Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Regret-Driven Portfolios: LLM-Guided Smart Clustering for Optimal Allocation

Muhammad Abro, Hassan Jaleel

TL;DR

This work addresses the risk–return trade-off in medium- to long-horizon portfolios for risk-averse investors and institutions. It proposes a no-regret online learning framework that integrates sentiment-driven LLM insights with hedging to dynamically adjust a diversified asset universe. Key contributions include a four-module architecture combining data preparation, sentiment gating, LLM-guided hedging, and regret-guided allocation, along with empirical evaluation showing substantial improvements in risk-adjusted performance across two asset universes and sensitivity analyses of cadence, lookback, and signal conditioning. The framework offers a practical, low-frequency, risk-aware approach for institutional asset management and demonstrates how online learning and NLP-driven signals can be integrated into portfolio construction for robust performance.

Abstract

We attempt to mitigate the persistent tradeoff between risk and return in medium- to long-term portfolio management. This paper proposes a novel LLM-guided no-regret portfolio allocation framework that integrates online learning dynamics, market sentiment indicators, and large language model (LLM)-based hedging to construct high-Sharpe ratio portfolios tailored for risk-averse investors and institutional fund managers. Our approach builds on a follow-the-leader approach, enriched with sentiment-based trade filtering and LLM-driven downside protection. Empirical results demonstrate that our method outperforms a SPY buy-and-hold baseline by 69% in annualized returns and 119% in Sharpe ratio.

Regret-Driven Portfolios: LLM-Guided Smart Clustering for Optimal Allocation

TL;DR

This work addresses the risk–return trade-off in medium- to long-horizon portfolios for risk-averse investors and institutions. It proposes a no-regret online learning framework that integrates sentiment-driven LLM insights with hedging to dynamically adjust a diversified asset universe. Key contributions include a four-module architecture combining data preparation, sentiment gating, LLM-guided hedging, and regret-guided allocation, along with empirical evaluation showing substantial improvements in risk-adjusted performance across two asset universes and sensitivity analyses of cadence, lookback, and signal conditioning. The framework offers a practical, low-frequency, risk-aware approach for institutional asset management and demonstrates how online learning and NLP-driven signals can be integrated into portfolio construction for robust performance.

Abstract

We attempt to mitigate the persistent tradeoff between risk and return in medium- to long-term portfolio management. This paper proposes a novel LLM-guided no-regret portfolio allocation framework that integrates online learning dynamics, market sentiment indicators, and large language model (LLM)-based hedging to construct high-Sharpe ratio portfolios tailored for risk-averse investors and institutional fund managers. Our approach builds on a follow-the-leader approach, enriched with sentiment-based trade filtering and LLM-driven downside protection. Empirical results demonstrate that our method outperforms a SPY buy-and-hold baseline by 69% in annualized returns and 119% in Sharpe ratio.
Paper Structure (21 sections, 6 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 21 sections, 6 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Asset Universe Construction and Data Sources
  5. Backtesting Period
  6. Evaluation Benchmarks and Metrics
  7. Architecture
  8. Sentiment filtering Module.
  9. LLM-Based Sector Selection and Hedging Module.
  10. Regret-Based Portfolio Construction Mechanism.
  11. Game-Theoretic Formulation
  12. Regret Guided Learning Objective
  13. Action Set Generation
  14. Parameters and Design Choices
  15. Backtesting Mechanics
  16. ...and 6 more sections

Figures (5)

  • Figure 1: The No Regret Learning strategy (red) significantly outperforms buy-and-hold benchmarks (SPX, gold, bonds) within the sector-based ETFs asset universe, demonstrating lower volatility and resilience during major market downturns (e.g., 2020 COVID-19 crash, 2022 inflation shock) than benchmarks like SPY, as well as online approaches like MAD (online) Konno1991 and Hedge (multiplicative weights) Helmbold1998.
  • Figure 2: The volatility and drawdown statistics for the NRL tests in Figure \ref{['nrlperformance']} (Jan 2011 - June 2024). Here we observe Regret Learning (Our approach) obtaining superior results with slightly lower volatility and signficantly lower drawdown.
  • Figure 3: Framework modules: (1) Data Prep, (2) Sentiment gating, (3) LLM hedging, (4) Regret-driven allocation.
  • Figure 4: The No Regret Learning strategy (purple) outperforms traditional buy-and-hold benchmarks (SPX, gold, bonds, static baseline) on COCKROACH through adaptive allocation and lower volatility, demonstrating superior long-term performance and robustness.
  • Figure 5: NRL's performance across multiple metrics as compared to SPY (benchmark) from January 2011 to June 2024, using the top configuration shown in Table \ref{['tab:strategy-performance']}. (separate scale used for returns)