DeltaHedge: A Multi-Agent Framework for Portfolio Options Optimization

In one sentence

Summary

The paper addresses the limitation of existing AI-driven portfolio management systems that largely ignore options trading for dynamic risk hedging. It proposes DeltaHedge, a hierarchical multi-agent system that coordinates specialized modules for forecasting, sentiment analysis, trading, and hedging. The framework aims to balance risk and return by integrating options into the decision-making loop, addressing the gap where traditional equity-focused methods fail to mitigate tail risks effectively in volatile markets. The system leverages the unique properties of options to provide downside protection while maintaining upside potential through coordinated agent interactions. The methodology employs a Forecasting Agent using the Informer transformer for price predictions and a Sentiment Analysis Agent using DistilRoBERTa for news sentiment. The core decision-making involves a Trading Agent using Proximal Policy Optimization (PPO) for equity allocation and a Hedging Agent that dynamically adjusts protective put positions to maintain a delta-neutral stance. An ensemble strategy concurrently trains PPO, A2C, and DDPG agents, selecting the best performer quarterly based on Sharpe ratio validation. Experiments are conducted on S&P 500, Apple, and Tesla data from 2010-2024, comparing against Buy-and-Hold, technical strategies, standalone RL, and LLM-based frameworks like FinAgent, using metrics like Sharpe, Sortino, and Maximum Drawdown. Results show DeltaHedge achieves a Sharpe ratio of 1.33 and Sortino ratio of 1.81 on the S&P 500, nearly doubling the next-best method, while capping maximum drawdowns at 10%. The ensemble hedging strategy proves robust across rising, falling, and volatile market regimes, outperforming single-agent hedgers and classical delta-hedge benchmarks. The system effectively times protection, lightening hedges when insurance is overpriced and switching to full protection during volatility spikes. Limitations include the focus on long-only positions, single-asset scope, and the assumption of sufficient liquidity for option execution, with future work targeting multi-asset portfolios and complex option structures.

Research problem

Traditional AI-driven portfolio management systems primarily focus on equity allocation, often overlooking the strategic advantages of options trading for dynamic risk hedging. This leaves portfolios exposed to discontinuous price jumps and tail risks that equity-only strategies cannot effectively mitigate. The paper addresses the need for a framework that systematically integrates options into AI-driven decision-making to enhance risk-adjusted returns and stabilize performance across varying market conditions.

Core contributions

• Introduction of DeltaHedge, a novel multi-agent framework integrating options-based hedging into AI-driven portfolio management.
• Development of an ensemble-based hedging strategy that dynamically selects among multiple RL agents (PPO, A2C, DDPG) to mitigate algorithmic biases and adapt to market regime shifts.
• Implementation of a hierarchical agent architecture with specialized roles for forecasting, sentiment analysis, trading, and hedging, coordinated via cross-attention mechanisms.
• Comprehensive benchmarking and ablation study demonstrating the superiority of the ensemble options-hedging approach over traditional strategies, standalone RL models, and LLM-based frameworks.
• Empirical validation on S&P 500, Apple, and Tesla data, showing significant improvements in Sharpe and Sortino ratios while reducing maximum drawdowns.

Method

• Forecasting Agent uses Informer transformer for 30-day price change predictions based on 60-day historical data.
• Sentiment Analysis Agent uses DistilRoBERTa to generate daily sentiment scores from financial news headlines.
• Trading Agent uses PPO to determine continuous equity allocation actions, maximizing portfolio Sharpe ratio.
• Hedging Agent uses an ensemble of PPO, A2C, and DDPG to dynamically adjust protective put positions for delta neutrality.
• Coordinator Agent synchronizes workflow, manages option expirations, and computes shared portfolio-level rewards.
• Cross-attention mechanism allows agents to share internal decision summaries for coordinated strategy adjustments.
• Quarterly retraining of hedging agents with a 90-day lookback and 30-day validation window for model selection.
• Transaction costs modeled as 0.2% for stocks and 0.7% fixed plus 0.5% proportional for options.

Results & findings

• DeltaHedge achieves a Sharpe ratio of 1.33 and Sortino ratio of 1.81 on the S&P 500, nearly doubling the next-best method.
• Total return of 121% with a maximum drawdown of only 10% on the S&P 500 during the testing period.
• Annualized volatility of 14%, lower than most benchmarks except the classic delta hedge, which sacrifices significant return.
• The ensemble hedging strategy outperforms single-agent hedgers (PPO, A2C, DDPG) by adapting to changing market conditions.
• DeltaHedge prevents breaches below previous cycle lows during the 2020 Covid crash and 2022 inflation-driven decline.
• Ablation studies confirm that both the options integration and the ensemble mechanism contribute significantly to performance gains.
• The system effectively times protection, reducing option costs during calm periods and increasing coverage during volatility spikes.

Data & experiments

• Daily price data (OHLCV) from Yahoo Finance for S&P 500, Apple, and Tesla.
• Options data from optionsdx.com, focusing on near-the-money protective puts with 30-day expiration.
• Sentiment data from EOD Historical Data, processed using DistilRoBERTa for financial sentiment classification.
• Volatility index (VIX) from Yahoo Finance to capture broader market uncertainty.
• Training period: 2010-2019 for S&P 500, 2016-2019 for Apple and Tesla; Testing period: 2020-2024.
• Comparison against Buy-and-Hold, KDJ with RSI filter, standalone PPO/DQN, and FinAgent (LLM-based framework).

Limitations & caveats

• The framework is limited to long-only positions, excluding short selling strategies.
• Experiments are conducted on single-asset portfolios (S&P 500, Apple, Tesla), not multi-asset diversification.
• Assumes sufficient liquidity for option execution, which may not hold in all market conditions.
• Relies on daily rebalancing, which may miss intraday opportunities or incur higher costs in high-frequency contexts.
• Sentiment data is derived from news headlines, potentially missing broader market sentiment indicators.
• The ensemble strategy requires computational resources for concurrent training and quarterly retraining.

Curator notes

Discovery status: institutional_trading_round2_promoted. Source links: From Natural Language to Executable Option Strategies via Large Language Models