

2 of 70 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.