PaCMoE: Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts for Stock Movement Prediction
- 1School of Informatics, Kogakuin University, Tokyo 163-8677, Japan
- 2Graduate School of Engineering, Kogakuin University, Tokyo 163-8677, Japan
Abstract
Stock movement prediction remains challenging because financial time series are nonstationary, noisy, and interdependent across heterogeneous information sources. Conventional multimodal forecasting methods often integrate price, text, and inter-firm relationships into a single representation, making it difficult to capture modality-pair-specific interactions and their varying relevance across market conditions. To address this issue, this study proposes Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts (PaCMoE), which explicitly decomposes multimodal interactions into three modality pairs—Price–Text, Price–Graph, and Text–Graph. Each pair is modeled by an independent Cross-Modal Attention expert, and a Dense Mixture-of-Experts router adaptively integrates the resulting representations according to the input. This design enables separate modeling of pairwise interactions and input-dependent weighting of their contributions. The modality encoders combine PatchTST for prices, FinBERT with canonical correlation analysis for text, and a single-relation R-GCN for daily correlation graphs. PaCMoE is evaluated using chronological data splitting, validation-based model selection, and ten random seeds on the ACL18 (StockNet) benchmark. The proposed method achieves an average accuracy of 60.69% and an MCC of 0.2200, which are comparable to or higher than the multi-seed values reported for LSTM-based Adv-ALSTM and Transformer-based DTML. Additional evaluations include single-run ablation studies, three-class prediction incorporating neutral days, precision–recall analysis, and visualizations of router and attention weights to examine component contributions and model behavior. Training takes approximately 30 min per run on a single NVIDIA GeForce RTX 3090. These results support the effectiveness of explicitly modeling pairwise cross-modal interactions and adaptively integrating their contributions for stock movement prediction on the ACL18 benchmark.
Related
- Mixture of Censored Partially Linear Experts Models for Symmetric or Asymmetric Data with or without Outlying Observations: A novel approach with two-piece scale-mixture of normal distributions
- The Limits of Speculation: Bounding Speculative Decoding in Mixture-of-Experts
- Non-asymptotic oracle inequalities for the Lasso in high-dimensional mixture of experts
- A Mixture-of-Experts Framework for Multi-Stream Quickest Change Detection
- When Does Sparse Routing Help? The Geometry of Expert Separation in Mixture of Experts
BibTeX
@article{nagata2026pacmoe,
title = {{PaCMoE: Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts for Stock Movement Prediction}},
author = {Eiji Nagata and Riku Motomochi and Qiu Chen},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14213948},
url = {https://www.mdpi.com/2227-7390/14/21/3948},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - PaCMoE: Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts for Stock Movement Prediction
AU - Eiji Nagata
AU - Riku Motomochi
AU - Qiu Chen
PY - 2026
DA - 2026/10/29
JO - Mathematics
DO - 10.3390/math14213948
UR - https://www.mdpi.com/2227-7390/14/21/3948
AB - Stock movement prediction remains challenging because financial time series are nonstationary, noisy, and interdependent across heterogeneous information sources. Conventional multimodal forecasting methods often integrate price, text, and inter-firm relationships into a single representation, making it difficult to capture modality-pair-specific interactions and their varying relevance across market conditions. To address this issue, this study proposes Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts (PaCMoE), which explicitly decomposes multimodal interactions into three modality pairs—Price–Text, Price–Graph, and Text–Graph. Each pair is modeled by an independent Cross-Modal Attention expert, and a Dense Mixture-of-Experts router adaptively integrates the resulting representations according to the input. This design enables separate modeling of pairwise interactions and input-dependent weighting of their contributions. The modality encoders combine PatchTST for prices, FinBERT with canonical correlation analysis for text, and a single-relation R-GCN for daily correlation graphs. PaCMoE is evaluated using chronological data splitting, validation-based model selection, and ten random seeds on the ACL18 (StockNet) benchmark. The proposed method achieves an average accuracy of 60.69% and an MCC of 0.2200, which are comparable to or higher than the multi-seed values reported for LSTM-based Adv-ALSTM and Transformer-based DTML. Additional evaluations include single-run ablation studies, three-class prediction incorporating neutral days, precision–recall analysis, and visualizations of router and attention weights to examine component contributions and model behavior. Training takes approximately 30 min per run on a single NVIDIA GeForce RTX 3090. These results support the effectiveness of explicitly modeling pairwise cross-modal interactions and adaptively integrating their contributions for stock movement prediction on the ACL18 benchmark.
SN - 2227-7390
ER -