Paper Overview
Field: NLP / Quantitative Finance Authors: Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini arXiv: 2508.03412
Summary
Large language models can extract richer signals from financial news than fixed sentiment lexicons, and recent work has explored feeding such signals into portfolio construction. This paper studies an uncertainty-aware construction in which model-predicted risk — decomposed into aleatoric and epistemic components — is fed directly into the portfolio allocator's covariance matrix, rather than treating portfolio risk as fixed or only adjusting expected returns.
The pipeline is evaluated on Russell 2000 stocks under three stock-selection mechanisms:
- Pure-alpha trigger: isolates stock-specific anomalous moves not explained by macro indicators.
- Pure-beta trigger: captures macro indicator moves before individual stocks trigger.
- Beta-intersection trigger: requires both channels to fire simultaneously.
- Across a grid of holding periods, the separated pure-alpha and pure-beta legs generally outperform the beta intersection on Sharpe ratio and returns.
- At the one-day horizon, pure beta works at low and medium transaction costs because it captures immediate lead-lag spillovers from liquid macro and sector indicators into small-cap exposure; this advantage vanishes at 100 basis points, where turnover and microstructure noise dominate.
- At the 40-day horizon, pure beta works for a different reason: slower macro repricing outweighs the firm-specific pure-alpha channel.
- The strongest conservative configuration is pure beta with GPT-4o mini sentiment, a Student-t objective, a 40-day holding period, and risk-parity allocation, reaching a Sharpe ratio of 2.33 at 100 basis points of transaction cost.
Key Findings
Implication
The choice of stock-selection mechanism and portfolio allocator is at least as important as the sentiment model itself, and separating firm-specific and macro exposure triggers is more informative than requiring both to trigger simultaneously.
---
*Auto-collected on 2026-08-14.*