Research map / Time series / Foundation models for time series
CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph
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 CI-STHPAN, a two-stage Transformer and hypergraph attention framework that pre-trains on stock time series with masked patch reconstruction and fine-tunes for stock ranking, using channel-independent DTW-based dynamic hypergraphs.
Problem
Quantitative stock selection on NASDAQ and NYSE faces non-stationary dynamics, distribution shift from channel mixing, difficulty with very long sequences, and reliance on predefined static stock relationships.
Approach
Combines patching, channel independence, and RevIN with a Transformer encoder pre-trained via masked patch reconstruction, followed by a hypergraph attention network over channel-wise DTW similarity hypergraphs, and fine-tuned with a regression plus pairwise ranking loss.
Key results
- On NASDAQ, CI-STHPAN reaches IRR 0.66 and SR 2.01, versus 0.53 IRR and 1.55 SR for the best baseline ALSP-TF.
- On NYSE, CI-STHPAN reaches IRR 0.79 and SR 2.14, versus 0.41 IRR and 1.24 SR for ALSP-TF.
- Transfer learning: pre-training on NASDAQ and fine-tuning on NYSE gives IRR 0.694 and SR 1.928 on NYSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the pretrained model rather than train with supervised learning (ablation) | CI-STHPAN · NYSE | IRR: 0.077 -> 0.789 | Helped |
| add channel independence (ablation) | CI-STHPAN with graph information and fine-tuning · NASDAQ | IRR: 0.293 -> 0.657 | Helped |
| add RevIN (ablation) | CI-STHPAN with graph information and fine-tuning · NYSE | IRR: 0.089 -> 0.488 | Helped |
| add channel independence and RevIN (ablation) | CI-STHPAN with graph information and supervised learning · NYSE | IRR: 0.175 -> 0.077 | Hurt |
| add DTW dynamic stock relations (ablation) | CI-STHPAN fine-tuning on NASDAQ · NASDAQ | IRR: 0.469 -> 0.657 | Helped |
| replace DTW stock relations with Wiki relations (ablation) | CI-STHPAN fine-tuning on NASDAQ · NASDAQ | IRR: 0.657 -> 0.591 | No clear effect |
| change DTW graph sparsity from K=20 to K=15 (ablation) | CI-STHPAN supervised on NASDAQ · NASDAQ | IRR: 0.491 -> 0.553 | Helped |
| change DTW graph sparsity from K=20 to K=15 (ablation) | CI-STHPAN fine-tuning on NASDAQ · NASDAQ | IRR: 0.657 -> 0.281 | Hurt |
| pre-train on NASDAQ instead of NYSE, then fine-tune on NYSE (ablation) | CI-STHPAN · NYSE | IRR: 0.789 -> 0.273 | Hurt |
| pre-train on NASDAQ instead of NYSE, then fine-tune on NYSE (ablation) | CI-STHPAN · NYSE | ICIR: 0.569 -> 0.457 | Hurt |
| use CI-STHPAN instead of ALSP-TF | stock selection · NYSE | IRR: 0.41 -> 0.79 | Helped |
| use CI-STHPAN instead of ALSP-TF | stock selection · NASDAQ | SR: 1.55 -> 2.01 | Helped |
Surprising findings
- Even without graph information, self-supervised Transformer pre-training with fine-tuning surpasses SOTA results on NYSE (IRR 0.479 supervised-only vs. fine-tuned variants, per ablation).
- Predefined industry/wiki graphs and combined graphs sometimes underperform the no-graph variant, while DTW-based dynamic hypergraphs perform better.
Limitations
- Relies on DTW similarity hypergraphs whose hyperparameters (K, patch length, loss factor) are selected by grid search on validation loss.
- Results on a single period (2017 test) and two markets; ablation results are mixed across settings (e.g., some fine-tuned variants underperform).
Experimental setup
Datasets: NASDAQ stock data 2013-2017 NYSE stock data 2013-2017 Wikidata relations sector-industry relations
Benchmarks: NASDAQ NYSE
Metrics: IRR Sharpe ratio IC ICIR
Baselines: ALSP-TF STHAN-SR RAT RSR-I iRDPG DQN HATR HMG-TF Adv-LSTM HATS
Builds on: PatchTST STHAN-SR Transformer encoder hypergraph convolution RSR
Closest papers by meaning
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.87)
- FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction (AAAI 2025 · similarity 0.86)
- MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction (AAAI 2024 · similarity 0.84)
- StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting (AAAI 2024 · similarity 0.84)
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.84)
- Towards Financial World Modeling (NeurIPS 2026 · similarity 0.84)
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.