CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction
Authors: Shuqi Li, Yuebo Sun, Yuxin Lin, Xin Gao, Shuo Shang, Rui Yan
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiment results show that Causal Stock outperforms the strong baselines for both news-driven multi-stock movement prediction and multi-stock movement prediction tasks on six real-world datasets collected from the US, China, Japan, and UK markets. |
| Researcher Affiliation | Academia | 1 Gaoling School of Artificial Intelligence, Renmin University of China 2 Peking University 3 King Abdullah University of Science and Technology 4 University of Electronic Science and Technology of China |
| Pseudocode | No | The paper describes the model and its components using mathematical equations and text, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Answer: [Yes] Justification: We release code and data in Git Hub. |
| Open Datasets | Yes | Dataset (Appendix C.1): We train and evaluate our model and baselines on six datasets: ACL18 [42], CMIN-US [23], CMIN-CN [23], KDD17 [45], NI225 [44], and FTSE100 [44]. |
| Dataset Splits | Yes | Table 3: Dataset Description ... ACL18 ... 2014/01/02-2015/08/02 (Train) 2015/08/03-2015/09/30 (Valid) 2015/10/01-2016/01/01 (Test) |
| Hardware Specification | Yes | Our model is implemented with Pytorch on 4 NVIDIA Tesla V100 and optimized by Adam [20]. |
| Software Dependencies | No | The paper mentions "Pytorch" as the implementation framework but does not specify a version number or other software dependencies with version numbers. |
| Experiment Setup | Yes | The learning rate is set as 1e 5 selected from [1e 3, 1e 4, 1e 5, 1e 6]. The time lag L is set as 5 selected from [3, 5, 7, 9]. We select the price encoder hidden size from [4, 8, 16] and get the best performance with size 4. The batch size is set as 32. The scalar weight λ is set to 0.01. For the traditional news encoder, the maximum word number in one piece of news and news number in one day are set to w = 20, l = 10, respectively. The embedding size of word and news are set to dw = 50, dm = 64, respectively. For the Lag-dependent temporal causal discovery module, λs = 1, hv and hu are all 1-layer MLPs. For the FCM part, the neural modules ζi, ℓand ψ are all 3-layer MLPs with hidden size 332. |