Research map / Probabilistic and scientific modeling / Causal discovery
CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CausalStock, a news-driven multi-stock movement predictor that learns lag-dependent temporal causal graphs with variational inference and an additive-noise functional causal model, using LLM-scored news denoising.
Problem
News-driven multi-stock movement prediction, where stock relations are modeled only as correlations and news text is noisy, so useful signals are hard to extract.
Approach
Discovers directed, lag-dependent temporal causal graphs among stocks end-to-end (variational Bernoulli posterior with sparsity and domain-knowledge priors), feeds them into a functional causal model for prediction, and encodes news via LLM-generated multi-dimensional scores (correlation, sentiment, importance, impact, duration) as denoised representations.
Key results
- On news-driven tasks, CausalStock reaches 63.42% ACC / 0.2172 MCC on ACL18, versus 62.69% / 0.2090 for the best baseline CMIN.
- On KDD17 (no news), CausalStock achieves 56.09% ACC / 0.1235 MCC, above the best baseline DTML at 53.53% / 0.0733.
- Removing the causal discovery module (w/o TCD) drops ACL18 accuracy to 51.08%, showing the causal graph is the main contributor.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal causal discovery (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 51.08 vs 63.42; 51.48 vs 54.64; 51.37 vs 56.19 | Helped |
| add the news encoder (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 58.10 vs 63.42; 53.16 vs 54.64; 54.16 vs 56.19 | Helped |
| model link non-existence likelihood (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 58.21 vs 63.42; 52.32 vs 54.64; 53.96 vs 56.19 | Helped |
| use lag-dependent rather than lag-independent TCD (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 59.19 vs 63.42; 52.93 vs 54.64; 54.97 vs 56.19 | Helped |
| add a variable-dependent causal mechanism (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 63.50 vs 63.42; 54.60 vs 54.64; 56.25 vs 56.19 | Mixed |
| replace denoised news encoder with Glove+Bi-GRU (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 60.78 vs 63.42; 53.87 vs 54.64; 55.13 vs 56.19 | Hurt |
| replace denoised news encoder with Bert (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 61.74 vs 63.42; 53.92 vs 54.64; 55.43 vs 56.19 | Hurt |
| replace denoised news encoder with Roberta (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 61.81 vs 63.42; 54.06 vs 54.64; 55.58 vs 56.19 | Hurt |
| replace denoised news encoder with FinBert (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 61.72 vs 63.42; 54.01 vs 54.64; 55.61 vs 56.19 | Hurt |
| replace denoised news encoder with FinGPT as a traditional news encoder (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 61.69 vs 63.42; 54.00 vs 54.64; 55.60 vs 56.19 | Hurt |
| replace denoised news encoder with Llama as a traditional news encoder (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 62.20 vs 63.42; 54.40 vs 54.64; 55.85 vs 56.19 | Hurt |
| use FinGPT as the denoised news encoder (ablation) | CausalStock · ACL18, CMIN-US, CMIN-CN | ACC: 61.92 vs 63.42; 54.30 vs 54.64; 55.67 vs 56.19 | Hurt |
Surprising findings
- Variable-dependent causal mechanism gives mixed gains across datasets and raises complexity from O(L×D^2) to O(L×D^4).
- Traditional news encoders (Glove+Bi-GRU, BERT, RoBERTa, FinBERT, FinGPT, Llama) are only modestly below the LLM-denoised encoder, with Llama-based traditional encoding close to the denoised version.
Limitations
- Causal graph assumed time-invariant (stationary); time-varying causal structure left to future work.
- Bernoulli link distribution may be insufficient for multi-level causal relationships.
- Gains from the variable-dependent extension are inconsistent across datasets, and the model struggles to scale to large stock universes.
Experimental setup
Datasets: ACL18 CMIN-US CMIN-CN KDD17 NI225 FTSE100
Benchmarks: ACL18 CMIN-US CMIN-CN KDD17 NI225 FTSE100
Metrics: ACC MCC APV Sharpe Ratio
Baselines: HAN StockNet PEN CMIN LSTM ALSTM Adv-ALSTM DTML
Builds on: DECI Rhino CMIN Adv-ALSTM FCM
Closest papers by meaning
- ECHO-GL: Earnings Calls-Driven Heterogeneous Graph Learning for Stock Movement Prediction (AAAI 2024 · similarity 0.85)
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.84)
- Towards the Disappearing Truth: Fine-Grained Joint Causal Influences Learning with Hidden Variable-Driven Causal Hypergraphs in Time Series (AAAI 2024 · similarity 0.84)
- MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction (AAAI 2024 · similarity 0.84)
- Causal Discovery from Event Sequences by Local Cause-Effect Attribution (NeurIPS 2024 · similarity 0.83)
- CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.