Can ChatGPT Forecast Stock Price Movements?

Can a large language model like ChatGPT predict stock price movements by analyzing news headlines?

That’s the question tackled by researchers Alejandro Lopez-Lira and Yuehua Tang in their recent paper, “Can ChatGPT Forecast Stock Price Movements?” The results are surprising and potentially disruptive for the future of algorithmic trading.

Related reading: Best trading indicators

How the Study Worked: Turning Headlines into Trading Signals

The researchers created a systematic trading strategy using ChatGPT-4 to score company-specific news headlines. The AI rated each headline as good or bad for the stock in question.

Based on these scores, they constructed daily long-short portfolios:

• Long: Buy stocks with positive news scores.
• Short: Sell stocks with negative news scores.

The strategy was rebalanced daily and handled news released both intraday and overnight, focusing on the latter, as over 80% of headlines were published outside regular trading hours.

Entry and Exit Timing: Overnight vs. Intraday News

The researchers split the news flow into:

• Overnight News (released after the previous market close or before 9 a.m. on the trading day)
• Intraday News (released during market hours)

If news came before 9 a.m., trades were executed at the market open and closed by the same day’s close.
If news came after the market closed, positions were entered at the next open and exited at the next close.

The Performance: How Well Did ChatGPT Do?

The paper tested four key strategies from October 2021 to December 2023:

1. Long GPT-4 Portfolio: Buy stocks with good news
2. Short GPT-4 Portfolio: Short stocks with bad news
3. Long-Short GPT-4 Portfolio: Buy good news, short bad news (self-financing)
4. Market Benchmark: Traditional value-weighted market index

Key Results (before transaction costs):

• The Long-Short GPT-4 Strategy earned an astonishing ~650% cumulative return.
• The Short Portfolio alone gained over 300%.
• The Long Portfolio gained around 70%.
• The market benchmark gained almost nothing over the same period.

This suggests that ChatGPT, without access to real-time prices or financial models, can still identify mispriced information in headlines—especially negative news.

Why the Short Side Worked So Well

Interestingly, the strategy’s performance was stronger on the short side, aligning with the paper’s theoretical model.

Negative news often causes delayed price reactions or overlooked risks, which ChatGPT seems to pick up effectively.

Limitations and Caveats

While the raw performance is impressive, it’s important to note:

• The results don’t include transaction costs.
• Performance spikes were partly due to low-news days and equal-weighting, which can be volatile.
• The strategy assumes immediate trade execution at open/close prices, which may not be fully realistic.

These issues highlight the need for robust risk management in real-world applications.

Conclusion: A New Era for AI in Financial Markets?

This paper suggests that ChatGPT—and potentially other large language models—can be more than just productivity tools.

They might represent the next frontier in AI-powered trading, especially for processing qualitative data like news headlines.