Leveraging Large Language Models for Institutional Investment Management

In one sentence

Summary

The paper addresses the gap in applying Large Language Models to complex institutional portfolio management by investigating their ability to predict price movements in a 40/60 stock-bond portfolio using macroeconomic indicators. The authors position this work within the context of institutional decision-making, where diverse investor attitudes and long-term perspectives are critical, contrasting with prior research focused on individual stock selection or individual investor behaviors. The core problem is determining whether LLMs can effectively translate economic data into actionable portfolio adjustments and how different 'personas' influence these predictions. The methodology employs GPT-4 to predict three-class portfolio movements (hold, rise, fall) based on ten days of seven economic indicators, including interest rates, VIX, and currency indices. The experimental design tests three personas (short, medium, and long-term investors) and two ensemble methods (mode and sensitive) across 593 weekdays from October 2021 to January 2024. The investment strategy adjusts position sizes based on predictions, comparing against baselines like buy-and-hold, continuous movement, and regression models. Performance is evaluated using Sharpe ratio, return, volatility, and maximum drawdown, with further analysis stratified by CPI trends to identify specific market conditions where LLMs excel. Results indicate that the mode ensemble across personas significantly improves prediction accuracy and F1-scores for detecting market declines, achieving a recall of 0.674 for downward movements. In terms of investment performance, LLM-based strategies outperform buy-and-hold in Sharpe ratio during periods of rising CPI, leveraging the model's knowledge to recognize broader trends beyond short-term noise. However, traditional strategies prove more effective during declining CPI trends or sharp market downturns, where LLMs sometimes respond too slowly. The study concludes that while LLMs enhance portfolio management, they require complementary strategies to optimize performance across varying market regimes, highlighting the value of persona-based ensembles in capturing diverse investment rationales.

Research problem

The paper investigates how Large Language Models can predict price movements in institutional stock-bond portfolios using economic indicators and whether incorporating different investor personas via ensemble methods improves prediction accuracy and investment performance. It addresses the underexplored potential of LLMs in complex, multi-asset portfolio management compared to individual stock selection.

Core contributions

• Designs prompts enabling LLMs to manage portfolios akin to institutional investors using macroeconomic data.
• Investigates the impact of different investor personas (short, medium, long-term) on LLM prediction accuracy.
• Demonstrates that mode ensemble methods across personas improve accuracy and decline-detection F1-scores.
• Quantifies periods where LLM-based strategies outperform buy-and-hold, specifically during rising CPI trends.

Method

• Input: 10-day history of 7 macroeconomic indicators for a 40/60 stock-bond portfolio.
• Model: GPT-4 (gpt-4-0613) with chain-of-thought and persona prompting.
• Personas: Short-term (individual, limited knowledge), Medium-term (institutional), Long-term (20-30 year horizon).
• Ensemble: Mode voting across 5 trials and across 3 personas to stabilize predictions.
• Strategy: Adjust position size (0.0 to 1.0) based on predicted class (hold, rise, fall).
• Evaluation: Monthly Sharpe ratio and win-ratio against buy-and-hold, stratified by CPI trend.

Results & findings

• Mode ensemble across personas achieved the highest accuracy (0.366) and F1-score for decline detection (0.484).
• LLM-based strategies outperformed buy-and-hold in Sharpe ratio during periods of rising CPI.
• Buy-and-hold strategy performed better during declining CPI trends or sharp market downturns.
• LLM personas showed distinct reasoning patterns: short-term focused on volatility and rates, long-term on growth potential.
• Sensitive ensemble method generally underperformed compared to mode ensemble in both accuracy and F1-score.
• LLM strategies effectively reduced positions during some sharp declines but were slower to react in others compared to technical baselines.

Data & experiments

• Dataset: 593 weekdays from October 2021 to January 2024, covering US stocks and government bonds.
• Features: Seven economic indicators including interest rates, VIX, and currency indices over a 10-day window.
• Benchmarks: Buy-and-hold, continuous movement, and linear regression strategies.
• Metrics: Sharpe ratio, return, volatility, maximum drawdown, and classification metrics (precision, recall, F1).
• Evaluation: Stratified by CPI trends (rising vs. falling) to assess strategy performance in different market regimes.

Limitations & caveats

• Limited to a specific 2.5-year period, potentially missing other market regimes.
• LLM predictions are based on numerical indicators, not unstructured text, limiting generalizability to news-based strategies.
• Performance varies significantly by market condition; LLMs do not dominate in all scenarios.
• GPT-4 version used is older (June 2023), which may not reflect current model capabilities.
• Ensemble method adds computational overhead and complexity to the trading pipeline.

Curator notes

Relevant to multi-asset allocation and institutional workflows.