💰 quantitative finance

Large Language Model-Driven Small-Capitalization Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals

This paper demonstrates that an uncertainty-aware portfolio construction strategy, which integrates Large Language Model-derived sentiment with macroeconomic and technical signals while explicitly decomposing risk into aleatoric and epistemic components, achieves superior performance by separating pure-alpha and pure-beta stock-selection regimes rather than requiring their simultaneous confirmation.

Original authors: Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Arman Khaledian

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Arman Khaledian

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

The Digital Crystal Ball and the Stock Market

Imagine the stock market as a giant, chaotic ocean. For decades, investors have tried to predict the waves using two main tools: looking at the water's surface (technical signals like price charts) and reading the weather reports (macroeconomic indicators like interest rates or oil prices). But there's a third, noisy element: the endless stream of news. In the past, computers tried to read this news using simple "lexicons," which are just lists of good words (like "profit") and bad words (like "loss"). It's like trying to understand a complex movie by only counting how many times the words "happy" or "sad" appear; you miss the plot, the irony, and the context.

Enter Large Language Models (LLMs). Think of these as super-smart digital readers that can actually understand the story behind the headlines, grasping nuance and context that simple word-counting misses. The big question researchers are asking is: If we give these super-readers a job to predict the future of small, risky companies (called "small-cap" stocks), can they help us build a better investment portfolio? The answer isn't just about what the news says, but how we use that information to decide which stocks to buy and how much risk to take.

The Story: Teaching a Robot to Read the Room

This paper is about a team of researchers who built a sophisticated trading system to test exactly this idea. They wanted to see if they could use an AI to read financial news, combine it with big economic data, and then use that information to make money trading small companies. But they didn't just want to guess the future; they wanted the AI to admit when it was unsure.

Here is how their system works, step-by-step:

1. The News Filter: Grouping the Noise
Imagine a breaking news story about a company. Dozens of news outlets might write about it, sometimes repeating the same facts over and over. If you count every single article, that one story might seem ten times more important than it really is. The researchers fixed this by using a "clustering" trick. They grouped similar articles together into one "story" before the AI read them. This way, the AI focuses on the importance of the event, not just how many reporters showed up to cover it.

2. The Two Types of Signals: The "Pure" Players
The researchers realized that stocks move for two very different reasons, and they treated them as separate players in a game:

Pure Alpha (The Company's Own Drama): This happens when a specific company does something weird—like a CEO quitting or a new product launch—that isn't explained by the general economy. It's like a specific actor in a play suddenly improvising a scene.
Pure Beta (The Wave of the Ocean): This happens when a big economic indicator (like interest rates or oil prices) moves, and a stock moves because it is sensitive to that change, even before the stock itself has done anything weird. It's like a surfer riding a wave before they even see the wave crest.

The team tested three strategies: betting only on the "Pure Alpha" drama, betting only on the "Pure Beta" waves, and betting only when both happened at the exact same time (the "Beta Intersection").

3. The Uncertainty-Aware Brain
Most trading systems treat risk as a fixed number, like a speed limit sign. This team's AI, however, calculates risk dynamically. It breaks risk down into two parts:

Aleatoric Uncertainty: The randomness of the market itself (the weather is just unpredictable).
Epistemic Uncertainty: The AI's own lack of confidence (the AI is saying, "I'm not sure I understand this story yet").

By feeding this "uncertainty score" directly into the math that decides how much money to put into each stock, the system becomes more cautious when it's confused and more aggressive when it's sure.

What They Found: The Surprising Results

The researchers ran their system on a massive list of small companies (the Russell 2000) over a specific period in 2025. They tested different holding periods (how long to keep a stock) and different transaction costs (the fees paid to trade).

The Big Surprise: Don't Wait for Agreement
The most important finding was that separating the signals worked much better than waiting for them to agree.

