DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction

Weijun Chen, Yanze WangAAAI 2025 · DOI

TL;DR

This paper introduces DHMoE, a novel hierarchical Mixture-of-Experts (MoE) model for stock prediction that emulates real-world trading firm workflows. By combining top-down teaching with bottom-up decision-making and leveraging diffusion models for expert parameter generation, DHMoE achieves superior cumulative and risk-adjusted returns compared to state-of-the-art methods.

What Problem Does This Paper Address?

Accurate stock prediction is crucial for investors, but the dynamic and noisy nature of financial markets, coupled with the complexity of integrating diverse data sources (like news and relational data), poses significant challenges. Existing deep learning models often rely on a single predictor or fail to effectively capture the hierarchical decision-making and knowledge transfer processes found in successful quantitative trading firms. These firms typically employ a top-down teaching approach where experienced experts guide less experienced ones, and a bottom-up decision-making process where individual analyses are synthesized into a final verdict. This paper aims to bridge this gap by designing a model that mimics these human-centric, hierarchical processes for more robust and profitable stock prediction.

Methods

The core innovation of DHMoE lies in its hierarchical, three-layer tree structure and its emulation of real-world trading paradigms.

  • Diffusion Generated Hierarchical Mixture-of-Experts (DHMoE): This is the overarching framework. It's structured as a tree where each node represents an expert.
  • Top-Down Teaching/Parameter Generation: The model uses a top-down approach to generate parameters for the experts.
  • Diffusion Inverted Transformer (DIT): This is the root expert. It leverages the generative capabilities of diffusion models to process multimodal inputs (stock time series, news, relational graphs) and generate conditional parameters for the lower-level experts. The DIT's ability to handle noise and stochasticity is highlighted as beneficial for stock market dynamics.
  • Conditional Diffusion: The DIT adapts diffusion models to generate parameters conditionally based on the input data, moving beyond unconditional generation.
  • Multimodal Embedding: Stock time series, news text (processed by BERT), and stock relational graphs are embedded into a unified representation to be used as conditional information.
  • Bottom-Up Decision-Making: This process synthesizes the predictions from the experts.
  • Middle-level and Bottom-level Experts: These experts are specialized for particular modalities. Bottom-level experts use lightweight FourierMLPs for efficient processing.
  • Sparse Synthesized Attention Mechanism (SSAM): This mechanism consolidates predictions from bottom-level experts to middle-level experts, and then from middle-level to the final top-level prediction. This hierarchical aggregation mirrors the decision-making process in trading firms.

Datasets

The paper evaluates DHMoE on three real-world datasets:

  • NASDAQ: Comprises stock news and data for 112 high-capitalization stocks.
  • NYSE: Includes news and data for 127 stocks listed on the New York Stock Exchange.
  • Ashare&HK: Contains Chinese news headlines and data for 80 top-traded stocks from Shanghai, Shenzhen, and Hong Kong exchanges.

The datasets cover a significant period, with training, validation, and testing splits defined.

Key Insights and Ideas

  • Emulating Human Trading Workflows: The most significant insight is that mimicking the hierarchical teaching and decision-making processes of human trading firms can lead to more effective stock prediction models. This moves beyond purely data-driven approaches.
  • Diffusion Models for Expert Parameter Generation: Using diffusion models as the top-level expert to generate parameters for lower-level experts is a novel application. This leverages their powerful generative and noise-handling capabilities for robust feature extraction and parameter allocation.
  • Hierarchical Expertise for Modality Specialization: The three-layer tree structure allows for specialized experts at different levels, ensuring that each modality (time series, news, relations) is processed effectively by dedicated components.
  • Bottom-Up Aggregation is Key: The ablation studies suggest that the bottom-up decision-making paradigm, which hierarchically aggregates expert opinions, provides the most significant performance gains, highlighting its importance for model robustness.

Key Results

DHMoE consistently outperforms state-of-the-art baselines across all three datasets in terms of both cumulative investment return ratio (IRR) and Sharpe Ratio (SR).

  • Overall Performance: DHMoE achieved an average relative performance gain of 17.37% in risk-adjusted returns and 20.08% in cumulative profits over the best baselines.
  • Ablation Study: Removing top-down teaching or bottom-up decision-making components led to performance degradation, with bottom-up decision-making showing a more substantial impact.
  • DIT's Superiority: Replacing the DIT with other parameter generation models (Transformer, iTransformer) resulted in lower performance, underscoring the benefits of the diffusion-based approach for the top-level expert. For instance, on NASDAQ, DHMoE achieved an SR of 1.24 and IRR of 0.64, compared to 1.07/0.49 for Transformer and 1.14/0.55 for iTransformer.

Limitations and Future Work

  • Parameter Sensitivity: The paper notes that the weighting coefficients for the loss function (λ and α) are crucial and can disrupt performance if set too high or too low, indicating a need for careful tuning.
  • Expert Number Sensitivity: While more experts generally improve performance, there's a point of diminishing returns, and the optimal number of middle-level and bottom-level experts needs to be determined.
  • Computational Cost: While the model is designed for efficiency, diffusion models and MoE architectures can still be computationally intensive, especially with a large number of experts.
  • Future Work: The authors suggest further exploration of diffusion models for generating parameters and the potential for extending the hierarchical structure to more complex scenarios. The flexibility of the MoE architecture also allows for easier scaling and adaptation to new data sources or prediction tasks.

Summary updated

BibTeX
@inproceedings{chen2025dhmoe,
  author = {Weijun Chen and Yanze Wang},
  title = {{DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i11.33250},
}
Figure