💰 quantitative finance

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

This paper proposes a novel LLM-guided no-regret portfolio allocation framework that integrates online learning, market sentiment, and large language model-driven hedging to significantly outperform the SPY buy-and-hold baseline in both annualized returns and Sharpe ratio for risk-averse investors.

Original authors: Muhammad Abro, Hassan Jaleel

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muhammad Abro, Hassan Jaleel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a ship trying to cross an ocean. Your goal is to get to the destination as fast as possible (high returns), but you also want to avoid hitting icebergs or getting caught in massive storms (high risk).

Most traditional ship captains (investors) face a tough choice: either they sail fast and risk crashing, or they sail slowly and safely but arrive late. This paper introduces a new kind of "smart autopilot" that tries to do both: sail fast while dodging the worst storms.

Here is how this new system works, broken down into simple concepts:

1. The "No-Regret" Autopilot

The core of this system is something called "No-Regret Learning."

Think of it like this: Imagine you are playing a video game where you have to pick a path every day. A "regretful" player looks back at the end of the week and thinks, "Oh no! If I had just picked the sunny path on Tuesday, I would have won!"

The "No-Regret" system is designed so that, over a long time, it doesn't matter which single path was actually the best. The system constantly adjusts its course to ensure that, in hindsight, it didn't miss out on the best possible outcome. It learns from its mistakes in real-time, slowly shifting its weight to the assets (like stocks, gold, or bonds) that are performing best, while avoiding the ones that are sinking.

2. The "Sentiment Weather Report"

The system doesn't just look at the numbers; it also checks the "mood" of the market. It uses a tool called the Fear & Greed Index, which measures how scared or greedy investors are feeling.

The Analogy: Imagine a co-pilot who looks out the window. If everyone is panicking (extreme fear), this co-pilot tells the autopilot to slow down or even drop anchor (sell everything and hold cash). If everyone is overly excited (extreme greed), it warns the ship to be careful. This acts as a safety brake to keep the ship from crashing during emotional market swings.

3. The "AI Navigator" (The LLM)

This is where the paper gets fancy. They used a Large Language Model (LLM)—a type of advanced AI that reads and understands text—as a special advisor.

What it does: Instead of asking the AI to pick specific winning stocks (which it isn't great at), they asked it to act like a strategic advisor. They fed the AI the "mood report" and asked, "Based on how scared or greedy people are right now, which sectors of the economy look like good places to hide, and which ones look like they need a shield?"
The Result: The AI suggested "hedges." Think of a hedge like an umbrella. If the AI senses a storm is coming in the "Technology" sector, it suggests buying assets in a different sector (like "Utilities" or "Gold") that usually stay calm when tech stocks crash. This doesn't necessarily make the ship go faster, but it keeps it from capsizing.

4. The Test Drive

The authors tested this system in two different "oceans":

The Simple Ocean: A mix of just four things: Cash, Gold, Bonds, and the S&P 500 (a big basket of top US companies).
The Complex Ocean: A mix of 51 different sectors (like Energy, Healthcare, Tech, etc.).

The Results:

Speed: The new system beat the standard "buy and hold" strategy (just buying the S&P 500 and doing nothing) by a huge margin. In the complex ocean, it returned 69% more money over the test period.
Safety: It didn't just go faster; it was smoother. When the market crashed (like during the pandemic or inflation shocks), this system fell much less than the standard market.
The "Sharpe Ratio": This is a score that measures how much reward you get for every unit of risk you take. The new system scored 119% higher than the standard market. In simple terms, it got you more profit for the same amount of worry.

5. How Often Does It Move?

You might think a system this smart would be trading every second. Surprisingly, it only moves once every quarter (every three months). This is similar to how many standard index funds operate. This is important because it means the system is practical and doesn't waste money on excessive trading fees.

The Bottom Line

The paper claims that by combining a math-based "no-regret" learning system with an AI that reads the market's emotional mood to suggest safety shields, investors can build a portfolio that grows faster than the average market while suffering fewer painful drops. It's like having a ship that automatically steers around storms without needing a human to constantly check the radar.

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

Problem Statement
The paper addresses the persistent challenge in medium- to long-term portfolio management of balancing risk and return. While passive index funds (e.g., S&P 500) have historically outperformed many active hedge funds on a nominal return basis, institutional and individual investors often prioritize risk-adjusted returns and downside protection over raw gains. Traditional active strategies frequently suffer from high volatility, infrequent rebalancing to avoid transaction costs, and a trade-off where higher returns necessitate higher risk. The authors propose a solution to achieve superior returns without incurring undue risk or excessive trading frequency, specifically targeting risk-averse investors and fund managers.

