AI Fund Selection: Build a Personal Investment Research Workflow in 10 Minutes
A practical guide to turning a general-purpose AI assistant into a configured fund research assistant. It explains why one-line prompts produce vague or unverifiable answers, introduces a three-layer setup of experts (role), skills (method), and connectors (data), and walks through a complete workflow: measurable screening, return attribution, portfolio evaluation, and verification. Includes a reusable prompt template, a YAML screening configuration, reproducible Python snippets for annualized return, maximum drawdown and Sharpe ratio, sample output tables, a five-step verification routine, and a list of common mistakes and risk boundaries. The data examples are drawn from the Chinese mutual fund market, and the content is research support rather than investment advice.
1. Why AI Fund Selection Fails: the Hard Part Is Configuration, Not Questioning
Ask a general-purpose model to “recommend a fund” and you will usually get something like “invest regularly in a broad-based index fund” — accurate, but useless. Ask “how has this fund performed over the last three years” and it may return numbers that look precise but cannot be traced to any source.
Three problems cause this:
- There are too many funds. The Chinese market alone has more than 10,000 publicly offered funds across equity, hybrid, bond, index, QDII (cross-border), fund-of-funds and other categories. Even inside “active equity” you find growth, value and balanced styles, so comparing headline returns means little on its own.
- The information is scattered. Historical NAVs, periodic reports, manager changes, asset-size shifts and fee adjustments live in different places, which makes systematic comparison hard for a retail investor.
- General models have limits. Their training data has a knowledge cutoff, so live NAVs and current holdings may be out of reach, and their probabilistic generation makes them prone to fabricating plausible-looking details — mixing up share-class codes, or inventing performance figures.
So the first step toward using AI for fund selection is not asking “which fund should I buy.” It is configuring the assistant into a research assistant with a defined role, a defined method, and defined data. The example below uses a generic AI assistant; the same structure maps onto other similar products.
2. Core Idea: a Three-Layer Configuration
A productive AI assistant is usually made of three parts:
- Expert (role) — decides from which perspective it works. It sets the professional direction, like a job description in a research process.
- Skill (method) — decides what steps it follows. It defines the operating procedure, like a reusable standard operating procedure (SOP).
- Connector (data) — decides what information it relies on. It defines where data comes from and whether it can refresh in real time.
A complete fund research workflow looks like this:
Define the objective → gather data → screen by rules → analyze in depth → evaluate the portfolio → human review → periodic re-run
The three layers answer “who does the work, how it is done, and on what basis.” Once configured, every later research session effectively invokes the same process.
3. A 10-Minute Setup Checklist
| Step | What to do | Goal |
|---|---|---|
| Install and sign in | Download the client or open the web version, sign in, and confirm which features are available | Get a working environment and know the capability boundaries |
| 1. Expert | In the finance/investing category, pick an expert whose research focus is fund research; read the description and switch if it does not fit | Fix the working perspective and professional scope |
| 2. Skills | Install three skills: fund screening, fund analysis, portfolio evaluation; invoke them as tasks require | Standardize a reusable research method |
| 3. Connectors | Connect market data and fund profile sources; confirm coverage and refresh frequency | Keep data current and traceable |
| Self-check | Test the output with one or two questions whose answers you already know | Run an acceptance test before real work |
3.1 The Expert: Give the AI a Clear Direction
Open the assistant and look at the left-hand menu, which normally lists experts, skills and connectors (names and positions vary between versions). For an expert, enter the finance and investing category, choose a role whose research focus is fund research, then read its description to confirm it is built for fund analysis rather than short-term stock trading or insurance planning. If the fit is poor, switch to another.
What if no ready-made expert fits? Most assistants support custom roles or custom prompts. Describe four things:
- Role: a fund researcher with many years of experience;
- Objective: screen and evaluate funds for investors at a specified risk level;
- Constraints: do not use data from unconnected sources; every conclusion must cite its source;
- Output: a table with fixed fields plus step-by-step explanations.
Why spend a minute on this? The same question produces visibly different output with and without a role:
- Without a role, answers tend to become lists of information, and analytical questions may be skipped entirely.
- With a role, the output carries a professional structure — for example, risk metrics first, then manager and holdings, and only then returns.
3.2 The Skills: Teach the AI How to Do It
Skills address methodology. The screening skill narrows the whole market down, the analysis skill performs a deep check on candidate funds, and the portfolio evaluation skill tests whether several funds work together sensibly. When installing, prefer skills that relate directly to fund research and whose descriptions are clear. Afterwards, if the platform supports it, inspect each skill’s input and output requirements to make sure it can handle your data.
Skills are typically used as a pipeline:
- Step 1: use the screening skill to reduce thousands of funds to a dozen or so candidates;
- Step 2: use the analysis skill to go deeper on each one, focusing on where returns come from and how stable they are;
- Step 3: use the portfolio evaluation skill to check how the candidates behave as a group.
