Review of ‘Artificial Intelligence in Day Trading: An Intraday Trading Framework with Economic Indicators and Large Language Model Analysis’ by Zhuokai Chen.
Artificial Intelligence & Machine Learning News
Author: Kenrick Melville
Date: July 31, 2025
Abstract
This review analyzes Zhuokai Chen’s study on integrating large language models (LLMs) into an intraday trading framework for the USTEC CFD (Nasdaq 100). The paper presents a hybrid strategy that combines a volatility based breakout system, triggered by macroeconomic announcements, with sentiment filtering powered by the DeepSeek Chat API. Over a seven‑year backtesting period, the LLM‑enhanced approach demonstrated notable improvements in profitability, achieving a cumulative return of 982.93%, an annualized return of 38.46%, and a Sharpe ratio of 2.11 compared to 1.67 for the baseline strategy. This review evaluates the methodology, contributions, and practical implications of the study, while also identifying its limitations. It highlights the paper’s innovative application of AI‑driven sentiment analysis and suggests areas for refinement before live deployment.
Introduction
Artificial intelligence has quickly reshaped the financial markets, especially in intraday trading. Large language models (LLMs) now play an increasingly important role in converting unstructured data, such as economic reports, into actionable trading signals. Zhuokai Chen’s paper introduces a two‑layered framework that merges a volatility breakout trading system with LLM based filtering to improve trade entries. This strategy focuses on scheduled U.S. macroeconomic events including Federal Open Market Committee (FOMC) announcements, Nonfarm Payrolls (NFP), and Consumer Price Index (CPI) releases blending quantitative indicators with sentiment analysis to sharpen trade selection. The framework attempts to filter out noise and align trades with prevailing market sentiment by employing structured prompts and JSON formatted outputs. This review examines the originality of this approach, the strength of its backtesting design, and its relevance for real‑world algorithmic trading.
Review
Novelty and Contribution
Chen’s work significantly contributes by embedding LLMs within an economic news‑driven trading strategy. Unlike prior efforts in machine learning and natural language processing (NLP), this framework uses real time, prompt‑engineered outputs from DeepSeek Chat to validate trading signals. These outputs provide confidence scores influencing execution decisions, forming a dual‑filtering system. This approach shows a practical, applied use of generative AI in systematic trading and adds to the growing literature on AI‑enhanced event‑driven strategies.
Methodology and Strategy Design
The core system employs a volatility breakout model, using a 200‑minute high/low range combined with market noise adjustments. A Kalman filter and a sigmoid function generate a Trading Indicator (TI) for trade entries. The LLM layer then filters these signals using structured prompts and returns JSON responses with confidence
1
levels. Trades are managed with a 1% stop loss and are closed at 15:50 ET to avoid overnight exposure. This layered design allows dynamic decision‑making by combining quantitative breakout signals with sentiment‑driven assessments.
Performance Results
Backtesting results show that the LLM‑enhanced strategy outperformed the backbone system and a buy‑and‑hold benchmark. Key metrics include:
• Cumulative Return: 982.93% (LLM) vs. 730.59% (baseline) vs. 205.42% (benchmark) • Annualized Return: 38.46% (LLM) vs. 33.53% (baseline)
• Sharpe Ratio: 2.11 (LLM) vs. 1.67 (baseline)
• Win Rate: 52.61% (LLM) vs. 49.74% (baseline)
• Maximum Drawdown: ‑14.1% (LLM) vs. ‑17.4% (baseline) vs. ‑35.5% (benchmark)
Notably, the LLM‑driven version executed fewer trades (1,401 vs. 1,717), indicating that its filtering increased selectivity and improved trade quality.
LLM Stability and Prompt Sensitivity
The study tested several prompt designs, with Prompt 3 emphasizing forecast deviations and sentiment producing the strongest Sharpe ratio (2.11) and maintaining consistency in retesting (2.07). This reinforces the value of thoughtful prompt engineering, although it also underscores the sensitivity of results to prompt design, which is critical for real world deployment.
Limitations and Practical Constraints
While promising, the study’s reliance on backtesting limits its real-world applicability. Key execution related factors such as slippage, API latency, and order delays were not modeled, which may overstate profitability. Additionally, the proprietary LLM based filtering poses risks of hallucination or output drift. Finally, the lack of publicly available code and datasets restricts reproducibility and peer validation.
Conclusion
Zhuokai Chen’s paper presents an innovative framework for combining LLMs with volatility driven trading strategies in an event driven context. The LLM enhanced approach delivered almost 1,000% cumulative returns and a Sharpe ratio above 2.0, underscoring the potential of AI powered sentiment analysis in quantitative trading. However, broader adoption of this method will require further validation, live simulation, and stress testing under extreme market conditions. Greater transparency, including the release of code and modeling of transaction costs, would also strengthen the study’s practical value.
Recommendation: Minor Revision.
While highly innovative, this study needs additional validation and practical enhancements before being considered for publication in a leading financial journal.



