Can ChatGPT, Claude and Gemini Pick Stocks? An Evidence-Based Portfolio Audit
Last updated: October 16, 2026
ChatGPT can explain a company's competitive advantage. Claude can organize financial research. Gemini can analyze business information and compare investment alternatives. But can any of these AI assistants consistently select stocks that outperform the market after accounting for risk, fees and taxes?
That is a much harder question than asking an AI chatbot to list 10 stocks it likes.
A convincing answer requires more than persuasive explanations, confident price targets or an impressive backtest. It requires a portfolio with a documented starting date, independently verifiable holdings, a fair benchmark and performance measured over the same period.
The distinction matters. Published studies have produced different findings depending on the model, market, portfolio-construction rules and evaluation period. Those findings are useful evidence, but they do not establish which current AI assistant will deliver the best investment returns.
Contents
- The initial verdict
- What published research actually shows
- How to test ChatGPT vs Claude vs Gemini fairly
- The portfolio performance scorecard
- Why AI stock-picking experiments can mislead
- How to interpret the results
- How investors should use AI
- Frequently asked questions
1. The initial verdict: AI stock selection is not proven to be a reliable market-beating strategy
Large language models can process financial text, summarize company information and generate structured investment theses. Research also suggests that some models can extract useful signals from financial information under particular experimental conditions.
However, the ability to analyze information is not the same as the ability to earn superior risk-adjusted returns in a real portfolio.
| Question | What the evidence supports | What remains unproven |
|---|---|---|
| Can AI generate a stock portfolio? | Yes. Chatbots can propose securities, investment rationales and portfolio weights. | Whether the resulting portfolio is suitable for a specific investor. |
| Can AI process financial information? | Yes. Studies show potential in selected financial-analysis and prediction tasks. | Whether its analysis consistently produces a tradable advantage after costs. |
| Can AI outperform the S&P 500? | Some experiments report favorable results under specific conditions. | Persistent outperformance across models, regimes and prospective holding periods. |
| Is ChatGPT better than Claude or Gemini? | Different models can behave differently on the same task. | A universal ranking of current versions based on independently verified investment returns. |
The practical conclusion is not that AI is useless for investing. It is that an investment process should be judged by evidence of performance, not the apparent sophistication of its reasoning.
2. What published research actually shows
Three strands of research illustrate why an AI portfolio audit needs careful controls.
Study A: Gemini 1.5 Flash and historical stock selection
A 2025 paper in Finance Research Letters, titled “Can AI beat a naive portfolio? An experiment with anonymized data”, examined Google's Gemini 1.5 Flash.
The researchers conducted approximately 30,000 simulations using anonymized information for 1,522 US companies over 20 years. They tested investment decisions across different horizons and different combinations of financial and price information.
Main finding: Gemini did not consistently outperform either a simple equal-weighted portfolio or the S&P 500 in returns and Sharpe ratios. The researchers also found that risk-adjusted performance weakened as the investment horizon increased.
Why it matters: A model can produce plausible company selections without demonstrating consistent superiority over a relatively simple benchmark.
Important limitation: The experiment evaluated Gemini 1.5 Flash under its specific historical-data design. It does not establish the performance of later Gemini versions, nor does it directly compare current Gemini, ChatGPT and Claude portfolios.
Study B: ChatGPT portfolios and investor risk preferences
A September 2025 study in Finance Research Letters, “Stock portfolio selection based on risk appetite: Evidence from ChatGPT”, investigated portfolio construction using ChatGPT.
The researchers examined portfolios designed for different risk preferences in US and European equity markets. They reported that GPT-4o performed best in the studied US market setting, while GPT-4 produced the highest returns in the studied European setting.
The study also examined how ChatGPT adjusted portfolios to different investor risk appetites and whether portfolio compositions were reproducible across repeated prompts.
Why it matters: Results may depend on the model version, market and portfolio-construction method. A conclusion about one model in one setting cannot automatically be extended to another.
Important limitation: These findings relate to the study's experimental settings and models. They do not prove that either model can deliver persistent, live-market outperformance after implementation costs.
Study C: Can ChatGPT predict market reactions to news?
An October 2026 paper in the Journal of Financial Economics, “Can ChatGPT forecast stock price movements? Return predictability and large language models”, examined GPT-4's ability to interpret news headlines using information published after the model's knowledge cutoff.
