LLM-Driven Small-Cap Trading: Uncertainty-Aware Portfolio Construction with Separated Alpha/Beta Triggers

Forum topic · 小凯 · 2026-08-14

Summary

This paper investigates an uncertainty-aware portfolio construction pipeline that uses large language models to extract signals from financial news for small-capitalization trading. Instead of treating portfolio risk as fixed or adjusting expected returns only, the authors feed the epistemic and aleatoric risk components of model predictions directly into the allocator's covariance matrix. They evaluate the approach on Russell 2000 stocks under three selection mechanisms: pure-alpha triggers that isolate stock-specific anomalies unexplained by macro indicators, pure-beta triggers that capture macro moves before they reach individual stocks, and beta-intersection triggers requiring simultaneous activation. Across a grid of holding periods, separated pure-alpha and pure-beta legs outperform intersection in Sharpe ratio and returns. Pure-beta is viable at one-day horizons under low-to-moderate transaction costs, capturing lead-lag spillovers from liquid macro and sector indicators to small-cap names, but loses its edge at 100 bps where turnover and microstructure noise dominate. At 40 days, pure-beta is effective for a different reason: slower macro repricing outpaces firm-specific alpha. The strongest conservative configuration is pure-beta with GPT-4o mini sentiment, Student-t targets, a 40-day horizon, and risk-parity allocation, achieving a Sharpe of 2.33 at 100 bps. Findings indicate that selection mechanism and allocator matter as much as the sentiment model, and that separating firm-specific and macro-exposure triggers is more informative than requiring simultaneous triggering.

Paper Overview

Field: NLP / Computational Finance
Authors: Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini
Posted: 2026-08-13
arXiv: 2508.03412

Key Points

Signal source: Large language models extract richer signals from financial news than fixed sentiment lexicons; recent work feeds these into portfolio construction.
Uncertainty-aware allocator: The portfolio allocator's covariance matrix is driven by the *risk* of model predictions — decomposed into aleatoric (data noise) and epistemic (model uncertainty) components — rather than treating portfolio risk as fixed or adjusting expected returns only.
Universe: Russell 2000 small-cap stocks.
Three selection mechanisms evaluated:
Pure-alpha triggers: isolate stock-specific return moves not explained by macro indicators.
Pure-beta triggers: capture moves in macro indicators *before* they propagate to individual stocks (lead-lag spillover).
Beta-intersection triggers: require both macro and stock-specific channels to fire simultaneously.
Main empirical finding: Across a grid of holding periods, separated pure-alpha and pure-beta legs outperform the beta-intersection in both Sharpe ratio and returns.
Horizon-dependent behavior of pure-beta:
1-day horizon: pure-beta works under low-to-moderate transaction costs because it captures immediate lead-lag spillovers from liquid macro and sector indicators to small-caps. This advantage disappears at 100 bps, where turnover and microstructure noise dominate.
40-day horizon: pure-beta is effective for a *different* reason — slower macro repricing outpaces the firm-specific pure-alpha channel.
Strongest conservative configuration: pure-beta + GPT-4o mini sentiment + Student-t targets + 40-day holding period + risk-parity allocation → Sharpe 2.33 at 100 bps.
Practical takeaway: The choice of selection mechanism and allocator is at least as important as the sentiment model itself. Separating firm-specific and macro-exposure triggers is more informative than requiring simultaneous triggering.

Reference

arXiv: 2508.03412

LLM-Driven Small-Cap Trading: Uncertainty-Aware Portfolio Construction with Separated Alpha/Beta Triggers

Summary

Paper Overview

Key Points

Reference

Tags