[Paper Review] GraphCNNpred: A stock market indices prediction using a Graph based deep learning system

Yuhui Jin|arXiv (Cornell University)|Jul 4, 2024

Stock Market Forecasting Methods6 citations

TL;DR

This study integrates graph neural networks with CNNpred to predict trends of major stock indices, achieving 4–15% F-measure gains over baselines and enabling trading simulations with Sharpe ratios over 3.

ABSTRACT

The application of deep learning techniques for predicting stock market prices is a prominent and widely researched topic in the field of data science. To effectively predict market trends, it is essential to utilize a diversified dataset. In this paper, we give a graph neural network based convolutional neural network (CNN) model, that can be applied on diverse source of data, in the attempt to extract features to predict the trends of indices of ext{S}\& ext{P} 500, NASDAQ, DJI, NYSE, and RUSSEL. The experiments show that the associated models improve the performance of prediction in all indices over the baseline algorithms by about $4\% ext{ to } 15\%$, in terms of F-measure. A trading simulation is generated from predictions and gained a Sharpe ratio of over 3.

Motivation & Objective

• Motivate improved stock index prediction using diverse data sources.
• Propose a Graph-CNN hybrid model (GAT/GCN + CNNpred) to extract daily and durational features.
• Evaluate prediction performance across major indices and compare with CNNpred baselines and pure graph methods.
• Demonstrate trading viability via Sharpe ratio and CEQ in simulated portfolios.

Proposed method

• Construct a feature correlation graph with 138 features as nodes and edges for correlations above 0.7.
• Use Graph Attention Network (GAT) or Graph Convolutional Network (GCN) layers to disseminate information across correlated features.
• Incorporate CNNpred-style durational feature extraction with CNN layers to capture temporal patterns.
• Aggregate graph and CNN features with pooling (mean, max, or fully connected) to form a final representation for prediction.
• Train with binary (0,1) or 0/1/2 labeling schemes for short-term returns, using a sigmoid final layer for probabilities or a softmax/CEQ for multiclass outputs.
• Evaluate via macro F-measure for 0/1 labeling and trading metrics (Sharpe, CEQ) for 0/1/2 labeling.

Experimental results

Research questions

• RQ1 Can a graph-based integration with CNNpred improve short-term stock index trend prediction over CNNpred alone and over pure graph methods?
• RQ2 How do different graph pooling strategies and GNN variants (GAT vs GCN) affect predictive performance across major indices?
• RQ3 What is the trading performance (Sharpe, CEQ) when predictions drive simulated long/short strategies for multiple indices?

Key findings

• GCN-CNNpred and GAT-CNNpred consistently outperform CNNpred baselines and pure graph baselines in mean F-measure across indices.
• Best single-index results (Table 1) show GCN-CNNpred achieving the highest F-measures on several indices (e.g., S&P 500: 0.5866; NASDAQ: 0.5885; NYSE: 0.5767; RUSSELL: 0.5954).
• Graph-based variants with CNN integration yield significant improvements (4% to 15% in F-measure) over baselines.
• The models enable trading simulations with Sharpe ratios above baseline strategies, indicating practical trading viability.
• Pooling choices (mean, max, fully connected) are evaluated and selected based on performance, contributing to robust predictions.