The authors reported that GPT-4 captured initial stock-market reactions and found information associated with subsequent price drift, particularly for smaller stocks and negative news. However, the paper distinguished the immediate market reaction from later price movements that might be more relevant to a tradable strategy.
Why it matters: Financial-language understanding may contain useful predictive information, but a prediction about a price reaction is not automatically a profitable trading signal.
Important limitation: This was a particular study of GPT-4 and news-based return predictability, not a live contest among today's three consumer AI assistants.
Evidence summary
| Research | Model or task | Key finding | What it does not establish |
|---|---|---|---|
| Finance Research Letters, 2025 | Gemini 1.5 Flash | No consistent outperformance over a naive portfolio or the S&P 500. | Performance of current Gemini, ChatGPT or Claude models. |
| Finance Research Letters, 2025 | GPT-4 and GPT-4o | Results varied by model and market in the studied portfolios. | Reliable future returns from today's ChatGPT service. |
| Journal of Financial Economics, 2026 | GPT-4 and news-based prediction | Evidence of predictive information in selected market reactions. | A profitable, current-model, multi-provider stock-picking strategy. |
Overall assessment: Published research supports further testing of AI-assisted investing. It does not yet provide a simple, defensible answer that ChatGPT, Claude or Gemini is the best stock picker for an ordinary investor.
3. How to test ChatGPT vs Claude vs Gemini fairly
A credible comparison should be designed before the portfolios are generated. That prevents researchers from changing the rules after seeing which model produced the most attractive results.
Step 1: Record the exact model versions
Do not label the test simply “ChatGPT versus Claude versus Gemini.” Each service may offer multiple models, reasoning settings, subscription tiers and research capabilities.
For every run, record:
- The provider and exact model identifier displayed or returned by the service.
- The date, time, subscription tier and reasoning setting.
- Whether web search, financial tools and other external sources were available.
- The complete prompt, unedited answer and source URLs supplied by the model.
Use the public consumer services if the objective is to evaluate what ordinary investors can access. An API-based experiment is also valid, but its findings should be identified as an API experiment rather than assumed to represent the consumer applications.
Step 2: Give all three models the same investment brief
The test should specify the same investable universe, portfolio size, investment horizon and risk constraints. Otherwise, differences in returns may reflect differences in instructions rather than differences in model quality.
For example, require each model to choose 10 liquid, US-listed common stocks for a simulated $10,000 portfolio, with equal weighting and a 12-month holding period.
Exclude options, leveraged products, cryptocurrencies and securities that do not meet the predefined liquidity requirements. State whether foreign listings, ADRs, preferred shares, ETFs and recently listed companies are eligible.
Ask for the model's investment rationale and supporting evidence, but do not allow it to change the rules or substitute new stocks after results start to emerge.
Step 3: Run repeated, independent trials
A single answer is insufficient for a model whose output may change with wording, context and sampling. Use five fresh conversations per provider as a modest initial experiment.
That creates 15 portfolios: five generated by ChatGPT, five by Claude and five by Gemini. Preserve every original output. Report each provider's median return, the range of results and the consistency of its selections.
Five runs are a starting point, not a guarantee of statistical power. A more rigorous study would determine its sample size in advance and use additional independent prompts or evaluation periods.
Do not keep only the best-performing portfolio from each service. That would introduce selection bias.
Step 4: Freeze the holdings before measuring performance
For a prospective test, generate all portfolios before the market data used to evaluate them becomes available.
One reproducible convention is to generate and archive the portfolios after the market closes on the starting date, then use the next trading session's opening prices for simulated purchases. Record the exact timestamps and apply the same convention to every model.
Set a fixed end date 12 months later. The primary test should be buy-and-hold, without discretionary substitutions or rebalancing. A separate experiment can test active rebalancing, but it must be evaluated as a distinct strategy.
Because this is a prospective test, interim results should be labeled as interim. A three-month result cannot establish 12-month outperformance.
Step 5: Use appropriate benchmarks
Comparing three AI portfolios only against one another is inadequate. All three could underperform a passive investment strategy.
Use at least two benchmarks:
- S&P 500 total return: Tests whether the portfolio outperforms a broad US large-cap equity benchmark, including reinvested dividends.
- Equal-weighted S&P 500: Helps distinguish security selection from the effects of equal weighting and exposure to smaller index constituents.
The equal-weighted benchmark must follow a documented rebalancing rule. For live implementation, investable proxies such as SPY and RSP can also be compared, with their fees and tracking differences accounted for.