The "Both" Strategy Failed: Requiring both the "Pure Alpha" drama and the "Pure Beta" wave to happen at the same time usually resulted in the worst performance. It was like waiting for the perfect storm and the perfect actor to show up simultaneously; you just missed too many good opportunities.
The "Separate" Strategies Won: The system made the most money when it treated the two signals as separate strategies.
- Short-Term (1 Day): The "Pure Beta" strategy (riding the economic wave) worked well when trading costs were low, but the advantage vanished when costs got high (100 basis points).
- Medium-Term (20–40 Days): The "Pure Beta" strategy shined again. It seems that big economic changes take a while to fully sink in for small companies, and the AI could catch this slow drift.
- Long-Term (5, 10, and 60 Days): The "Pure Alpha" strategy (the company's own drama) was the clear winner. It suggests that when a specific company has unique news, it takes time for the market to fully understand and price it in.

The Best Combination
The single most successful setup they found was a "conservative" one:

Strategy: Pure Beta (riding the economic wave).
AI Model: GPT-4o mini (a specific version of the AI).
Holding Period: 40 days.
Risk Management: Using a "Student-t target" (a math method that handles extreme market crashes better) and "Risk Parity" (balancing risk rather than just balancing money).
Result: This setup achieved a Sharpe ratio of 2.33 even with a high transaction cost of 100 basis points.

What This Means

The paper suggests that the secret to using AI for trading isn't just having a smarter reader; it's about how you organize the game. By separating the "company-specific" news from the "economy-wide" news, and by letting the AI's own uncertainty guide how much risk it takes, investors can do better than traditional methods.

However, the authors are careful to note that this is a simulation based on one year of data. They suggest these results are promising evidence, not a guaranteed "win" for the future. They also point out that their AI, while smart, still has a "knowledge cutoff" (it doesn't know things that happened after its training ended), and summarizing long news articles might lose some tiny details. But the core idea stands: in the noisy world of small stocks, knowing when to listen to the economy and when to listen to the company—and knowing when to be cautious—is just as important as the news itself.

Technical Summary: Large Language Model-Driven Small-Capitalization Trading

Problem Statement
While financial news drives market movements, traditional sentiment analysis using fixed lexicons often fails to capture negation, context, and domain-specific financial language. Although Large Language Models (LLMs) address these linguistic limitations, existing literature suffers from two primary gaps:

Redundant Coverage: Most studies score headlines as independent observations, allowing syndicated coverage of a single event to artificially overweight a signal based on frequency rather than informational value.
Static Risk Modeling: Sentiment is typically treated solely as an input for expected returns, while portfolio risk (covariance) remains a historical, signal-independent estimate. This ignores the predictive model's ability to provide information regarding uncertainty and cross-asset covariance.

Furthermore, LLM-based trading faces validation challenges regarding "look-ahead bias" (where a model's training data overlaps the evaluation period) and "distraction effects" (where general company knowledge distorts sentiment judgments).

Methodology
The authors propose an end-to-end pipeline that integrates news-summary sentiment, macroeconomic indicators, and technical signals into an uncertainty-aware portfolio construction framework.

