ChatGPT Improves Momentum Portfolios’ Risk-Adjusted Returns

If you follow stocks, the hard part is telling whether yesterday’s winners still have room to run. This paper tests whether ChatGPT can help by reading firm-specific news in real time and judging whether it supports a stock’s recent momentum, a strategy that buys shares with strong past returns. The authors combine daily U.S. equity returns for S&P 500 constituents with high-frequency news data, then send prompt-engineered queries to ChatGPT right before a stock enters a momentum portfolio. The model assigns scores that affect both which stocks are chosen and how heavily they are weighted. The result is an LLM-enhanced momentum strategy that beats a standard long-only momentum benchmark, with higher Sharpe and Sortino ratios both in-sample and in a truly out-of-sample period after ChatGPT’s pre-training cut-off. The gains hold up under transaction costs, different prompt designs, and portfolio constraints, and they are strongest for concentrated, high-conviction portfolios. The takeaway is simple: in this setting, ChatGPT acts like a real-time reader of financial news, adding extra value to a familiar investing rule.

A hot stock can still be a tired stock. That is the trap in momentum investing. You buy yesterday's winners and hope the run keeps going. This study puts ChatGPT in that job. It feeds the model firm news right before a stock enters a momentum portfolio. Then the model judges whether the news still backs the price trend. If you have ever watched a stock jump on one headline, you know the feeling. News and price do not always move at the same speed. That gap is where the idea tries to earn its keep. The surprise is that a text model can help a price rule without replacing it. It acts like a fast reader in the middle of a trading system.

When price momentum meets the news stream

The main result is simple. ChatGPT-backed momentum beat a plain long-only momentum benchmark. A long-only strategy buys shares, but never bets against them. The edge showed up in the Sharpe ratio. That ratio measures return per unit of risk. It also showed up in the Sortino ratio. That ratio looks at bad swings only. Those gains appeared in the main test and after the model's pre-training cut-off. A pre-training cut-off is the last date in a model's training set. They stayed after transaction costs, prompt changes, and portfolio limits. The strongest lift came from concentrated, high-conviction portfolios. A better score meant a larger bet. A weaker score meant less weight. The score also changed which stocks got in. That shift mattered.

How ChatGPT reads the trade

The setup joins two feeds. One feed holds daily U.S. returns for S&P 500 stocks. The other feed brings in high-frequency news. A large language model, or LLM, is a program that reads and writes text. Prompt-engineered queries are carefully written instructions. They tell ChatGPT when a stock is about to enter the portfolio. ChatGPT then scores whether recent news supports the old price trend. Those scores shape both selection and weight. That makes the model a news judge, not a market oracle. And it does it only when the trade is about to happen. The point is to use text before the price move settles.

Why the news filter matters

The useful idea here is not a magic stock picker. It is a news filter that sits inside a known momentum rule. That matters because the same trade can look strong in price and weak in words. The score adds a second read on the same name. The gains were strongest when the portfolio was concentrated. So the method looks most useful when each bet really counts. It kept working after costs and trading rules. That makes it easier to imagine using in live systems. This result points to a simple role for ChatGPT: read, score, and size.

What the next test should prove

The cleanest next test is another truly out-of-sample slice of the same S&P 500 stock set. That is the hardest check for any text model. It asks if the news score still helps after the pre-training cut-off. If it does, a momentum desk can keep its old price rule and add a live news judge. The surprise is not that ChatGPT replaces the trade. The surprise is that it can sharpen one that already works. That makes ChatGPT a helper inside momentum, not a shortcut around it.