Use matching dates, currency conventions and dividend treatment. Report a model portfolio's sector concentration so readers can determine whether its result largely reflects a concentrated bet on a rising sector.
Step 6: Include costs and real-world constraints
A backtest that ignores trading costs can overstate profitability. Record simulated commissions, bid-ask spreads, slippage, dividends, corporate actions and any portfolio turnover.
State whether returns are before or after taxes. Tax treatment depends on the investor's jurisdiction and circumstances; one universal tax assumption is not suitable for all readers.
Keep a gross-return calculation for analytical comparison and an explicitly defined net-return calculation for practical interpretation. Use identical assumptions across all three models.
4. A standardized prompt for the experiment
The following template can be supplied to each service in a new conversation. Fill in the dates and model details before running the experiment.
You are participating in a controlled stock-selection experiment. This is a simulated portfolio, not a request for personalized financial advice. Portfolio start date: [DATE] Information cutoff: [DATE AND TIME] Investment universe: US-listed common stocks meeting the predefined liquidity and eligibility rules. Initial simulated capital: US$10,000. Holding period: 12 months. Number of stocks: 10. Portfolio weights: Equal weight. Rebalancing: None during the primary test. Permitted assets: Common stocks only. Excluded: ETFs, options, leveraged products and cryptocurrency. Select 10 stocks that you believe offer the strongest risk-adjusted investment prospects over this horizon. For each stock, provide: 1. Ticker and full company name. 2. Investment thesis and relevant business drivers. 3. Valuation considerations and material downside risks. 4. The principal evidence supporting the thesis. 5. Source URLs and publication dates where available. 6. Conditions that would invalidate the thesis. State material uncertainties and identify any information you cannot independently verify. Do not invent financial figures, citations or current prices. Use only information available by the specified cutoff. Do not change the portfolio after submission. Return a machine-readable table of the 10 selected stocks, the proposed thesis for each, and a concise portfolio-level risk assessment.
If a service cannot comply with the historical information cutoff, record the limitation rather than silently treating the result as a valid historical forecast.
In the primary stock-selection test, use equal weights regardless of any suggested weights. A secondary test can evaluate the models' allocation decisions separately. This avoids confusing stock-selection ability with position-sizing ability.
5. The performance scorecard: what should we measure?
The winning portfolio should not be determined by its headline return alone. Different metrics reveal different aspects of investment quality.
| Metric | What it measures | Why it matters |
|---|---|---|
| Total return | Change in portfolio value, including dividends and applicable costs under the stated method. | Shows whether the initial investment grew. |
| Excess return | Portfolio total return minus benchmark total return over the same period. | Measures outperformance against the chosen benchmark. |
| Maximum drawdown | Largest peak-to-trough decline in portfolio value. | Shows the severity of losses investors may have experienced. |
| Volatility | Variation in periodic returns. | Helps compare the risk taken to earn a return. |
| Sharpe ratio | Excess return relative to return volatility, calculated using a stated risk-free rate and consistent frequency. | Provides one measure of risk-adjusted performance. |
| Turnover and costs | Trading required and its estimated implementation costs. | Frequent changes can erode apparent gains. |
| Sector concentration | Exposure to particular industries, themes or correlated companies. | Reveals whether the result depends on a narrow market bet. |
| Reproducibility | Consistency of picks and results across repeated independent runs. | Helps distinguish repeatable output from a one-off answer. |
How results should be reported
The scorecard below is a reporting template, not a record of observed results. It should remain unfilled until the experiment has been conducted and the returns verified.
| Measure | ChatGPT | Claude | Gemini | Benchmark |
|---|---|---|---|---|
| Median 12-month total return | Pending test | Pending test | Pending test | To be measured |
| Median excess return | Pending test | Pending test | Pending test | Reference: 0% |
| Median maximum drawdown | Pending test | Pending test | Pending test | To be measured |
| Median Sharpe ratio | Pending test | Pending test | Pending test | To be measured |
| Five-run return range | Pending test | Pending test | Pending test | Not applicable |
The primary comparison should use each provider's median portfolio return across the repeated runs. Also report the individual returns and a clearly defined measure of dispersion. Five observations are too few to support sweeping statistical claims, and overlapping results should not be presented as proof of a meaningful difference.
6. Why AI stock-picking experiments can mislead
Problem 1: A convincing explanation can hide a weak investment thesis
Language models are designed to generate coherent text. A stock can therefore receive a polished explanation featuring market size, competitive positioning, an industry catalyst and long-term growth potential without being attractively valued at the purchase price.