Data and Stock Selection:
- The universe consists of Russell 2000 equities.
- The pipeline decomposes the "macro-beta trigger" into three distinct selection regimes based on z-score thresholds for stock returns ($SS$) and indicator returns ($SI$):
  - Pure Alpha ( $SS \setminus SI$ ): Isolates firm-specific abnormal moves not contemporaneously explained by macro indicators.
  - Pure Beta ( $SI \setminus SS$ ): Captures anticipatory macro moves where an indicator has shifted for a levered stock, but the stock itself has not yet reacted.
  - Beta Intersection ( $SI \cap SS$ ): Requires both channels to fire simultaneously.
- Candidates are ranked using in-sample BUY events and capped at 50 names.
Sentiment Processing:
- Clustering: To prevent syndicated coverage from dominating signals, articles are embedded and clustered via single-linkage agglomerative clustering (cosine distance threshold 0.90) within a 30-day rolling window. Only the article nearest to the cluster centroid is retained as the representative story.
- Scoring: The representative articles are scored using LLMs (primarily GPT-4o mini, with robustness checks using FinBERT, Mistral 7B, and Llama 2). Models utilize knowledge cutoffs preceding the test period (e.g., October 2023 for GPT-4o mini) to mitigate look-ahead bias.
- Normalization: Sentiment is transformed into a normalized surprise via entity-prior correction and strictly trailing rolling demeaning.
Joint Return and Risk Prediction:
- A multimodal network processes news-summary features and daily OHLCV-derived technical indicators.
- Instead of treating risk as a separate historical quantity, the model predicts the full conditional return distribution (mean $\hat{\mu}_t$ and covariance $\hat{\Sigma}_t$ ).
- Uncertainty Decomposition: The model distinguishes between:
  - Aleatoric Uncertainty: Market randomness, captured via a low-rank plus diagonal covariance head.
  - Epistemic Uncertainty: Model confidence, captured via Monte Carlo dropout at inference. The spread across stochastic passes estimates the model's uncertainty.
- These are combined into a total predictive covariance matrix: $\hat{\Sigma}_t = \hat{\Sigma}^A_t + \hat{\Sigma}^E_t$ .
Portfolio Construction:
- The predicted moments ( $\mu_t, \Sigma_t$ ) are fed directly into a constrained Mean-Variance Optimizer (MVO) with transaction cost regularization.
- The optimizer is penalized more heavily when the model itself is uncertain (high epistemic covariance), not just when assets are historically volatile.
- Benchmarks include Equal Weight, Risk Parity, Hierarchical Risk Parity, and Black-Litterman variants.

Key Results
Evaluated on Russell 2000 equities (2025 out-of-sample) across various holding periods (1 to 60 days) and transaction costs (up to 100 bps):

Regime Superiority: Separating the macro-beta trigger into "Pure Alpha" and "Pure Beta" generally outperforms the "Beta Intersection" (requiring both triggers). The intersection strategy is often weaker because it excludes early macro spillovers and idiosyncratic stock events.
Horizon Dependence:
- Short Term (1 day): Pure Beta can be effective at low costs by capturing immediate lead-lag spillovers from liquid macro indicators to small caps, but this advantage vanishes at 100 bps due to turnover and microstructure noise.
- Medium Term (5, 10, 60 days): Pure Alpha dominates. This suggests firm-specific information (earnings, contracts, governance) diffuses slowly in small caps, allowing for delayed repricing.
- Medium-Long Term (20, 40 days): Pure Beta dominates again, suggesting a slower macro repricing channel where aggregate shocks are gradually mapped into heterogeneous firm fundamentals.
Best Performance: The strongest conservative configuration identified was Pure Beta using GPT-4o mini sentiment, a Student-t target distribution, a 40-day holding period, and Risk Parity allocation. This achieved a Sharpe ratio of 2.33 at 100 bps transaction costs.
Robustness: The 60-day Pure Alpha advantage remained robust across the full range of transaction costs (0 to 100 bps), suggesting it is not an artifact of specific cost assumptions.

Significance and Claims
The paper claims that the stock-selection regime (how the tradable set is defined) and the allocator choice matter at least as much as the sentiment model itself.

Decomposition is Informative: Separating firm-specific triggers (Pure Alpha) from macro-exposure triggers (Pure Beta) yields more actionable signals than requiring them to fire simultaneously.
Uncertainty Integration: Feeding model-predicted risk (decomposed into aleatoric and epistemic components) directly into the covariance matrix of portfolio allocators is a critical advancement over treating risk as a fixed, historical quantity.
Contextual Sentiment: Clustering related news stories before scoring prevents repetitive syndicated coverage from distorting signals, ensuring the sentiment reflects the story's importance rather than its frequency.

The authors conclude that LLM news sentiment is most effective when embedded in a broader architecture that controls the opportunity set, distinguishes between macro and idiosyncratic channels, and utilizes predicted uncertainty in portfolio construction, rather than serving merely as an expected-return overlay.

The Digital Crystal Ball and the Stock Market

The Story: Teaching a Robot to Read the Room

What They Found: The Surprising Results

What This Means

Technical Summary: Large Language Model-Driven Small-Capitalization Trading

More like this