If the platform offers no ready-made skills, write the steps into the prompt instead — the effect is similar.
3.3 The Connectors: Bring the Data In
Skills determine the method; connectors determine how fresh and how trustworthy the information is. Connect at least three types of data:
- Live market data: current levels and performance for indices and fund NAVs;
- Fund profiles: category, size, fees, the manager’s start date and other basic facts;
- Periodic reports: holdings, size changes and strategy commentary from quarterly, semi-annual and annual reports.
When connecting, verify three things:
- Coverage — does it include the fund categories you care about (cross-border funds, index funds, and so on);
- Refresh frequency — is NAV data updated daily;
- Traceability — can every data point be mapped back to a source.
What if no connector is available? Two workarounds:
- Upload documents such as fund periodic reports and contract summaries into a knowledge base or workspace and use them as the analysis basis;
- State explicitly in the prompt: “label the source and cutoff date for every data point; write ‘not retrieved’ for anything missing; do not estimate.”
Be clear about the trade-off: information obtained this way is bounded by the documents’ cutoff dates, so it is less current than a live connector.
3.4 Self-Check: Validate Against Known Answers
Before starting real research, test with one or two questions you already know the answer to:
- “Roughly what percentile is the CSI 300 index’s current P/E within its last five years?”
- “For a fund you know well, which year did the current manager start running it?”
If the output is well structured, clearly sourced, and broadly matches what you know, you are ready to work. If it is vague or plainly wrong, go back and check whether the expert fits, the skills are installed correctly, and the connectors are actually working.
4. Start With a Good Question: Turn “What to Buy” Into Executable Steps
Configuration is only the foundation. When you ask, stating the objective, the constraints and the output format in one go noticeably improves quality. You can adapt this prompt directly:
[Objective]Screen equity funds for an investor with a moderate risk profile (R3 on the commonfive-level scale) and build a candidate pool.[Screening criteria]1. Type: active equity or equity-biased hybrid;2. At least 5 years since inception; current manager at least 3 years in the role;3. Assets between 200 million and 10 billion (CNY);4. Top 30% among peers over the past 5 years; maximum drawdown no worse than -35%;5. Management fee plus custody fee no more than 1.5% in total.[Analysis requirements]1. For each fund, break down the sources of return: sector allocation, securityselection, and changes in position size, with each contribution quantified;2. Identify the style: large / mid / small cap, and growth / value / balanced;3. Propose a three-fund portfolio, explaining correlation, concentration andrebalancing rules.[Output requirements]1. Output the screening result table first, with fixed fields:fund name, code, type, 1/3/5-year return, maximum drawdown, Sharpe ratio,manager tenure, assets, total fee, data source;2. Label the source and cutoff date for every data point;3. Write "not retrieved" for missing data; do not estimate;4. Close with a separate section: "Limitations and uncertainties of this proposal".[Execution steps]Step 1: call the screening skill to get the candidate list -> Step 2: call the analysisskill on each fund -> Step 3: call the portfolio evaluation skill on the candidategroup -> Step 4: compile the report.
This prompt has four strengths:
- The objective is specific: an R3 risk level is more precise than “somewhat conservative”;
- The conditions are measurable: tenure, drawdown caps and fee caps can all be checked directly;
- The steps are traceable: it is obvious which skill each step calls;
- The output is verifiable: sources, cutoff dates and missing values all have explicit handling rules.
5. Screening: Shrink “Ten Thousand Funds” to “A Dozen”
Follow a “hard filter first, then rank” order: use objective conditions to eliminate obviously unsuitable funds, then rank what remains by risk-adjusted return.
| Dimension | Key metrics | What to watch |
|---|---|---|
| Basic eligibility | Type, years since inception, size range | Too small carries liquidation risk; too large may reduce flexibility to reposition |
| Performance | Period returns, annual win rate, peer ranking percentile | Percentiles are more stable than absolute rank numbers |
| Risk control | Maximum drawdown, volatility, downside risk | Drawdown largely determines the holding experience |
| Risk-adjusted return | Sharpe ratio, Calmar ratio | Return earned per unit of risk taken |
| Fund company | Fixed income and equity research teams | Active products depend heavily on the platform |
| Fund manager | Tenure, number of products managed, change frequency | Whether the tenure window matches the evaluation window matters |
| Holdings profile | Top-holding concentration, sector concentration, turnover | Higher concentration means a sharper style and possibly higher volatility |
| Cost | Management fee, custody fee, sales service fee | Share classes A, C, and so on suit different holding periods |
| Size trend | Size growth or decline, changes in institutional ownership | A supplementary signal of how the market is voting with money |
A few definitions and caveats:
- Peer ranking percentile: rank all funds in the same category by return, then divide a fund’s position by the total number of peers. A top-30% ranking roughly means a percentile below 30; calculation conventions differ slightly between data providers.
