Evaluating Financial Relational Graphs: Interpretation Before Prediction

TL;DR

The paper tackles the interpretability gap in financial graph-based forecasting by decoupling graph evaluation from downstream tasks and introducing SPNews for dynamic relationship graphs. It defines a time-evolving graph set $\mathcal{G}=\{\mathcal{G}_t\}$ where edges $E_t(A,B)=(\mu_t^{(A,B)},\nu_t^{(A,B)})$ are built from news co-occurrence with thresholding and normalization. A novel, task-independent evaluation framework (FRI) with Return Correlation Stability, Event Detection, and Edge Factor Model components assesses graphs via $CSS$, $AECR$, $\Delta_{\beta}$, and $\Delta_{DCC}$ metrics, providing interpretability beyond predictive performance. Experiments show SPNews-based dynamic graphs yield more interpretable relationships and can outperform traditional correlation-based graphs, offering practical value for constructing robust, explainable financial graphs. The release of SPNews and the FRI framework aids reproducibility and enables broader exploration of news-informed, time-varying financial relationships.

Abstract

Accurate and robust stock trend forecasting has been a crucial and challenging task, as stock price changes are influenced by multiple factors. Graph neural network-based methods have recently achieved remarkable success in this domain by constructing stock relationship graphs that reflect internal factors and relationships between stocks. However, most of these methods rely on predefined factors to construct static stock relationship graphs due to the lack of suitable datasets, failing to capture the dynamic changes in stock relationships. Moreover, the evaluation of relationship graphs in these methods is often tied to the performance of neural network models on downstream tasks, leading to confusion and imprecision. To address these issues, we introduce the SPNews dataset, collected based on S\&P 500 Index stocks, to facilitate the construction of dynamic relationship graphs. Furthermore, we propose a novel set of financial relationship graph evaluation methods that are independent of downstream tasks. By using the relationship graph to explain historical financial phenomena, we assess its validity before constructing a graph neural network, ensuring the graph's effectiveness in capturing relevant financial relationships. Experimental results demonstrate that our evaluation methods can effectively differentiate between various financial relationship graphs, yielding more interpretable results compared to traditional approaches. We make our source code publicly available on GitHub to promote reproducibility and further research in this area.

Figures (2)

• Figure 1: (a): Dynamic relationship graph set $\mathcal{G}$. (b): relationship factor ($HML_R$) construction. (c): $HML_R$ evaluation. (b) and (c) corresponds to Algorithm \ref{['alg:factor-construct']} & \ref{['alg:factor-test']} respectively. A, B, C, D, P, and Q are examples of nodes(companies). In each matrix of (b)(3) and (c)(3), each column is the rolling correlation coefficient of a node pair. The 3 matrices in (b)(3) have different shapes, that are $[T-\epsilon, n_{high}], [T-\epsilon, n_{medium}], [T-\epsilon, n_{low}]$ respectively, where $n_{high}, n_{medium}, n_{low}$ represent the number of node pairs in $group_{high}, group_{medium}$, and $group_{low}$. $T$ and $\epsilon$ follow the definition in Table \ref{['tab:symbol']}. The 10 matrices in (c)(3) have the same shape $[T-\epsilon, 100]$.
• Figure 2: Apple-JP Morgan Case Study. Line plot: the rolling return correlation. Scatter points: edge index $\mu_t^{(AAPL, JPM)}$.