[20240626_LabSeminar_Huy]MGGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction.pptx

MGGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction

1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: [email protected] 2024-06- 26 MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction Hao Qian et al. AAAI-2024: Proceedings of the Thirty-Eight Conference on Artificial Intelligence

2. OUTLINE MOTIVATION INTRODUCTION METHODOLOGY EXPERIMENT & RESULT CONCLUSION

3. MOTIVATION The stock market is a crucial component of the financial system Offer investors a marketplace to trade shares of a wide range of assets . Predicting movement of stock prices is not only important, but also challenges because of dynamic and intricate relations from various aspects: buy and sell, economic indicators, reports, news, politic, etc. Overview Previous works focus on two approaches: Traditional sequential methods: capture the temporal patterns of stock movement by sequential extraction technique (statistic, machine learning, or deep learning such as LSTM, GRU, etc ). Graph-based models: corporate heterogeneous information explicitly from data or implicitly mine it from textual data to capture the interdependence of stocks.

4. MOTIVATION Overview: Limitation Multifacetedness : The stock price movement influence: multiple relations among stocks, industries, investment banks, etc. Previous graph-based methods only utilized single relations between stocks, ignoring the potential of incorporating other complex relations as auxiliary information. Temporal: The movement of stock prices and multifaceted relations among stocks are exhibit temporal evolution . Relationships among stocks change over time due to factors: investment banks trading stocks, common shareholders coholding stocks, and companies releasing products into new industries. Previous works not consider this relation: historical trends and evolving relationships.

5. INTRODUCTION Discuss the multifacetedness and temporal in the context of stock investment prediction tasks : Provide insights on modeling complex stock relations based on empirical evidence. Propose to capture the multifaceted and temporal evolution nature of stocks with a multi-relational dynamic graph - Multi-relational Dynamic Graph Neural Network (MDGNN) Generate a comprehensive representation of the stock markets. Contribution

6. METHODOLOGY Problem Definition A Dynamic Graph Neural Network (DGNN) Capture relationships between stocks are multifaceted and changing daily. : a multi-relational graph snapshot at trading day and is the total number of snapshots. : closing price at trading day of stock node The ground-truth label of stock on trading day t: return between two consecutive trading days.   Problem: formulate the stock prediction as a node regression task utilizes DGNN to learn a scoring function parameterized by . minimizing the loss function:   benchmark is return of the benchmark index on trading day t.  

7. METHODOLOGY Main Architecture

8. METHODOLOGY Intra-day Graph Snapshot Multi-relational Graph Construction: Integrate relations from industry, investment banks, and stock pairs to establish a multi-relational graph. Industry Graph: performance of a company and its corresponding stock is closely tied to the industry where it operates. represent the stock as and the industry as , while the connection between them is denoted as . contains features that encode supply, demand, competition, and regulatory connections. Investment Bank Graph: act as market makers for stocks provide liquidity to the market by buying and selling stocks. extract the buy, sell, research, and advisory relations to capture the wield significant influence over stock prices. denote the investment bank as and the connection between that and stock as . Stock Graph: stocks have a great impact on other stocks because of the interconnectedness of the stock market and the various factors that can affect stock prices Identify relationships between stocks based on factors such as sector, ownership, and co-holding relations.  

9. METHODOLOGY Intra-day Graph Snapshot Multi-relational Graph Construction: To create the multiplex relations as edges , first is gathering daily trading data and textual data macroeconomic reports, financial news, financial statements, and research reports. Employ financial lexicons and syntactic matching methods to build edges between pairs of entities. Rule-based with lexical and syntactic matching types. Financial lexicon: specialized dictionary or vocabulary of financial terms, concepts, and jargon. Syntactic matching: analyze the grammatical structure and patterns in text to identify relationships between entities.

