Paper: arXiv 2411.00782

Authors: Qianggang Ding, Haochen Shi, Jiadong Guo, Bang Liu

Abstract

The integration of Artificial Intelligence (AI) in the financial domain has opened new avenues for quantitative trading, particularly through the use of Large Language Models (LLMs). However, the challenge of effectively synthesizing insights from diverse data sources and integrating both structured and unstructured data persists. This paper presents TradeExpert, a novel framework that employs a mix of experts (MoE) approach, using four specialized LLMs, each analyzing distinct sources of financial data, including news articles, market data, alpha factors, and fundamental data. The insights of these expert LLMs are further synthesized by a General Expert LLM to make a final prediction or decision. With specific prompts, TradeExpert can be switched between the prediction mode and the ranking mode for stock movement prediction and quantitative stock trading, respectively. In addition to existing benchmarks, we also release a large-scale financial dataset to comprehensively evaluate TradeExpert’s effectiveness. Our experimental results demonstrate TradeExpert’s superior performance across all trading scenarios.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; reported Sharpe ratio of 5.01 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)

Why this score: The paper primarily focuses on engineering a complex LLM-based framework with prompting and fine-tuning, which is architecturally heavy but mathematically lightweight; however, it demonstrates high empirical rigor by releasing a large-scale dataset, reporting specific backtesting metrics (Annualized Return, Sharpe Ratio, etc.), and conducting ablation studies.

Research Flowchart

  flowchart TD
  Goal["Research Goal: Improve quantitative trading<br>using LLMs on structured/unstructured data"] --> Data["Data Inputs:<br>News, Market Data, Alpha Factors, Fundamentals"]
  Data --> Strategy["Methodology: MoE Framework<br>4 Expert LLMs + 1 General Expert"]
  Strategy --> Mode["Switch Modes via Prompting:<br>Prediction Mode or Ranking Mode"]
  Mode --> Process["Computational Process:<br>Experts analyze specific data, General Expert synthesizes"]
  Process --> Outcome["Key Findings: TradeExpert shows<br>superior performance across trading scenarios"]

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 TradExpert, a novel Mixture-of-Experts framework that integrates four specialized LLMs and a General Expert LLM for stock movement prediction and trading.

Source sentence
We present TradExpert, a novel framework that employs an MoE approach, integrating four LLMs each specialized to analyze distinct sources of financial data, imitating the structured division of labors seen in the real world.

Assumes 2

  • Backtesting is performed over the calendar year 2023 using the same period for testing. execution
    Source sentence
    The backtesting period is the same as the testing period of our datasets, which ranges from January 1, 2023, to December 31, 2023.
  • Trading decisions are based on the Top‑K ranked stocks without accounting for slippage or market impact. execution
    Source sentence
    We implement a Buy-and-Hold trading strategy on the Top‑K stocks ranked by TradExpert.

Data

Backtest sample: 2023-01 – 2023-12 · DOW 30 · equities · daily

  • CIKM18 benchmark public
    Source sentence
    Benchmark Datasets: We use publicly available benchmark datasets in stock movement prediction research including CIKM18 Wu et al. (2018), ACL18 Xu & Cohen (2018), and BigData22 Soun et al. (2022) datasets.
  • ACL18 benchmark public
    Source sentence
    Benchmark Datasets: We use publicly available benchmark datasets in stock movement prediction research including CIKM18 Wu et al. (2018), ACL18 Xu & Cohen (2018), and BigData22 Soun et al. (2022) datasets.
  • BigData22 benchmark public
    Source sentence
    Benchmark Datasets: We use publicly available benchmark datasets in stock movement prediction research including CIKM18 Wu et al. (2018), ACL18 Xu & Cohen (2018), and BigData22 Soun et al. (2022) datasets.
  • Proprietary dataset (OHLCV, news, alpha, fundamentals) other proprietary
    Source sentence
    Proprietary Datasets: We also utilize our proprietary datasets, which include extensive historical OHLCV data, news articles, alpha factors, and fundamental metrics for a comprehensive analysis.
  • DOW 30 backtest universe market-prices public
    2023 · DOW 30
    Source sentence
    To reduce computational costs in backtesting, we evaluated all methods on datasets with stocks on the DOW 30 list, a subset of the S&P 500, with around 30 stocks.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

Reported performance (TradExpert): Sharpe 5.01 (benchmark 1.22) · annual return 49.8% · max drawdown 6.6%

