The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →AI can pick stocks, but the evidence does not show that it can reliably pick winners or consistently beat the market. Results vary by system and test: a 2025 experiment found that Gemini 1.5 Flash did not consistently outperform simple benchmarks, while separate studies reported strong historical or experimental returns under different methods. Those findings are not interchangeable—and none guarantees what an investor will earn.
Can AI pick stocks or beat the market?
“AI stock picking” can mean very different things. A general-purpose chatbot asked to name stocks is not the same as a purpose-built machine-learning system trained on structured financial and price data. Their inputs, objectives, and testing methods can differ, so a result from one system does not establish that another will work.
The available studies do not support either extreme: that AI stock picking always fails, or that a general-purpose AI can reliably beat a market benchmark. They show results tied to particular models, data, periods, and assumptions.
Gemini 1.5 Flash: no consistent benchmark win
A 2025 study tested Gemini 1.5 Flash with anonymized and randomized U.S. company information, using financial data, price data, or both and investment horizons from one to 36 months. The authors reported that the model did not consistently outperform either an equal-weighted portfolio or the S&P 500, and that risk-adjusted performance declined at longer horizons. The finding is specific to that experiment; it does not establish how every AI system performs. Perlin, Foguesatto, Müller, and Righi, Finance Research Letters (2025).
Recommended Free Tools
A historical machine-learning strategy: strong returns, costly turnover
A 2024 paper evaluated machine-learning stock-selection strategies against S&P 500 benchmarks over 2002–2021. In the authors’ historical evaluation, an equally weighted benchmark returned 11.1% per year and the value-weighted S&P 500 returned 6.4% per year; the paper’s 50-stock ensemble strategy returned 20.8% per year. These are historical study results, not a current investor outcome or a forecast. Tobias Wolff, Journal of Forecasting (2024).
The same study reported annual portfolio turnover of 9.9 to 31.1 times, depending on model and portfolio size, and break-even transaction costs of 5 to 21 basis points. Its authors caution that low transaction costs are needed to exploit the strategies. A backtest return that does not account for trading costs and implementation is not the same as a return an investor could keep. Wolff (2024).
A GPT-4-based framework: promising experimental result
MarketSenseAI, a framework evaluated in a 2024 study, reported excess alpha of 10–30% and cumulative returns of up to 72% in a 15-month S&P 100 experiment. This is the authors’ finding in that experimental setup; it does not show that retail users can reproduce the result, predict future performance, or settle the different result from the Gemini experiment. Fatouros et al., Neural Computing and Applications (2024).
Why do AI stock-picking results differ?
A performance figure is meaningful only in the context of how it was produced. Before comparing two claims, check what the system saw, what it was asked to do, and what costs and risks the reported return includes.
- System type: A general-purpose large language model is not equivalent to a supervised or other purpose-built machine-learning strategy.
- Data and timing: Anonymized or point-in-time data can test a different question from recognizable historical company data. Data availability and timing affect whether a test resembles a real investment decision.
- Evaluation method: A backtest, an out-of-sample test, and live performance are different kinds of evidence. A historical result alone does not establish future performance.
- Universe and benchmark: Results depend on which stocks and geography are included and whether the comparison is an equal-weighted portfolio, a broad index, or another benchmark.
- Horizon and risk: A result over a short period may not hold over a longer one. Returns also need context such as volatility, drawdown, and risk-adjusted performance.
- Trading costs and turnover: Frequent portfolio changes can make fees and execution costs decisive, as the turnover and break-even cost estimates in Wolff’s study illustrate.
- Concentration and diversification: A concentrated portfolio can behave differently from a diversified one; return alone does not describe the risk taken to achieve it.
Are AI trading bots a scam?
Not necessarily. The fact that a strategy uses AI does not by itself make it fraudulent, and the regulator alert does not say every AI investment service is a scam. But promises of certain profits are a serious warning sign. A joint investor alert from the SEC, NASAA, and FINRA gives examples of promotions such as “Our proprietary AI trading system can’t lose!” and “Use AI to Pick Guaranteed Stock Winners!” It cautions investors to be wary of claims that AI can guarantee amazing investment returns. SEC, NASAA, and FINRA, “Artificial Intelligence (AI) and Investment Fraud: Investor Alert” (January 25, 2024).
Before sending money or relying on a service, use the alert’s practical checks:
Rank #4
- Check the registration and disciplinary history of the investment professional.
- Review public-company disclosures through the SEC’s EDGAR system.
- Be cautious of unregistered platforms, unrealistic return promises, false claims, and promotions that may involve pump-and-dump activity.
- Do not make an investment decision solely because a celebrity endorses a product or service.
How should an investor assess an AI stock-picking claim?
Treat a performance claim as a test to scrutinize, not as proof that a tool can predict winning stocks. Ask for enough detail to determine whether the result applies to the service and investing conditions being offered.
- Identify the system: Is it a general-purpose chatbot, a rules-based product, or a purpose-built model trained on structured data?
- Check the evidence type: Is the claim based on a backtest, an out-of-sample evaluation, or observed live results? A historical result should not be presented as a guaranteed future outcome.
- Inspect the comparison: Find the market universe, dates, investment horizon, and benchmark. A return without these details is difficult to interpret.
- Look beyond headline return: Ask about risk-adjusted performance, volatility, drawdowns, diversification, turnover, and the treatment of transaction costs and fees.
- Verify the people and platform: Use the regulator alert’s registration and disciplinary-history checks, and review relevant public-company filings through EDGAR.
- Reject guarantees: No claim that an AI system cannot lose or guarantees stock winners should substitute for independent verification and sound judgment.
What the evidence supports
Research has produced mixed results: one anonymized-data experiment found no consistent benchmark outperformance, while other studies reported strong results in specific historical or experimental settings. Differences in models, datasets, benchmarks, horizons, risk measures, and trading assumptions prevent those results from proving that AI stock picking generally works—or that it universally fails.
For an investor, the useful distinction is between a narrowly described study result and a sales promise. A past experiment can be worth examining; it cannot guarantee that a consumer product will reproduce its performance or that a future portfolio will beat the market.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
