How to Use AI Tools for Portfolio Optimization: Asset Allocation and Risk Balance
A practical afternoon test plan for checking AI portfolio calculations and risk scenarios while separating documented vendor features from unverified accuracy claims.
Sources checked 2 Oct 2026
Use ChatGPT, Claude, and Gemini to draft a portfolio-optimization workflow: prepare a file, request transparent calculations and scenarios, then verify every output. Their official pages document features, limits, data settings, and warnings, but they do not establish that any assistant can allocate assets or balance risk correctly. Chat Picker has not tested them for this use; this page is an afternoon test plan, not a ranking or advice on finances, taxes, or law.
Know these limits before you start
File handling comes first: use de-identified test data, not a full account statement.
- ChatGPT: OpenAI's File Uploads FAQ says spreadsheets are limited to approximately 50 MB, Free users are limited to three file uploads per day, and every file has a 512 MB hard limit, as read on October 1, 2026.
- Claude: Anthropic's file-upload page sets chat uploads at 500 MB per file and project files at 30 MB; uploading XLSX files requires code execution and file creation, as read on October 1, 2026.
- Gemini: Google's Gemini file-help page documents up to 10 supported files per prompt, with a 100 MB limit for each non-video file and 2 GB for each video, as read on October 1, 2026.
OpenAI says ChatGPT may sound confident when wrong and urges users to verify important information. Anthropic says not to rely on Claude as the only source of truth and to inspect original sources. Gemini Apps may show sources in a side panel, while Google's guidelines describe LLM outputs as probabilistic and acknowledge training-data limits. A citation is not a reproduced calculation.
Before using real holdings, review account controls. OpenAI's pricing page says an opt-out is available on Free, Go, Plus, and Pro. Anthropic's pricing page says opt-out is available on Free, Pro, and Max, while Team is not trained on by default. Google's plans page linked to data-handling information but stated no training setting, so Chat Picker's matrix marks it “Not verified.” The site's method explains that rule.
OpenAI's usage policies prohibit automated high-stakes decisions in sensitive areas without human review. Anthropic says elevated-risk use should integrate human expertise, with qualified-professional review of covered recommendations. Google says Gemini should not produce inaccurate outputs that could significantly harm someone's finances, while also noting data limits and probabilistic outputs. These warnings do not validate a portfolio recommendation.
The test plan
-
Check Sharpe-ratio arithmetic. Prepare
portfolio_returns.csvwith dated rows, portfolio returns, risk-free rates, and return frequency. Give this instruction: “Use only the attached file. Show every formula, unit conversion, and row count used for an annualized Sharpe ratio; flag missing inputs before calculating.” Recalculate it independently. Log mismatches, hidden assumptions, and sensitivity to a corrected input. -
Test scenarios. Prepare
allocation_inputs.csvwith current weights, candidate assets, constraints, and labeled baseline and stress scenarios. Instruction: “For every scenario, propose alternative weight sets and separate supplied inputs from assumptions. State when expected return, volatility, or drawdown cannot be calculated.” Log invented inputs, weights that breach constraints, and explanations that change when an assumption changes. -
Inspect tail-risk hedges. Prepare
options_positions.csvwith the underlying, expiration, strike, premium, quantity, and contract terms. Instruction: “Explain each position's payoff under labeled terminal-price cases. Do not infer volatility, Greeks, liquidity, or costs; list every missing input.” Log fabricated contract terms, inconsistent payoffs, and confusion between protection and speculation. -
Test tax-loss harvesting. Prepare
tax_lots.csvwith acquisition date, cost basis, current value, unrealized gain or loss, account type, and tax year; put the relevant jurisdiction in a separate note. Instruction: “Identify what cannot be determined, then explain possible steps conditionally. Do not assert eligibility or a tax result.” Log invented basis figures, ignored dates, and unqualified tax conclusions. -
Compare rebalancing rules. Prepare
current_values.csvandtarget_weights.csv. Instruction: “Calculate weight gaps from the attached files. Compare a drift-band rule with calendar rebalancing, and list the cost and tax inputs each would require.” Recalculate every weight. Log arithmetic errors, arbitrary thresholds, and omitted constraints. -
Check plain-language output. Prepare a short reviewer brief stating the portfolio objective, horizon, liquidity need, and risk limits, without recommending a trade. Instruction: “Create a checklist containing assumptions, calculations, uncertainty, missing data, and unresolved questions. Do not issue a trade instruction.” Log changed numbers, buried caveats, vague action verbs, and claims that cannot be traced to an input.
How to read your results
Log the input, prompt, source check, caveat, and result for every item. Mark each item Pass, Review, or Stop. Use Pass only when you reproduced every material number and traced every material assumption. Use Review for missing context or weak explanation. Use Stop for invented data, contradictory arithmetic, or output that bypasses required human review.
Do not average these labels into a score. A Stop rules out that workflow until the cause is checked; repeated Review items show what to revise in the prompt or plan. Base your decision on your log, the data controls you accept, and the matrix features you actually need.
Where the plans differ
Open the ChatGPT, Claude, and Gemini comparison matrix and read the rows for Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Context window in the app, How usage limits are described, Training on your chats, and Ads. Check the file limits above separately because they affect whether your test material fits.
As read on October 1, 2026, OpenAI's pricing page lists ChatGPT Plus at $20 per month; Anthropic's pricing page lists Claude Pro at $20 monthly or $17 per month with annual billing and $200 up front; and Google's US AI plans page lists Google AI Pro at $19.99 per month. The same pages do not publish exact message counts for ChatGPT or Claude, while the matrix marks Gemini's app token context as “Not verified.” Confirm current prices and limits before subscribing; Chat Picker's method leaves unconfirmed cells blank rather than filling them in.
Sources