A Review on Graph Neural Network Methods in Financial Applications

TL;DR

Financial data exhibit rich relational structure that is often heterogeneous and time-varying, posing unique modeling challenges. The survey categorizes financial graphs by construction and type, reviews feature processing for sequential and textual data, and maps GNN techniques to homogeneous, directed, bipartite, multi-relational, and dynamic graphs across key financial applications. Key contributions include a systematic graph taxonomy, a consolidated view of x-applications (stock movement, loan defaults, recommender systems, fraud detection, event prediction), and a discussion of evaluation, explainability, data availability, and scalability challenges with future directions. The work serves as a practical guide for researchers and practitioners to implement, compare, and extend GNN-based solutions in finance while highlighting gaps in reproducibility and graph-quality assessment.

Abstract

With multiple components and relations, financial data are often presented as graph data, since it could represent both the individual features and the complicated relations. Due to the complexity and volatility of the financial market, the graph constructed on the financial data is often heterogeneous or time-varying, which imposes challenges on modeling technology. Among the graph modeling technologies, graph neural network (GNN) models are able to handle the complex graph structure and achieve great performance and thus could be used to solve financial tasks. In this work, we provide a comprehensive review of GNN models in recent financial context. We first categorize the commonly-used financial graphs and summarize the feature processing step for each node. Then we summarize the GNN methodology for each graph type, application in each area, and propose some potential research areas.

Figures (7)

• Figure 1: Workflow for stock movement prediction task using GNN methodology. The graph construction and feature processing steps present stock information in a graph and a feature matrix, which is then used as the input for the GNN model. In the graph, nodes are connected if there exist some relationships between stocks, such as supplier, competitor, shake-holder, etc. A multi-layer perception layer (MLP) is used to output the price prediction result.
• Figure 2: Graph categorization based on graph characteristics. Each color of the circle represents a node type and each color of the line represents an edge type. Arrows represent directed edges. A homogeneous graph is a graph with one type of node and one type of edge. A directed graph is a graph with directed edges. A bipartite graph is a graph with two types of nodes and edges only exist between nodes of different types. A multi-relation graph has edges with different types. A dynamic graph is a sequence of graphs.
• Figure 3: Feature processing for sequential features and textual information. For sequential numerical features, recurrent neural network (RNN) based approaches are commonly used to capture the temporal dependencies. For text features, it is often processed utilizing natural language processing (NLP) methods including word embedding, sentence embedding, and language models, to convert the unstructured data to structured ones.
• Figure 4: Graph neural network models for different graph types. The term $\text{Conv}$ denotes graph convolution process. The term $\text{MLP}$ denotes the multi-layer perception. The term $\text{Agg}$ denotes the aggregation process. The term $\text{Split}$ denotes data splitting according to its characteristics.
• Figure :

Theorems & Definitions (8)

• Definition 2.1: Graph
• Definition 2.2: Adjacency matrix
• Definition 2.3: Undirected graph and Directed graph
• Remark
• Definition 2.4: Bipartite graph
• Definition 2.5: Homogeneous graph and Heterogeneous graph
• Definition 2.6: Multi-relation graph
• Definition 2.7: Dynamic graph