Good companies are not necessarily good investments at every price. A rigorous evaluation separates business quality, expected future cash flows, valuation and risk.
Problem 2: Historical data can leak information from the future
If a model is asked to reconstruct a portfolio using historical prices or company information, it may have encountered some of that information during training. Supplying an old date does not guarantee that the model has forgotten subsequent events.
A credible backtest needs to control information availability. Historical results should be treated cautiously unless the researchers can demonstrate that the information available to the model could realistically have been known at the time.
A prospective test is often easier to audit because the portfolio is frozen before the evaluation period begins, although it still requires enough time and observations to support meaningful conclusions.
Problem 3: Survivorship bias can inflate performance
A backtest that uses only companies that exist today can exclude bankrupt, acquired, delisted or otherwise unsuccessful companies. The remaining universe may make historical stock selection appear easier than it really was.
Historical tests should use a point-in-time eligible universe, including relevant delisted securities and corporate actions where data permits.
Problem 4: One exceptional run is not a reliable strategy
AI outputs may vary between prompts or independent sessions. Reporting only the best return rewards lucky selection and conceals instability.
Keep every run, disclose the variation, and avoid selecting prompts or model settings retrospectively because they produce attractive performance.
Problem 5: Benchmark selection can change the conclusion
A growth-heavy portfolio may outperform during a technology rally but underperform when market leadership changes. A portfolio of smaller companies can also behave differently from the market-cap-weighted S&P 500.
At least one broad benchmark and one benchmark that helps isolate the effects of weighting are appropriate. Report exposures alongside returns rather than attributing every difference to superior AI reasoning.
Problem 6: A strong backtest may not survive trading
Bid-ask spreads, slippage, taxes, trading restrictions and portfolio turnover can reduce returns. Small positions in less-liquid companies may be particularly difficult to implement at assumed prices.
A result is more informative when it survives reasonable cost assumptions and can be replicated with publicly available price data.
7. AI stock picks versus AI-managed ETFs
There is another distinction worth making. A chatbot that produces a list of 10 stocks is not the same as an investment fund that uses AI to manage money.
AI-managed funds may combine proprietary models, structured financial data, portfolio constraints, risk controls and trading systems. Their results cannot automatically be attributed to a public chatbot such as ChatGPT, Claude or Gemini.
In an October 14, 2026, Reuters report, some fully AI-managed ETFs were described as having recently outperformed the S&P 500. The report also highlighted their short performance histories, fees and concerns about human oversight. Collectively, the small group of funds described managed around $30 million compared with an ETF industry worth nearly $17 trillion.
Those reported results are a reason to investigate, not sufficient evidence that chatbot-generated portfolios can reliably beat a passive benchmark.
Source: Reuters — AI-powered stock funds bet bots can pick stocks faster and better than humans.
8. How to interpret the results when the test is complete
Once all portfolios have reached the predefined evaluation date, conclusions should follow the evidence rather than the desired headline.
Outcome A: One model wins on return and risk-adjusted performance
Report the magnitude of the difference, its consistency across runs, the drawdown profile and the effect of transaction costs. A single 12-month win still does not establish persistent superiority.
Outcome B: The models produce similar results
Conclude that the experiment did not establish a meaningful performance difference under the chosen rules. Compare research quality, reproducibility and portfolio risks as secondary findings.
Outcome C: All three underperform the benchmark
State that none demonstrated outperformance during the period. Examine whether valuation, sector exposure, turnover or other common biases could explain the results.
Outcome D: One model performs well but takes much greater risk
Do not automatically declare it the winner. Compare risk-adjusted performance and drawdowns, and explain whether the additional return compensated for the additional risk.
Outcome E: Results vary substantially between runs
Highlight the instability. If small changes in the conversation create substantially different portfolios, the system may be less useful as a standalone investment process than a single winning portfolio suggests.
What would count as convincing evidence?
A strong result would have several characteristics:
- Positive excess returns against appropriate benchmarks, measured net of stated implementation costs.
- Acceptable risk-adjusted performance and drawdowns, not merely the highest headline return.
- Reasonable consistency across independent runs and market conditions.
- A transparent methodology, accessible source data and results that independent researchers can reproduce.
- Prospective confirmation over multiple periods, rather than an isolated winning year.
Even these conditions would support a conclusion about the tested strategy, period and model versions, not a guarantee of future returns.
