Background: LLM Research in Quantitative Trading

Large Language Models (LLMs) are changing the research paradigm in quantitative trading. This document summarizes the main research directions and representative works of LLMs in the quantitative field.

1. LLM Application Scenarios in Quantitative Trading

2. Representative Research Works

2.1 FinGPT (2023)

Source: Columbia University + AI4Finance

Position: Open-source financial large language model

Core Features:

• Fine-tuned on LLaMA
• Trained on financial corpus
• Supports sentiment analysis, Q&A, summarization
• Open-source and available

Architecture:

Base Model (LLaMA)
    |
Financial Corpus Pre-training (Financial News, SEC Filings)
    |
Task Fine-tuning (Sentiment, QA, Summarization)
    |
FinGPT

Limitations:

• Mainly for NLP tasks
• Cannot directly generate trading signals
• Insufficient real-time capability

Link: https://github.com/AI4Finance-Foundation/FinGPT

2.2 BloombergGPT (2023)

Source: Bloomberg

Position: 50-billion parameter finance-specific LLM

Core Features:

• Proprietary model, not open-source
• Training data includes Bloomberg's exclusive data
• Outperforms general models on financial NLP tasks

Training Data:

Performance: Leads GPT-3 on financial sentiment analysis, NER, and other tasks

Limitation: Not open-source, cannot be used directly

2.3 FLAG-Trader (2024-2025)

Source: Academic research

Position: LLM + Reinforcement Learning trading system

Core Innovation:

• LLM generates trading hypotheses
• RL validates and optimizes
• Feedback loop improves LLM

Architecture:

Market Data + News
    |
LLM (Generate trading hypotheses)
    |
RL Agent (Execute and evaluate)
    |
Reward Feedback
    |
LLM Improvement (Learn from feedback)

Advantages:

• Combines LLM's reasoning ability with RL's optimization capability
• Good interpretability
• Can handle text + numerical multimodal data

2.4 TradingGPT / QuantGPT Series

Position: Using GPT-4 and similar models for trading decisions

Method:

• Directly use GPT-4 API
• Design specific prompts
• Generate trading recommendations

Typical Prompt:

You are a quantitative analyst. Based on the following market data:
- BTC 24-hour change: +5.2%
- RSI(14): 72
- Volume: Up 30% from yesterday

Please analyze the current market state and provide trading recommendations.

Limitations:

• Dependent on API, high latency
• High cost
• Hallucination issues

2.5 MM-DREX (2024)

Position: Multimodal trading system

Core Innovation:

• Directly "sees" candlestick charts
• Combines text + images
• Visual pattern recognition

Input:

Text: News, indicator values
Image: Candlestick charts, technical analysis charts
    |
Multimodal LLM
    |
Trading Decision

Advantage: Mimics human traders' ability to "read charts"

3. Main Research Directions

3.1 Enhanced Sentiment Analysis

Traditional Method:

# Simple dictionary approach
positive_words = ['bullish', 'surge', 'rally']
negative_words = ['bearish', 'crash', 'plunge']

LLM Method:

# Context understanding
prompt = """
Analyze the market sentiment of the following news (-1 to 1):
"Fed signals potential rate cuts, but warns inflation remains sticky"

Please consider:
1. Direct impact
2. Implied expectations
3. Possible market reaction
"""

Advantage: Understands complex, contradictory information

3.2 Factor Discovery and Hypothesis Generation

Traditional Method: Manual factor design

LLM Method:

prompt = """
Based on the following market patterns, propose potentially effective quantitative factors:
1. Small-cap stocks outperform large-cap over the long term
2. High momentum stocks show continuation
3. Low volatility stocks have better risk-adjusted returns

Please propose 3 new factor hypotheses, including:
- Factor definition
- Theoretical basis
- Potential risks
"""

Advantage: Rapidly generate many hypotheses for testing

3.3 Code Generation

Applications:

• Strategy code generation
• Data processing scripts
• Visualization code

Example:

prompt = """
Implement a dual moving average strategy in Python:
- Short-term MA: 10 days
- Long-term MA: 30 days
- Golden cross = buy, death cross = sell
- Use pandas for data processing
"""

Note: Generated code requires human review

3.4 Market Reasoning

Application: Multi-step logical reasoning

Example:

Question: If the Fed raises rates, what's the impact on tech stocks?

LLM Reasoning Chain:
1. Rate hike -> Interest rates rise
2. Rising rates -> Higher discount rate
3. Higher discount rate -> Future cash flow present value decreases
4. Tech stocks depend on future growth -> Valuations more affected
5. Conclusion: Tech stocks may fall, especially high-valuation growth stocks

4. Current Limitations

4.1 Hallucination Problem

LLMs may generate plausible-sounding but incorrect analysis:

• Fabricated data
• Incorrect causal relationships
• Overconfident predictions

Coping Strategy: Always verify with real data

4.2 Insufficient Real-Time Capability

• API call latency: 100-500ms
• Model inference time: 1-10 seconds
• Not suitable for high-frequency trading

4.3 Cost Issues

4.4 Cannot Predict Prices

LLM Can Do:

• Understand news
• Analyze sentiment
• Generate hypotheses

LLM Cannot Do:

• Predict tomorrow's price
• Guarantee profits
• Replace traditional quantitative models

5. Practical Recommendations

5.1 Tasks Suitable for LLM

5.2 Recommended Architecture

Traditional Quantitative Model (Primary)
    ^
LLM Enhancement Layer (Auxiliary)
- Sentiment signals
- News filtering
- Hypothesis generation
    ^
Raw Data

5.3 Open-Source Options

6. Future Trends

1. Multimodal Fusion: Text + Numerical + Image
2. Agent-ification: LLM as the "brain" of trading Agents
3. Real-time Processing: Lower latency inference
4. Specialization: More finance-specific models
5. Compliance: Meeting regulatory explainability requirements

Core Principle: LLMs are powerful tools, but not magic. They can enhance your analytical capabilities but cannot replace solid quantitative foundations and strict risk control.