[Paper Review] Exploring Graph Neural Networks for Stock Market Predictions with Rolling Window Analysis

Daiki Matsunaga, Toyotaro Suzumura+1 more|arXiv (Cornell University)|Sep 23, 2019

Stock Market Forecasting Methods57 citations

TL;DR

The paper evaluates graph neural networks that use company knowledge graphs to predict Nikkei 225 stock prices, using rolling window backtesting to compare against LSTM and benchmarks, with notable gains for all and certain relation types.

ABSTRACT

Recently, there has been a surge of interest in the use of machine learning to help aid in the accurate predictions of financial markets. Despite the exciting advances in this cross-section of finance and AI, many of the current approaches are limited to using technical analysis to capture historical trends of each stock price and thus limited to certain experimental setups to obtain good prediction results. On the other hand, professional investors additionally use their rich knowledge of inter-market and inter-company relations to map the connectivity of companies and events, and use this map to make better market predictions. For instance, they would predict the movement of a certain company's stock price based not only on its former stock price trends but also on the performance of its suppliers or customers, the overall industry, macroeconomic factors and trade policies. This paper investigates the effectiveness of work at the intersection of market predictions and graph neural networks, which hold the potential to mimic the ways in which investors make decisions by incorporating company knowledge graphs directly into the predictive model. The main goal of this work is to test the validity of this approach across different markets and longer time horizons for backtesting using rolling window analysis. In this work, we concentrate on the prediction of individual stock prices in the Japanese Nikkei 225 market over a period of roughly 20 years. For the knowledge graph, we use the Nikkei Value Search data, which is a rich dataset showing mainly supplier relations among Japanese and foreign companies. Our preliminary results show a 29.5% increase and a 2.2-fold increase in the return ratio and Sharpe ratio, respectively, when compared to the market benchmark, as well as a 6.32% increase and 1.3-fold increase, respectively, compared to the baseline LSTM model.

Motivation & Objective

• Motivate the use of knowledge graphs to capture inter-company and inter-market relationships in stock prediction.
• Assess the generalizability of graph neural networks across markets and long time horizons using rolling window backtesting.
• Evaluate how different knowledge-graph relation types affect predictive performance on stock prices.
• Investigate end-to-end training of a temporal graph neural network for regression on future returns.

Proposed method

• Use Graph Convolutional Networks (GCN) with node features derived from an LSTM processing historical price data.
• Incorporate temporal dynamics by weighting neighbor contributions with a dot-product based relation strength (Equations 3).
• Construct seven adjacency matrices from first- and second-order company relations (supplier, customer, partner, shareholder, common-industry, common-customer, common-supplier) and test an all-types adjacency A (K relation types).
• Apply rolling window analysis (train 2000 timesteps, test 200 timesteps, slide window) for backtesting over roughly 20 years of data.
• Train end-to-end, jointly optimizing RNN (LSTM) and temporal graph convolution layers with a final regression head.
• Evaluate performance using yearly return ratio and Sharpe ratio against Buy-Hold and baselines.

Experimental results

Research questions

• RQ1 Are knowledge-graph datasets of company relations useful for stock prediction across different markets and longer time horizons?
• RQ2 Do temporal graph convolutions with dynamic relation strengths improve predictive performance beyond static graphs?
• RQ3 Which relation types (e.g., customer-of, supplier-of, common-supplier) most improve predictive accuracy in the Nikkei 225 context?
• RQ4 Can end-to-end training of RNN + GCN architectures outperform standard LSTM and market benchmarks over rolling windows?

Key findings

• All graph-based models outperform the Buy-Hold benchmark in return ratio and Sharpe ratio.
• The all-relations graph yields the best performance among the tested graphs (Return Ratio 29.95%, Sharpe 0.37).
• The customer-of relation independently improves predictions (Return 27.52%, Sharpe 0.35), outperforming LSTM in some metrics.
• LSTM is outperformed by certain graph types (e.g., customer-of, all) on Sharpe and returns for Nikkei 225 data.
• Second-order and multi-relation graphs help mitigate sparsity and improve generalization over long horizons.