9. How investors should use ChatGPT, Claude and Gemini
The most practical approach is to use AI to improve the research process rather than delegate the entire investment decision to a chatbot.
Use AI to investigate companies
Ask for an analysis of revenue growth, margins, cash generation, debt, competitive advantages, valuation assumptions and potential thesis-breaking events. Verify reported figures against company filings and other authoritative sources.
Ask what could go wrong
Request a bearish thesis for every proposed holding. Investigate customer concentration, competitive pressure, dilution, debt refinancing, regulatory changes and the possibility that projected growth is already reflected in the share price.
Use a common framework to compare candidates
Apply the same valuation measures, financial criteria and risk questions to each company. Ask the model to separate reported facts from estimates, assumptions and speculative claims.
Keep portfolio construction independent of the chatbot
Determine the role of individual stocks within the overall portfolio, the maximum acceptable concentration, the investment horizon and the circumstances that would trigger a review. A good single-stock idea can still be inappropriate when it creates excessive portfolio risk.
Compare the resulting strategy with a diversified, low-cost fund appropriate for the investor's objectives. Passive investing also carries market risk, but it provides a meaningful reference point against which active stock selection can be evaluated.
For related reading, see our guides to stocks to consider researching in 2026 and ETF selection and portfolio construction.
10. Frequently asked questions
Is ChatGPT better at picking stocks than Claude or Gemini?
There is no established universal winner based on comparable, independently verified forward portfolio returns from the current versions of all three services. Results depend on the model version, information provided, market, portfolio constraints and testing method.
Can ChatGPT predict stock prices accurately?
Research suggests that some language models can extract predictive information from financial text and market news under specific conditions. That does not mean they can consistently predict prices or generate profitable trades after costs. Prediction accuracy and investment profitability are different measures.
Can AI beat the S&P 500?
Some experimental portfolios and AI-managed funds have reported periods of outperformance. Other research has found that AI did not consistently beat simple benchmarks. A credible conclusion requires appropriate risk and cost adjustments, comparable dates and longer-term evidence.
Should I buy stocks recommended by an AI chatbot?
Do not buy a stock solely because a chatbot recommends it. Independently verify the company's financial information, valuation, risks and suitability for your portfolio. AI-generated citations and numerical claims can be wrong or incomplete.
Is a historical AI stock-picking backtest trustworthy?
Not automatically. Check for look-ahead bias, survivorship bias, unrealistic execution prices, transaction-cost omissions and excessive strategy optimization. A prospective test with frozen holdings provides a more direct way to evaluate a strategy's forward performance.
How long should an AI portfolio be tested?
A 12-month period is a useful initial evaluation horizon, but one year is not enough to establish durable skill across market conditions. Multiple years, different market regimes and a larger number of independent trials provide stronger evidence, provided the test is designed appropriately.
Can AI replace a financial adviser or professional investment analyst?
AI can support research and analysis, but it does not automatically provide regulated advice, fiduciary oversight or accountability for investment losses. Its outputs require independent verification and human judgment.
Conclusion: Test the portfolio, not the chatbot's confidence
ChatGPT, Claude and Gemini can make investment research faster and more structured. Published studies suggest that language models can identify useful information and sometimes produce promising portfolio results. They also show why results must be tied to the model, task and experimental design rather than generalized to every investment decision.
The decisive question is not which chatbot produces the most convincing list of stocks. It is which strategy, if any, delivers reproducible returns after accounting for risk, costs and the benchmark an investor could reasonably have owned instead.
Until a controlled comparison of current model versions has been completed and its results independently checked, declaring a winner would be premature.
References and further reading
- “Can AI beat a naive portfolio? An experiment with anonymized data.” Finance Research Letters, 2025. https://doi.org/10.1016/j.frl.2025.107126
- “Stock portfolio selection based on risk appetite: Evidence from ChatGPT.” Finance Research Letters, 2025. https://doi.org/10.1016/j.frl.2025.107517
- “Can ChatGPT forecast stock price movements? Return predictability and large language models.” Journal of Financial Economics, 2026. https://doi.org/10.1016/j.jfineco.2026.104335
- Reuters. “AI-powered stock funds bet bots can pick stocks faster and better than humans.” October 14, 2026. Read the report
Related on One Day Advisor:
- Best AI for Investment Research in 2026: The Complete Comparison Guide
Editorial standard: Distinguish model-generated claims from independently verified financial data. Update this article when a controlled portfolio test produces verified results, and preserve the original methodology when reporting revisions.
Comments