Paper: arXiv 2501.09636
Authors: Kuan-Ming Liu, Ming-Chih Lo
Abstract
Recent advances in deep learning and large language models (LLMs) have facilitated the deployment of the mixture-of-experts (MoE) mechanism in the stock investment domain. While these models have demonstrated promising trading performance, they are often unimodal, neglecting the wealth of information available in other modalities, such as textual data. Moreover, the traditional neural network-based router selection mechanism fails to consider contextual and real-world nuances, resulting in suboptimal expert selection. To address these limitations, we propose LLMoE, a novel framework that employs LLMs as the router within the MoE architecture. Specifically, we replace the conventional neural network-based router with LLMs, leveraging their extensive world knowledge and reasoning capabilities to select experts based on historical price data and stock news. This approach provides a more effective and interpretable selection mechanism. Our experiments on multimodal real-world stock datasets demonstrate that LLMoE outperforms state-of-the-art MoE models and other deep neural network approaches. Additionally, the flexible architecture of LLMoE allows for easy adaptation to various downstream tasks.
Complexity vs Empirical Score
- Math Complexity: 3.5/10
- Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
- Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)
Why this score: The paper involves advanced ML architecture (MoE with LLM router) but the mathematics is presented at a conceptual level without dense derivations, keeping complexity moderate. Empirical rigor is high due to the use of real-world datasets (MSFT, AAPL), detailed feature engineering, backtesting metrics (Sharpe, Sortino, MDD, etc.), and multiple baselines.
Research Flowchart
flowchart TD A["Research Goal<br>Improve MoE Routing for Stock Investment<br>Integrating Multimodal Data & Context"] --> B["Data Collection<br>Historical Price & Stock News"] B --> C["Framework: LLMoE<br>LLM-as-Router replaces NN Router"] C --> D["Input Processing<br>Multimodal Features"] D --> E["LLM Reasoning & Routing<br>Select Experts based on Context"] E --> F["Expert Execution<br>Specialized Models Process Data"] F --> G["Output Prediction<br>Stock Prediction / Trading Signal"] G --> H["Key Findings<br>Outperforms SOTA MoE & DNNs<br>Better Interpretability & Flexibility"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes LLMoE, a novel framework that uses large language models as routers within a mixture-of-experts architecture for stock trading.
Source sentence
we propose LLMoE, a novel framework that employs LLMs as the router within the MoE architecture.
Assumes 3
- Trading strategy invests all cash on positive predictions and liquidates all holdings on negative predictions (All-in All-out). execution
Source sentence
We employ an ”All-in All-out” strategy, where all available cash is invested when the expert predicts positive price movements, and all holdings are liquidated when the expert predicts negative price movements.
- Historical stock prices and news headlines provide sufficient multimodal information for prediction. data
Source sentence
LLMoE processes historical stock prices and news headlines through an LLM-based router, which provides a comprehensive overview of the current instance.
- The LLM router can effectively reason over combined numerical and textual inputs to select experts. model
Source sentence
leveraging their extensive world knowledge and reasoning capabilities to select experts based on historical price data and stock news.
Data
MSFT stocks · equities · daily
- MSFT dataset market-prices unknown
Source sentence
The MSFT dataset, with significant missing news days, presents a challenging test for handling incomplete data
- AAPL dataset market-prices unknown
Source sentence
the AAPL dataset provides a more complete environment
Costs
Not modeled
Finds 2
Sharpe 2.14 · total return 65.4% · max drawdown 26.1%
- LLMoE significantly outperforms baseline models on Total Return, Sharpe Ratio, and Calmar Ratio. performance
Source sentence
Our LLMoE model significantly outperformed other baseline models in key metrics, including Total Return (TR), Sharpe Ratio (SR), and Calmar Ratio (CR)
- LLMoE achieves the highest Total Return (31.43±11.46), Sharpe Ratio (1.17±0.33), and Calmar Ratio (2.12±1.10) on the MSFT dataset. performance
Source sentence
LLMoE 31.43±11.46 1.17±0.33 2.12±1.10
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 67 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive : MM-DREX: Multimodal-Driven Dynamic Routing of LLM Experts …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (15)
- Mastering Stock Markets with Efficient Mixture of Diversified Trading Experts (2023) — link
- Accurate Multivariate Stock Movement Prediction via Data-Axis Transformer with Multi-Level Contexts (2021) — link
- Deep Attentive Learning for Stock Movement Prediction From Social Media Text and Company Correlations (2020) — link
- A Deep Learning Framework for News-oriented Stock Trend Prediction (2018) — link
- MIGA: Mixture-of-Experts with Group Aggregation for Stock Market Prediction (2024) — link
- An Attention-Based Multi-Gate Mixture-of-Experts Model for Quantitative Stock Selection (2023) — link
- Automate Strategy Finding with LLM in Quant investment — → 2409.06289
- Artificial Intelligence in Accounting and Finance: Challenges and Opportunities (2023) — link
and 7 more
- VWAP execution as an optimal strategy (2015) — link
- Are Transformers Effective for Time Series Forecasting? (2022) — link
- An Analysis of Linear Time Series Forecasting Models (2024) — link
- REST: Relational Event-driven Stock Trend Forecasting (2021) — link
- Accurate Multivariate Stock Movement Prediction via Data-Axis Transformer with Multi-Level Contexts (2021) — link
- TradExpert: Revolutionizing Trading with Mixture of Expert LLMs — → 2411.00782
- Optimization of Traditional Stock Market Strategies Using the LSTM Hybrid Approach (2024) — link
2 of 15 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.