MultiHedge: Adaptive Coordination via Retrieval-Augmented Control

1. Problem Statement

Financial markets are characterized by frequent regime shifts (e.g., transitions from bull to bear markets) and volatility bursts, which render traditional risk management strategies unstable. Existing approaches face two primary limitations:

• Rule-based/Static Strategies: Established option strategies (e.g., collars, straddles) often fail to adapt dynamically to changing conditions if not coordinated over time.
• End-to-End Learning (RL/LLMs): Purely learning-based models (like Deep RL) often struggle with non-stationarity and distribution shifts because they optimize over historical stationary reward landscapes. Conversely, standalone Large Language Models (LLMs) can produce unstable, unconstrained, or hallucinated decisions when used directly for trading.

The core research question is: Can retrieval-augmented coordination improve the robustness of modular decision pipelines under distributional shifts and regime instability?

2. Methodology: The MultiHedge Architecture

The authors propose MultiHedge, a hybrid computational architecture that separates stochastic reasoning from symbolic execution. The system is formulated as a Partially Observable Markov Decision Process (POMDP) where the goal is to maximize risk-adjusted returns while maintaining bounded risk.

Key Architectural Components

1. Episodic Memory (Retrieval-Augmented):

   • The system maintains a non-parametric memory buffer storing past state–action–outcome tuples.
   • During decision-making, the system retrieves the top-$k$ most similar historical episodes using cosine similarity on embeddings.
   • This acts as a "non-parametric prior," grounding current decisions in empirically validated historical precedents rather than purely parametric generalization.

2. Modular Coordination Pipeline:

   • Market Analyzer: Classifies the current market regime (Bull, Bear, Volatile) and analyzes sentiment.
   • Portfolio Manager (LLM Coordinator): An LLM (GPT-4.1) processes multimodal signals and retrieved historical analogues to generate structured allocation candidates. It acts as a bounded heuristic approximator.
   • Asset Managers (Symbolic Agents): Specialized agents execute canonical option strategies based on the LLM's allocation:
     • Collar Agent: Caps downside (Buy put + sell call).
     • Straddle Agent: Profits from volatility (Buy ATM put + call).
     • Delta Agent: Constructs delta-neutral mixes.
   • Deterministic Safety Layer: Enforces hard constraints (liquidity, budget, risk limits). If drawdown thresholds are breached, this layer overrides the LLM with defensive policies to ensure stability.

3. Decision Loop:

   • Retrieve: Query memory for similar past market states.
   • Reason: LLM synthesizes a strategy based on retrieved outcomes and current signals.
   • Act: Deterministic agents execute the specific option trades.
   • Reflect: The outcome (reward) is calculated and stored back into the memory buffer for future learning.

3. Key Contributions

• Structured Sequential Decision Formulation: The paper formulates retrieval-conditioned coordination as a structured problem where LLMs serve as reasoning modules within a constrained optimization loop, rather than end-to-end traders.
• Hybrid Modular Architecture: It introduces a reproducible workflow that decouples reasoning (LLM), action generation (structured allocation), and execution (symbolic option primitives).
• Controlled Computational Study: The authors provide empirical evidence isolating the role of memory and architectural design versus model scale in achieving robustness.

4. Experimental Results

The system was evaluated on U.S. large-cap equities (AAPL, TSLA, NVDA) from 2021–2023, with memory calibrated on 2016–2020 data. It was compared against Buy & Hold, Equal Weight, PPO (Reinforcement Learning), FinAgent, and FinRL.

Performance Metrics (MultiHedge vs. Best Baseline)

Ablation Studies

The authors conducted ablation studies to determine the source of performance gains:

• Removing Memory: Performance degraded significantly. Max Drawdown nearly tripled (16.22% $\to$ 46.18%), and CVaR worsened. This confirms that memory is the dominant factor for robustness.
• Reducing Model Scale (GPT-4o Mini): Performance dropped moderately compared to removing memory, indicating that while model capacity matters, architectural design (memory + modularity) is more critical than raw model size.
• Stochasticity (Temperature 0.7): Introducing randomness degraded performance, proving that deterministic, constrained inference is essential for stability in sequential financial decisions.

5. Significance and Conclusion

The paper demonstrates that retrieval-augmented coordination is a superior mechanism for financial decision-making under uncertainty compared to simply scaling up model parameters or relying on end-to-end reinforcement learning.

• Robustness over Scale: The study concludes that memory and modular architectural design play a central role in robustness. Grounding LLM reasoning in episodic memory allows the system to adapt to regime shifts without requiring retraining.
• Safety and Interpretability: By separating the "brain" (LLM reasoning) from the "hands" (deterministic execution), MultiHedge ensures that decisions remain interpretable, auditable, and bounded by strict risk constraints.
• Future Directions: The authors suggest extending the architecture to include term-structure features, integrating risk-sensitive objectives, and generalizing to multi-agent coordination protocols.

In summary, MultiHedge proves that combining retrieval-based memory with modular, symbolic execution creates a more stable and robust financial agent than current state-of-the-art learning-based or purely rule-based systems.