Methodology
The authors propose a novel framework that integrates no-regret online learning with Large Language Model (LLM) guidance. The approach is structured as a sequential decision-making problem where an agent dynamically selects portfolio weights to minimize "regret"—the difference between the agent's cumulative performance and the best fixed portfolio in hindsight.

The framework consists of four primary components:

Data Preparation & Sentiment Filtering: The system processes asset price data and market sentiment indices (specifically CNN's Fear & Greed Index). A binary sentiment filter acts as a coarse-grained risk control, triggering either portfolio rebalancing or full liquidation during extreme sentiment conditions.
LLM-Guided Hedging and Clustering: An LLM interprets historical sentiment trends to suggest sector-level adjustments.
- Dynamic Clustering: The LLM is prompted with a 12-month sentiment time series to recommend promising sectors. To mitigate variance, the prompt is repeated five times, and the most frequent recommendations are selected.
- Pseudo-Hedging: The LLM is queried to identify complementary sectors that act as hedges against the selected clusters, aiming to reduce volatility and improve the Sharpe ratio.
Regret-Guided Allocation (The Core Engine): The allocation mechanism utilizes a "Follow-the-Leader" (FTL) inspired strategy.
- Game-Theoretic Formulation: The problem is modeled as a single-agent game where the agent selects a portfolio vector $a(t)$ over $n$ assets at each timestep.
- Action Generation: Instead of maintaining weighted mixtures of experts (as in classical Hedge algorithms), the system generates candidate actions by perturbing the current baseline allocation. Specifically, it modifies the weight of a single asset within predefined bounds and adjusts remaining allocations to satisfy the sum-to-one constraint.
- Selection: The agent selects the candidate action that has achieved the highest absolute performance (e.g., maximizing Sharpe, Calmar, or Sortino ratios) over a specific lookback window ( $k$ ).
Backtesting Mechanics: The system simulates trading with realistic constraints, including transaction costs, delisting handling (blacklisting assets), and daily price tracking for bonds and cash.

Key Contributions
The paper outlines three main contributions:

Greedy Regret-based Portfolio Strategy: A novel optimization framework that dynamically rebalances assets to minimize hindsight regret. Unlike traditional static methods, this approach adapts to shifting market dynamics while explicitly optimizing for risk-adjusted performance metrics.
LLM-Guided Clustering and Hedging: The integration of LLM insights directly into the portfolio construction pipeline. The authors demonstrate that while LLMs may struggle to predict high-return sectors solely from sentiment data, they are effective at identifying complementary hedging sectors, thereby systematically linking behavioral finance signals to allocation decisions.
Empirical Outperformance: The framework was validated on two distinct asset universes: a minimalist "COCKROACH" mix (Cash, Gold, Bonds, SPY) and a complex set of 51 sector-based ETFs. The study provides extensive sensitivity analyses regarding rebalancing frequency, lookback windows, and sentiment filters.

Results
Empirical results from backtesting (January 2011 – June 2024 for the ETF universe) demonstrate significant outperformance over traditional benchmarks:

Returns and Risk-Adjusted Metrics: The proposed strategy outperformed the SPY buy-and-hold baseline by 68.6% in cumulative returns and achieved a 119% higher Sharpe ratio.
Drawdown Control: The strategy exhibited significantly lower maximum drawdown (approx. 16%) compared to SPY (34%) over the same period, highlighting superior downside protection.
LLM Impact: LLM-guided hedging improved risk-adjusted returns substantially, boosting Sharpe ratios by up to 63% in specific configurations (quarterly rebalancing). Conversely, dynamic sector clustering alone did not outperform the full asset universe, suggesting LLMs are better suited for hedging than for primary sector selection in this context.
Rebalancing Frequency: A quarterly rebalancing schedule was found to be optimal, striking a balance between capturing medium-term trends and minimizing transaction costs.

Significance and Claims
The paper claims that its framework offers a robust, adaptive solution for low-frequency asset management that successfully integrates online learning with behavioral signals. By constraining the search space to incremental adjustments around a diversified baseline and utilizing LLMs for risk-offsetting rather than pure alpha generation, the method produces portfolios that consistently outperform standard benchmarks with lower volatility.

The authors position this work as a bridge between classical optimization, no-regret learning theory, and modern NLP capabilities. They assert that the approach is particularly valuable for risk-averse investors and institutional managers seeking strategies that offer downside protection and low correlation to market swings without the high fees or underperformance often associated with traditional active management. The study concludes that while limitations exist (such as fixed-frequency rebalancing and LLM predictive constraints), the integration of "no-regret" updates with lightweight sentiment conditioning represents a significant step forward in practical portfolio construction.