- Annual win rate: the share of the past N calendar years in which the fund beat its benchmark or the peer median.
- Manager tenure: count only the continuous period managing this fund; performance before and after a manager change must be excluded.
- Size and liquidation: fund contracts commonly stipulate that if net assets fall below a threshold (a typical figure in the Chinese market is CNY 50 million) for a number of consecutive trading days, or if the number of holders becomes too small, the manager must publish an advisory notice or propose a remedy; in extreme cases liquidation proceedings may follow.
- Share classes: Class A typically charges an upfront subscription fee; Class C typically accrues a daily sales service fee and charges no subscription fee. Generally C is cheaper for short holding periods and A for long ones — the contract and fee notices govern. Always compare performance within the same share class.
Use Configuration Instead of Repeated Conversation
Turn the screening criteria into a configuration file and let the AI execute it. Conditions become versionable and reproducible:
# Example screening configuration; adjust to your own situationfund_type: [equity-biased hybrid, equity]min_years_since_inception: 5 # years since fund inceptionmin_manager_tenure: 3 # years the current manager has run this fundaum_range_yi: [2, 100] # assets in units of 100 million (CNY)ranking_percentile_max: 30 # max peer percentile for 5-year returnmax_drawdown_limit: 0.35 # max drawdown capannual_win_rate_min: 0.6 # min annual win rate over 5 years (vs. benchmark or peer median)fee_total_max: 0.015 # management fee + custody fee capsharpe_min: 0.8 # minimum Sharpe ratio
Let the AI Produce Reproducible Calculation Code
Valuation-style metrics are a good fit for generated code: run it, and every number can be traced back to its computation and independently rechecked.
TRADING_DAYS = 244 # illustrative value; adjust to the actual trading calendar of the marketdef annualized_return(nav):"""Annualized return over the period, computed from a unit NAV series."""years = len(nav) / TRADING_DAYSreturn (nav[-1] / nav[0]) ** (1 / years) - 1def max_drawdown(nav):"""Maximum drawdown over the period; returned as a negative number."""peak, worst = nav[0], 0.0for value in nav:peak = max(peak, value) # update the running peakworst = min(worst, value / peak - 1)return worstdef sharpe_ratio(daily_returns, risk_free_rate=0.02):"""Sharpe ratio: annualized excess return / annualized volatility."""excess_mean = sum(daily_returns) / len(daily_returns) - risk_free_rate / TRADING_DAYSvariance = sum((r - sum(daily_returns) / len(daily_returns)) ** 2for r in daily_returns) / (len(daily_returns) - 1)volatility = variance ** 0.5 * TRADING_DAYS ** 0.5return excess_mean / volatility * TRADING_DAYS ** 0.5
Three things to watch when running code like this:
- Dividend adjustment: a distribution makes the unit NAV drop; convert to adjusted NAV before computing returns, or the return will be understated;
- Sample length: the Sharpe ratio is sensitive to sample length, so use the same measurement window across funds;
- Consistent frequency: the risk-free rate convention must match the frequency of the returns.
The screening table should end up looking like this (all figures below are illustrative):
| Fund | Type | 5Y annualized | Max drawdown | Sharpe | Manager tenure | Assets (CNY bn) | Fee (annual) |
|---|---|---|---|---|---|---|---|
| Fund A | Equity-biased hybrid | 12.4% | -28.6% | 0.82 | 6.5 years | 4.52 | 1.50% |
| Fund B | Equity-biased hybrid | 10.8% | -22.1% | 0.91 | 4.2 years | 3.17 | 1.40% |
| Fund C | Equity | 15.1% | -34.8% | 0.79 | 3.8 years | 1.89 | 1.50% |
On return alone, C leads. On drawdown and Sharpe, B is steadier. A is fairly balanced across the board. Which one actually suits you depends on your own risk tolerance.
Step Out of the Recency Trap
AI inherits common cognitive biases too — for example, favoring whatever has risen most recently. Address it in the prompt: ask it to “set aside the funds ranked in the top 5% over the past year and analyze them separately,” and to answer “if the best performers of the past year underperform the following year, what are the likely reasons?” Explanations such as mean reversion, style rotation, or inflows diluting strategy capacity help build a calmer judgment.
6. Deep Analysis: Unpack the Return, Detect Style Drift
Once candidates are shortlisted, the question becomes “where did this fund’s return come from?” The analysis step can cover three layers:
- Performance attribution: using the sector allocations and top holdings disclosed in periodic reports, determine whether excess return came mainly from sector allocation (which segments were overweighted), security selection (stock-picking within a segment), or position timing. Excess return that relies persistently on a single sector is usually less durable than excess return from diversified sources.
