Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction

Authors: Heyuan Wang, Shun Li, Tengjiao Wang, Jiayi Zheng

IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three real-world stock market datasets validate the effectiveness of our model.
Researcher Affiliation Academia 1School of EECS, Peking University 2University of International Relations 3Institute of Computational Social Science, Peking University(Qingdao)
Pseudocode No No pseudocode or clearly labeled algorithm blocks were found in the paper.
Open Source Code No The paper does not provide any explicit statement about releasing open-source code or a link to a code repository for the methodology described.
Open Datasets Yes We verify HATR on three real-world datasets for comprehensive evaluation. Table 1 shows the detailed statistics. The first dataset comprises stocks from the well-know CSI-300 Composite Index... The second dataset is targeted at stocks from popular S&P500 Composite Index... The third dataset [Li et al., 2020] is from Tokyo Stock Exchange, including 95 stocks with the largest market capitalization in Japan from the TOPIX-100 Index. We collect the daily quote data, industry and capital information from Wind-Financial Terminal1. To mine topicality relations, we detect co-occurrence stock pairs in user reviews from a popular Chinese investment forum Xueqiu2 for the CSI dataset, and collect first- and second-order linkages from Wikidata3 for the SPX and Topix datasets.
Dataset Splits Yes The training/validation/test sets are strictly split in chronological order to avoid data leakage problems. Table 1: # Split Records 683:171:139 (CSI), 787:197:169 (SPX), 814:204:144 (Topix).
Hardware Specification Yes Parameters are tuned using Adam optimizer [Kingma and Ba, 2014] on a single NVIDIA Titan Xp GPU for 100 epochs, the learning rate is 0.0005 and the batch size is 200.
Software Dependencies No The paper mentions 'Adam optimizer' but does not specify version numbers for any programming languages or software libraries used for implementation.
Experiment Setup Yes In our experiments, a 4-layer stacking hierarchy with the dilation list of {1-2-3-4} is employed for temporal representations. The window size and the number of gated convolution kernels at each layer were set to 3 and 32. The dimensions of randomly initialized stock ID embeddings and node embeddings were set to 20 for Topix and 30 for CSI and SPX, the target-specific query for attending to important temporal scales has a dimension of 16. The finite step K for graph diffusions is set to 2. We apply dropout [Srivastava et al., 2014] at the end of each layer to mitigate overfitting and the drop rate is 0.3. Parameters are tuned using Adam optimizer [Kingma and Ba, 2014] on a single NVIDIA Titan Xp GPU for 100 epochs, the learning rate is 0.0005 and the batch size is 200.