Paper: arXiv 2306.14222

Authors: Haohan Zhang, Fengrui Hua, Chengjin Xu, Hao Kong, Ruiting Zuo, Jian Guo

Abstract

The rapid advancement of Large Language Models (LLMs) has spurred discussions about their potential to enhance quantitative trading strategies. LLMs excel in analyzing sentiments about listed companies from financial news, providing critical insights for trading decisions. However, the performance of LLMs in this task varies substantially due to their inherent characteristics. This paper introduces a standardized experimental procedure for comprehensive evaluations. We detail the methodology using three distinct LLMs, each embodying a unique approach to performance enhancement, applied specifically to the task of sentiment factor extraction from large volumes of Chinese news summaries. Subsequently, we develop quantitative trading strategies using these sentiment factors and conduct back-tests in realistic scenarios. Our results will offer perspectives about the performances of Large Language Models applied to extracting sentiments from Chinese news texts.

Complexity vs Empirical Score

  • Math Complexity: 2.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper relies primarily on standard NLP methods and backtesting procedures without advanced mathematical derivations, yet it demonstrates high empirical rigor through a large-scale dataset (394k news items) and realistic trading simulations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Can LLMs Predict Chinese Stock Price<br>Movements via Sentiment Analysis?"] --> B["Methodology & Data<br>Three LLMs<br>Chinese Financial News Summaries"]
  B --> C["Computational Process<br>1. Sentiment Factor Extraction<br>2. Quantitative Strategy Development"]
  C --> D["Back-Testing<br>Simulate Trading in Realistic Scenarios"]
  D --> E["Key Findings & Outcomes<br>LLM Comparative Performance<br>Trading Strategy Viability"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a standardized experimental procedure that combines sentiment extraction from Chinese news with realistic back‑tests of quantitative trading strategies using three distinct LLMs.

Source sentence
We propose an innovative approach that combines sentiment extraction with realistic back-tests of quantitative strategies.

Assumes 6

  • Portfolios are adjusted daily and only at market open (9:30 am Beijing Time). execution
    Source sentence
    Portfolios are adjusted daily and only at market open, 9:30 am Beijing Time (UTC+8:00)
  • Only news generated or acquired before market open are used for sentiment factors. data
    Source sentence
    We only use news generated or acquired before market open.
  • Trades are executed at the Volume‑Weighted Average Price (VWAP) between 9:30 am and 9:35 am. execution
    Source sentence
    we have implemented a more realistic approach by utilizing the Volume-Weighted Average Price (VWAP) between 9:30 am and 9:35 am.
  • A transaction fee of 0.15% of transaction value is applied to all trades. execution
    Source sentence
    We avoid being overly-optimistic about our simulated returns by imposing a transaction fee of 0.15% of transaction value.
Show 2 more
  • The CSI 300 index is used as the benchmark for calculating excess returns. market
    Source sentence
    We use the CSI 300 index as basis when calculating the excess returns.
  • Sentiment scores are mapped to numeric values: GOOD=1, NOT SURE=0, BAD=‑1, and averaged across sources for the same day. model
    Source sentence
    We translate the response into numeric values with GOOD NEWS being assigned as 1, NOT SURE being assigned as 0 and BAD NEWS being assigned as -1.

Data

  • Chinese news summaries text-news proprietary
    Source sentence
    We offer Table 1 to illustrate the distribution of these prime sources within our data-set.
  • News summary items text-news proprietary
    Source sentence
    After the responses for all 394,429 items of news summary have been given.

Costs

Modeled Costs are part of the reported headline results.

0.15% of transaction value (0.05% commission + 0.1% stamp duty) applied to all trades.

Source sentence
We avoid being overly-optimistic about our simulated returns by imposing a transaction fee of 0.15% of transaction value. This includes a 0.05% commission charged by the stock brokerage firms and a 0.1% stamp duty fee paid to the stock exchanges.

Finds 2

  • Erlangshen‑110M outperformed both ChatGPT (GPT‑3.5) and Chinese‑FinBERT on all main performance metrics, achieving the highest annual excess return (24.14%) and Sharpe ratio (0.6841). performance
    Source sentence
    outperformed GPT‑3.5 across all metrics in this task.
  • The continuous sentiment score (0‑1) from Erlangshen‑110M provides better differentiation among news items than the discrete three‑level rating used by GPT‑3.5. other
    Source sentence
    GPT‑3.5 was tasked with returning a discrete sentiment rating from three options (-1, 0, 1), whereas Erlangshen‑110M outputs a continuous score between 0 and 1.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 10 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 10 papers in the archive (9 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.