Study finds AI trading strategies underperform buy-and-hold investing over 20-year period

For anyone who’s spent the last two years watching AI promise to revolutionize everything from radiology to restaurant reviews, here’s a sobering data point from the world of finance: a comprehensive academic study found that large-language-model trading strategies failed to beat the oldest trick in the investing playbook. Just buying and holding.

The research, titled “Can LLM-based Financial Investing Strategies Outperform the Market in Long Run?”, tested AI-driven trading approaches across more than 20 years of market data and over 100 stock symbols. The conclusion was blunt. LLM strategies consistently underperformed a simple buy-and-hold benchmark.

What went wrong with the AI traders

During bullish market phases, the LLM strategies were excessively conservative. They effectively left money on the table by failing to capture the full upside of rising markets.

Then, when markets turned bearish, the models flipped to being overly aggressive. Inadequate risk controls meant they absorbed significant losses precisely when capital preservation mattered most.

Why this study matters more than previous AI trading research

Previous research on LLM investing strategies typically tested on short timeframes, often under two years, and used a limited selection of assets, sometimes fewer than 10 to 30 symbols.

The researchers, including scholars from the University of Edinburgh and UCLA, built a backtesting framework called FINSABER specifically designed to address the biases that plagued earlier studies. Survivorship bias, where failed companies are excluded from historical datasets, making past returns look artificially rosy. Look-ahead bias, where models inadvertently use future information they wouldn’t have had in real time. Data snooping, where strategies are tweaked until they fit historical data but crumble in live markets.

The paper was accepted for presentation at KDD ’26, one of the premier conferences in data science and knowledge discovery, scheduled for August 2026.

The crypto-shaped hole in the research

One notable limitation: the study focused entirely on traditional equity markets. No Bitcoin, no Ethereum, no memecoins. Crypto assets were completely absent from the analysis.

The study doesn’t prove AI trading works in crypto. It simply didn’t test it.

What this means for investors

For retail investors evaluating AI-powered trading platforms, whether in equities or crypto, the study is a useful reality check. Short-term backtests and demo accounts can paint a misleading picture. A strategy that looks brilliant over six months might quietly bleed capital over six years.