10. METHODOLOGY Intra-day Graph Snapshot Hierarchical Multi-relational Graph Embedding Layer: Define a meta-paths: Stock-Stock (SS), Stock-Bank-Stock (SBS), and Stock-Industry-Industry-Stock (SIIS). Utilize the multi-head attention of GAT with average pooling to update the target node’s representation: Add more attention to focus on the most relevant representations for the target node between different meta-paths and reduce redundancy. where denotes the neighborhood nodes of the target node , represents the concatenation operation, and W is the shared projection matrix. is the shared projection matrix and K is the total number of heads.   , , and : meta-paths “SS”, “SBS”, and “SIIS”.   where W is a learnable matrix and is the representation of node after incorporating multiple relations among nodes.  

11. METHODOLOGY Inter-day Temporal Extraction Layer Node representation from trading day t and preceding trading days. Applying Transformer structure to extract temporal evolution of graph snapshots’ propagation within a time window size   Employ forward mask to prevent positions in input sequence from attending to subsequent positions. a static, non-learnable bias to the query-key dot product. an inductive bias in favor of recent events, as it imposes a penalty on attention scores between distant query key pairs. where m is a slope parameter, P is the position bias, and M is the forward mask matrix. Z combines the significant evolving patterns extracted from the historical data between time period and . denotes the representation of stock node on trading day .  

12. METHODOLOGY Inter-day Temporal Extraction Layer Prediction Layer: estimate the probability that a given stock will yield a positive return on trading day based on the stock’s representation :  

13. EXPERIMENT AND RESULT EXPERIMENT SETTINGs Dataset: China’s stock market: CSI100 and CSI300 index. Baselines: Deep Learning: MLP, LSTM, and Transformer [1]. Graph methods: GCN, GAT, RGCN[2], HAN[3], HGT[4], EvolveGCN [5], HTGNN[6]. [1] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. [2] Schlichtkrull , M., Kipf , T. N., Bloem, P., Van Den Berg, R., Titov, I., & Welling, M. (2018). Modeling relational data with graph convolutional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 (pp. 593-607). Springer International Publishing. [3] Han, Y., Kim, J., & Enke , D. (2023). A machine learning trading system for the stock market based on N-period Min-Max labeling using XGBoost . Expert Systems with Applications, 211, 118581. [4] Hu, Z., Dong, Y., Wang, K., & Sun, Y. (2020, April). Heterogeneous graph transformer. In Proceedings of the web conference 2020 (pp. 2704-2710). [5] Pareja, A., Domeniconi , G., Chen, J., Ma, T., Suzumura , T., Kanezashi , H., ... & Leiserson , C. (2020, April). Evolvegcn : Evolving graph convolutional networks for dynamic graphs. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 04, pp. 5363-5370). [6] Fan, Y., Ju, M., Zhang, C., & Ye, Y. (2022). Heterogeneous temporal graph neural network. In Proceedings of the 2022 SIAM International Conference on Data Mining (SDM) (pp. 657-665). Society for Industrial and Applied Mathematics. Measurement : Information Coefficient (IC): overall ranking performance. Information Ratio (IR): divides the excess return of a portfolio by its tracking error. Cumulative Return (CR): accumulated portfolio return based on the prediction score. Precision@K : whether the excess returns of TopK stocks outperform the benchmark index.

14. EXPERIMENT AND RESULT RESULT – Overall Perfor mance

15. EXPERIMENT AND RESULT Result: Case Study Visualization A research report on investment bank holdings reveals a rising credit pulse trend: An increase in proportion of bank holdings by international investment institutions. To analyze, focus on Chengdu Bank (601838.SH) and its subgraph: Chengdu Bank(601838.SH), Nanjing Bank (601009.SH), Jiangsu Bank(600919.SH), and Vanke A (000002.SZ). Fig(b): average stock change rates for 4 selected stocks. Fig(c): average stock change rates of three key industries.

16. CONCLUSION F ormally define the multifacetedness and temporal patterns of stocks through empirical analysis for the first time. P ropose a novel hierarchical multi relational dynamic graph framework for modeling stock investment prediction C onstructing a multi-relational graph for each trading day and generating a set of discrete graph snapshots within the specified lookback window size. For intra-day graph snapshot, design a hierarchical multi-relational graph embedding layer to aggregate neighbor nodes within a specific meta-path and adaptively integrate the stock representation from distinct meta-paths. I ncorporate transformer structure to aggregate the temporal evolving patterns of stocks. Summarization