AI stock picking has been touted as a shortcut to riches, but a 10-month study by Spain's CNMV (the National Securities Market Commission) reveals that large language models (LLMs) are not inherently bad investors—they are bad at following vague instructions, which is exactly how most people use them.
The study, conducted by CNMV researchers Ricardo Crisóstomo and Diana Mykhalyuk, put four popular AI models to the test: ChatGPT, Gemini, DeepSeek, and Perplexity. From April 2025 to January 2026, each model was asked monthly to identify the five best-performing stocks in the Ibex 35 index to buy and the five worst-performing to short sell. The results were then compared against actual market performance, with no historical data cherry-picking—the real market was the sole arbiter.
How the CNMV Study Worked
The methodology was straightforward yet demanding. Each month, the models received the same prompt: pick the top five and bottom five Ibex 35 stocks for the coming month. The researchers then tracked the actual returns of those selections. This live-testing approach eliminated the benefit of hindsight and provided a realistic assessment of the models' predictive capabilities.

What makes this study particularly relevant is its focus on real-world application. Unlike backtests that can be gamed, this experiment forced the models to make forward-looking predictions in a dynamic market environment.

Key Findings: AI Models Evolved Over Time
One of the most significant aspects of the study was the evolution of the models' performance. According to the excerpt, the models showed noticeable improvement over the 10-month period. This suggests that LLMs can learn and adapt when given consistent feedback, even without explicit retraining.
However, the study also highlighted a critical flaw: the models struggled with vague instructions. When asked simply to "pick the best stocks," they often relied on superficial patterns or generic advice rather than deep financial analysis. This is precisely how most retail investors use AI—with broad, unspecific prompts—leading to suboptimal results.
- Vague prompts lead to poor performance: Models need specific, detailed instructions to provide useful investment insights.
- Models can improve with feedback: Over time, the AI adapted to the task, suggesting a learning curve.
- Real market testing is crucial: The study used live market data, avoiding the pitfalls of historical backtesting.
- AI is not a magic bullet: Investors should not expect consistent outperformance without careful prompt engineering and human oversight.
Why Vague Instructions Sabotage AI Stock Picks
The CNMV study underscores a fundamental issue: LLMs are not financial analysts. They are pattern-recognition engines that generate text based on probabilities. When asked to pick stocks without context, they may default to well-known companies or recent trends, which doesn't necessarily translate to future returns.
For example, if a user asks, "What stocks should I buy?" the model might list Apple, Microsoft, and Amazon—solid companies, but not necessarily the best short-term picks. Without specific criteria like sector, risk tolerance, or time horizon, the AI's recommendations are essentially educated guesses.
The Role of Prompt Engineering
The study implies that prompt engineering is critical. Investors who provide detailed instructions—such as "Identify five Ibex 35 stocks with strong momentum and low volatility for a one-month holding period"—are likely to get more actionable insights. However, even then, the AI's limitations must be acknowledged.
Comparison of AI Models in the Study
The table below summarizes the four models tested and their general characteristics based on the study's context. The longer history behind the story is traced in CNMV's AI Stock Picks: 10-Month Study Reveals Surprises.
| Model | Developer | Key Strength | Key Weakness |
|---|---|---|---|
| ChatGPT | OpenAI | Broad knowledge base | May default to popular stocks |
| Gemini | Integration with real-time data | Inconsistent with vague prompts | |
| DeepSeek | DeepSeek | Strong analytical reasoning | Limited financial-specific training |
| Perplexity | Perplexity AI | Real-time web search | May over-rely on recent news |
While the study did not provide a direct ranking, it noted that all models showed improvement over time, suggesting that continuous interaction and feedback can enhance performance. The background to this story is set out in Best Free AI Tools for Students: NotebookLM, Perplexity. A related thread runs through MOVA Robots at IFA 2026: Touch, Watering, and Fruit Picking. Earlier coverage in Best Value Smartphone 2026: Why Picking a Winner Is Harder Than Ever shows how the argument developed.
What This Means for Retail Investors
The CNMV's study serves as a cautionary tale for anyone considering AI-driven stock picking. The message is clear: AI can be a useful tool, but it is not a substitute for due diligence. Investors should use AI as an assistant, not an oracle. A closer look at the same subject appears in Best Tablets for Students: Top 7 Picks for Studying. Anyone following the wider picture can start with Idena Toy Reviews: Best Picks & Top Picks 2026. The same theme surfaces in Clementoni Toy Reviews: Best 2026 Picks, which takes a different angle.
Moreover, the study highlights the importance of specificity. Instead of asking, "What should I buy?" investors should frame questions with clear parameters. For instance: "Based on recent earnings reports and technical indicators, which five Ibex 35 stocks are likely to outperform in the next month?" This approach gives the AI a fighting chance to provide value. Context on how this played out before sits in Ai Video: latest articles. A companion piece, US and China AI Race: Stanford Report Shows Gap Nearly Closed, approaches the same subject from another angle.
FAQ
What was the main conclusion of the CNMV study on AI investing?
The study found that LLMs are not inherently bad investors; they perform poorly when given vague instructions, which is how most people use them. With specific prompts, they can improve over time.
Which AI models were tested in the CNMV study?
The study tested ChatGPT, Gemini, DeepSeek, and Perplexity over a 10-month period from April 2025 to January 2026.
Can AI models like ChatGPT really predict stock market winners?
While the models showed some improvement over time, they are not reliable predictors. Their performance depends heavily on the quality of instructions and market conditions. They should be used as tools, not as sole decision-makers.
How should investors use AI for stock picking?
Investors should provide detailed, specific prompts that include criteria like risk tolerance, time horizon, and sector focus. AI should complement, not replace, fundamental analysis and professional advice.