  • TradExpert achieved an Annualized Return of 49.79%, Annualized Volatility of 9.95%, and Sharpe Ratio of 5.01 on the DOW 30 backtest, outperforming all baselines. performance
    Source sentence
    TradExpert, our proposed model, significantly outperformed all other models with an exceptional AR of 49.79% and the lowest AV of 9.95%. This combination yielded an outstanding Sharpe ratio of 5.01, indicating a high return per unit of risk.
  • In stock movement prediction, TradExpert-NM consistently outperformed other models across all datasets except for MCC on the ACL18. predictive-accuracy
    Source sentence
    Our proposed TradExpert-NM, utilizing a mixture of expert LLMs approach, consistently outperformed other models across all datasets except for MCC on the ACL18, showcasing its superior performance.
  • Ablation study shows removing the Market expert drops AR to 30.87%, while removing other experts reduces performance to varying degrees. performance
    Source sentence
    w/o Market 30.87% 16.43% 1.88 13.29% w/o News 31.92% 18.36% 1.74 13.04% w/o Alpha 41.65% 11.38% 3.66 8.94% w/o Fundamental 44.32% 10.68% 4.15 7.82%

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 9 papers in the archive (9 by independent authors): To Trade or Not to Trade: An Agentic Approach to Estimating … (uses its method or data); Exploring the Synergy of Quantitative Factors and Newsflow … (uses its method or data)

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 (57)

  1. Blackboard Systems (1986)
  2. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education (1983)
  3. Classification Problem Solving (1984)
  4. New Ways to Make Microcircuits Smaller (1980) — link
  5. New Ways to Make Microcircuits Smaller—Duplicate Entry (1980)
  6. Strategic explanations for a diagnostic consultation system (1984) — link
  7. Strategic Explanations in Consultation—Duplicate (1983)
  8. Poligon: A System for Parallel Problem Solving (1986)
and 49 more
  1. Transfer of Rule-Based Expertise through a Tutorial Dialogue (1979)
  2. The Engineering of Qualitative Models (2021)
  3. Attention Is All You Need (2017) — link
  4. Pluto: The ‘Other’ Red Planet (2015)
  5. FinGPT: Open-Source Financial Large Language Models — → 2306.06031
  6. Finbert: Financial sentiment analysis with pre-trained language models (2019) — link
  7. Bloomberggpt: A large language model for finance (2023) — link
  8. When FLUE Meets FLANG: Benchmarks and Large Pretrained Language Model for Financial Domain (2022)
  9. Bbt-fin: Comprehensive construction of chinese financial domain pre-trained language model, corpus and benchmark (2023) — link
  10. PIXIU: a large language model, instruction data and evaluation benchmark for finance (2023)
  11. Investlm: A large language model for investment using financial domain instruction tuning — → 2309.13064
  12. FinReport: Explainable Stock Earnings Forecasting via News Factor Analyzing Model (2024)
  13. AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework (2024)
  14. Astock: A new dataset and automated stock trading based on stock-specific news analyzing model (2022) — link
  15. Stock movement prediction from tweets and historical prices (2018)
  16. Accurate stock movement prediction with self-supervised learning from sparse noisy tweets (2022)
  17. Can chatgpt forecast stock price movements? return predictability and large language models — → 2304.07619
  18. ChatGPT informed graph neural network for stock movement prediction — → 2306.03763
  19. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (2024)
  20. Hybrid deep sequential modeling for social text-driven stock prediction (2018)
  21. Llama 2: Open foundation and fine-tuned chat models (2023) — link
  22. Gpt-4 technical report (2023) — link
  23. Gemini: a family of highly capable multimodal models (2023) — link
  24. The Llama 3 Herd of Models (2024) — link
  25. InternLM2 Technical Report (2024) — link
  26. The falcon series of open language models (2023) — link
  27. Mistral 7B (2023) — link
  28. Deep reinforcement learning for automated stock trading: An ensemble strategy (2020)
  29. FinRL: A deep reinforcement learning library for automated stock trading in quantitative finance (2020) — link
  30. ModAugNet: A new forecasting framework for stock market index value with an overfitting prevention LSTM module and a prediction LSTM module (2018)
  31. Financial time series forecasting using cnn and transformer — → 2304.04912
  32. Llama: Open and efficient foundation language models (2023) — link
  33. Language models are few-shot learners (2020) — link
  34. Mixture-of-experts meets instruction tuning: A winning combination for large language models (2023) — link
  35. Glam: Efficient scaling of language models with mixture-of-experts (2022)
  36. Learning factored representations in a deep mixture of experts (2013) — link
  37. Time series feature extraction on basis of scalable hypothesis tests (tsfresh–a python package) (2018)
  38. Lo RA : Low-Rank Adaptation of Large Language Models (2022) — link
  39. Lightgbm: A highly efficient gradient boosting decision tree (2017)
  40. Aligning with human judgement: The role of pairwise preference in large language model evaluators (2024) — link
  41. Random forests (2001)
  42. Classification and regression trees (2011)
  43. Support-vector networks (1995)
  44. Asynchronous methods for deep reinforcement learning (2016)
  45. Proximal policy optimization algorithms (2017) — link
  46. DeepTrader: a deep reinforcememant learning approach for risk-return balanced portfolio management with market conditions Embedding (2021)
  47. Reinforcement-learning based portfolio management with augmented asset movement prediction states (2020)
  48. A deep reinforcement learning framework for the financial portfolio management problem (2017) — link
  49. Chain-of-thought prompting elicits reasoning in large language models (2022)

5 of 57 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.