AI vs Stock Analysts

This paper evaluates ChatGPT-4o’s ability to perform complex financial data analysis, including zero-shot prompting, time series plotting, risk/return metrics (e.g., Sharpe ratios, CAPM beta), and ARMA-GARCH modeling. Using daily return data for the 30 Dow Jones firms (2019–2023), the authors benchmarked GPT-4o’s responses against Stata. The model showed a strong grasp of financial concepts, generated usable Python code, and offered insightful narrative analysis. However, it often required user corrections for code bugs, naming mismatches, and visualization limitations. The study concludes that GPT-4o has high potential for financial research and practice, provided users validate outputs carefully.

View Paper

This study investigates whether ChatGPT can be used to forecast stock price movements by analyzing financial news headlines. The authors prompt ChatGPT-4 to classify news as positive, negative, or neutral for specific S&P 500 stocks, and use this sentiment as the basis for a simulated trading strategy. Back-testing with post-2021 news shows impressive results: a cumulative return between 545% and 655% over two years and a daily alpha of approximately 0.38%. The findings suggest that large language models can extract meaningful predictive signals from unstructured text, though results are based on simulations without accounting for transaction costs or market impact.

View Paper

Assistant professor Alejandro Lopez-Lira (University of Florida) conducted both academic simulations and live experiments using ChatGPT, DeepSeek, and Grok to test their stock-picking capabilities. In a 2023 back-test using over 134,000 news headlines post-2021, ChatGPT achieved a 650% cumulative return with a daily alpha of 0.38%, outperforming benchmarks. In collaboration with the Autopilot investment app, Lopez-Lira now manages real portfolios based on AI model outputs. The models analyze macroeconomic data, corporate fundamentals, and recent headlines to assign investment scores and build portfolios of stocks and ETFs. While promising, the article notes practical limitations like outdated training data, model hallucinations, and real-world trading frictions.

View Article