Adaptive and Explainable Margin Trading via LLM and RL

We introduce the source code of our paper "Adaptive and Explainable Margin Trading via Large Language Models on Portfolio Management" (ICAIF 2024).

Key Features

• Adaptive Adjustments: Dynamically reallocates funds between long and short positions based on evolving market conditions.
• Explainable Reasoning: Provides transparent explanations for market forecasts and position adjustments.
• Flexible Integration: Accommodates various external data sources and data types, including time series and textual data.
• Improved Performance: Achieves up to three times the return and doubles the Sharpe ratio compared to benchmarks.
• Publicly Available: All data and code are publicly available in the data/ folder.

Framework Overview

Explainable Market Forecasting/Reasoning Pipeline

• Utilizes LLMs to process diverse external data sources.
• Provides clear reasoning paths for optimal adjustment ratios.
• Flexible to incorporate various LLMs and data types.

Position Reallocation Stage

• Interacts with a pre-trained RL model to enhance decision-making.
• Regularly adjusts portfolio states for optimal long-short position ratios.
• Enhances transparency and trust in financial decisions.

Usage

Environment setup

• FinRL
• Python 3.10
• pandas
• peft
• torch
• openai
• anthropic
• transformers

Market Forecasting/Reasoning Pipeline

python inference_open_source_llm.py --model Llama-3-70B

Position Reallocation Stage

python interaction_llm_rl.py

Datasets

Data Source

The Dow Jones Industrial Average (DJIA) is selected as the portfolio pool. We follow Margin Trader Paper for the training, validation, and testing periods. Note that the test period is extended to 2020/5 - 2024/2 in our paper, as it includes complex economic fluctuations marked by a significant rise (COVID-19 pandemic recovery) and subsequent variations (supply chain disruptions and inflation). The price data of companies in DJIA for RL is sourced from Yahoo Finance.

Additionally, two distinct external data sources are collected and tested to evaluate their impact on near-future (six-month) US market trend prediction:

1. Macroeconomic Indicator Time Series Dataset: Comprises monthly time series data for 21 key US economic metrics. The inflation rate data is sourced from the US Inflation Calculator, while the remaining macroeconomic indicators are obtained from the Federal Reserve Bank of St. Louis. Data with daily or quarterly features has been appropriately downsampled or upsampled to a monthly frequency.
2. Microeconomic Firm-Specific News Dataset: Includes daily news data specific to the 30 companies listed on the DJIA. News is gathered from various sources, such as company announcements, earnings reports, and other significant events. The data is retrieved by ticker from the Stock News API, which indexes articles and video content from reputable sources including CNBC, Reuters, MarketWatch, Seeking Alpha, Bloomberg, and The Motley Fool.

Conclusion

Our adaptive and explainable framework for portfolio management represents a significant advancement in financial strategies, offering secure and efficient methods for dynamic long-short position adjustments using LLMs and RL. Its adaptability and transparent reasoning enhance decision-making, ensuring both profitability and stability in diverse market conditions.

During training, the best model along with predictions and labels will be saved to the result directory.

Citation

If you use this code for your research, please kindly cite our paper:

@inproceedings{gu2024adaptive,
  title={Adaptive and Explainable Margin Trading via Large Language Models on Portfolio Management},
  author={Gu, Jingyi and Ye, Junyi and Wang, Guiling and Yin, Wenpeng},
  booktitle={Proceedings of the 5th ACM International Conference on AI in Finance},
  pages={248--256},
  year={2024}
}