Keywords
How to Cite
Abstract
This study develops an end-to-end trading framework that integrates large language model–based news sentiment with conventional momentum and valuation factors in China’s A-share market. Financial and policy-relevant articles from People’s Daily are collected through an automated crawler, time-stamped, deduplicated, and aligned with the trading calendar to prevent look-ahead bias. A LoRA-fine-tuned LLaMA model converts each article into a continuous sentiment score, which is aggregated into daily market- and sector-level sentiment indices. These indices are combined with momentum and value signals to generate cross-sectional stock scores, while portfolio weights are determined through mean–variance optimization subject to turnover, sector-neutrality, transaction-cost, T+1 settlement, and price-limit constraints. The framework is evaluated on CSI 300 constituents over 2022–2024 using a simulated but market-calibrated dataset designed to approximate realistic trading conditions. The sentiment-augmented strategy records an annualized return of 18.5%, a Sharpe ratio of 1.42, and a maximum drawdown of 10.5%, outperforming both a momentum–value benchmark and the CSI 300 index. The average information coefficient rises from 0.05 for the conventional factor model to 0.10 after sentiment integration, indicating incremental cross-sectional predictive content. The findings suggest that policy-oriented media narratives may provide economically relevant information beyond traditional price and fundamental variables. However, because the empirical results are derived from simulated data, they should be interpreted as evidence of methodological feasibility rather than proof of deployable alpha. Future research should validate the framework using audited real-time data, multiple news sources, longer market cycles, and live-execution tests, and examine signal stability across regulatory regimes and distinct macroeconomic shocks.