Research map / Image generation / Fast diffusion sampling
DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock 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 DHMoE, a three-layer tree mixture-of-experts for multimodal stock ranking whose root expert is a conditional diffusion-based Inverted Transformer that generates parameters for lower-level experts, aggregated bottom-up via sparse synthesized attention.
Problem
Single neural predictors for stock prediction adapt poorly to diverse market conditions and struggle to integrate multi-source information (time series, news, stock relations).
Approach
Emulate top-down teaching and bottom-up decision-making of quantitative firms with a hierarchical MoE tree in which a diffusion-based Inverted Transformer (DIT) root generates expert parameters top-down from multimodal conditions, and a sparse synthesized attention mechanism (SSAM) aggregates expert predictions bottom-up.
Key results
- DHMoE achieves SR 1.24 / IRR 0.64 on NASDAQ, 1.08 / 0.46 on NYSE, and 1.17 / 0.45 on Ashare&HK, the best among all compared methods.
- Reports an average relative gain of 17.37% in risk-adjusted returns and 20.08% in cumulative profits over the best baselines.
- Replacing the DIT root with a vanilla Transformer or iTransformer lowers results (e.g., NASDAQ SR 1.07 and 1.14 vs 1.24).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DHMoE instead of STHAN-SR | stock ranking · NASDAQ test | SR: 1.10 -> 1.24 | Helped |
| use DHMoE instead of STHAN-SR | stock ranking · NASDAQ test | IRR: 0.53 -> 0.64 | Helped |
| use DHMoE instead of THGNN | stock ranking · NYSE test | SR: 0.90 -> 1.08 | Helped |
| use DHMoE instead of THGNN | stock ranking · NYSE test | IRR: 0.38 -> 0.46 | Helped |
| use DHMoE instead of AlphaMix | stock ranking · Ashare&HK test | SR: 0.98 -> 1.17 | Helped |
| use DHMoE instead of AlphaMix | stock ranking · Ashare&HK test | IRR: 0.38 -> 0.45 | Helped |
| replace DIT with Transformer for parameter generation (ablation) | DHMoE · NASDAQ | SR: 1.07 -> 1.24 | Helped |
| replace DIT with iTransformer for parameter generation (ablation) | DHMoE · NASDAQ | IRR: 0.55 -> 0.64 | Helped |
| replace DIT with Transformer for parameter generation (ablation) | DHMoE · NYSE | SR: 0.92 -> 1.08 | Helped |
| replace DIT with iTransformer for parameter generation (ablation) | DHMoE · NYSE | IRR: 0.41 -> 0.46 | Helped |
| replace DIT with Transformer for parameter generation (ablation) | DHMoE · Ashare&HK | SR: 0.91 -> 1.17 | Helped |
| remove top-down parameter transfer and independently initialize bottom-level experts (ablation) | DHMoE · NASDAQ, NYSE, and Ashare&HK | investment return ratio (IRR) | Helped |
Surprising findings
- The bottom-up decision-making (SSAM aggregation) contributes the primary gain in ablations, larger than the top-down teaching or the DIT root alone.
Limitations
- Evaluation limited to three stock datasets over a single test period per dataset.
- Relies on diffusion and multimodal conditioning, adding complexity and hyperparameter sensitivity (expert counts, loss weights, diffusion steps).
Experimental setup
Datasets: NASDAQ NYSE Ashare&HK
Benchmarks: NASDAQ NYSE Ashare&HK
Metrics: Sharpe Ratio investment return ratio nDCG@5
Baselines: Adv-ALSTM StockNet MAN-SF LSTM-RGCN SFM THGNN AMMoE AlphaMix iRDPG RAT RSR STHAN-SR
Builds on: Mixture-of-Experts DDPM iTransformer FourierMLP SGC AlphaMix
Closest papers by meaning
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.85)
- StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting (AAAI 2024 · similarity 0.85)
- Filtered not Mixed: Filtering-Based Online Gating for Mixture of Large Language Models (ICLR 2025 · similarity 0.85)
- 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)
- Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking (NeurIPS 2025 · 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.