- Style identification: from the market-cap and sector distribution of top holdings, judge whether the fund is large cap or mid/small cap, growth or value. A common technique is to plot the fund’s style exposure relative to its benchmark as a time series. For example, if the contract specifies “equity assets between 60% and 95% of fund assets” but the actual equity position sits near the ceiling year after year, the style is aggressive. Sector concentration drifting persistently outside the contract’s stated investment scope can also constitute style drift and deserves attention.
- Manager and team: mark the current manager’s start date on the NAV curve and measure performance before and after it separately. High returns achieved with less than three years in the role carry limited evidentiary weight. Also look at how many products the manager runs and of what type, and how often manager-change announcements appear.
One rule to enforce during analysis:
Do not generate content about a manager's personal history, internal decision-makingprocesses, or non-public information by inference. Such content may only quote theoriginal wording of public materials such as periodic reports and announcements,with the source cited. If it cannot be found, write "not disclosed in public sources".
7. Portfolio Evaluation: Good Funds Do Not Guarantee a Good Portfolio
No matter how good a single fund is, several near-identical funds held together will not reduce risk. The portfolio evaluation step should check the following:
- Correlation: compute the correlation matrix of the candidates’ daily returns over the past three years. High correlation usually means similar style exposure and limited diversification benefit;
- Concentration: the combined weight of the portfolio’s top three holdings as a share of NAV, and the largest single-sector weight;
- Style complementarity: for instance, a balanced-style fund as the core holding and a sector-theme fund as a satellite holding, with an explicit cap on the satellite;
- Drawdown scenarios: use historical market data to approximate the portfolio’s drawdown if a single sector sells off or the whole market corrects — in effect, a stress test;
- Rebalancing rules: set a deviation threshold or a fixed interval (such as every six months) and return to target weights when the deviation is breached.
Illustrative output:
| Portfolio | Composition | Avg. 3Y correlation | Top-3 holdings weight | Max drawdown (illustrative) |
|---|---|---|---|---|
| Portfolio 1 | A + C | 0.76 | 58% | around -30% |
| Portfolio 2 | B + bond fund D | 0.32 | 41% | around -14% |
Both funds in Portfolio 1 are equity products with similar price action; adding more equity exposure amplifies a single risk. Adding a bond allocation in Portfolio 2 lowers correlation and volatility, which better matches an R3 moderate investor.
8. Verification and Backtesting: Do Not Let AI Conclusions “Look Right”
AI output is a hypothesis to be verified, not a conclusion. Five steps:
- Sample re-check: pick 3–5 data points at random and verify each against the original source — periodic reports, NAV announcements, and so on;
- Recompute the code: recalculate key metrics with the generated code. If results differ from the AI’s figures, the recomputed values win; then look for data-source or dividend-adjustment problems;
- Adversarial test: have the AI play a “devil’s advocate analyst” and list three reasons the selected funds might disappoint — an unfavorable market style, elevated sector valuations, rapid asset growth, and so on;
- Backtest: run the strategy over historical data, watching specifically for look-ahead bias (using information in the backtest that was not available at the time) and survivorship bias (a sample containing only funds still operating today, ignoring liquidated products);
- Periodic review: rerun the same prompt and configuration each quarter, compare old and new results, and fix conditions that have stopped working.
9. Risk Boundaries and Common Mistakes
| Common mistake | Risk | What to do |
|---|---|---|
| Treating AI as a stock-tipping tool | Ignoring your own risk tolerance; chasing rallies and selling dips | Use AI only for research support; make decisions in light of your finances and risk appetite |
| Looking only at return rankings | Overlooking drawdown, volatility and style exposure | Use risk-adjusted metrics such as the Sharpe and Calmar ratios |
| Trusting historical performance | Past performance does not predict the future; a “former champion” effect exists | Focus on the durability of the return sources and the stability of the manager |
| Skipping data verification | The model may fabricate data or cite stale information | Insist on source labels, sample re-checks and code recomputation |
| Exposing sensitive information in the chat | Uploading ID numbers, bank card numbers, trading account details creates privacy risk | Upload only public material relevant to the research; redact sensitive information yourself |
| Switching funds frequently | Higher transaction costs and the risk of missing the rebound | Set rebalancing rules and follow them with discipline |
And keep one point in mind throughout: AI can help you work faster, organize information and supply an analytical framework, but it cannot bear the consequences of your investment decisions, and it does not constitute investment advice.
Conclusion
The value of using AI for fund selection lies less in how powerful the model is and more in whether you have configured it into a dependable research assistant. Spend ten minutes setting up the expert, the skills and the connectors, then run one research cycle through measurable screening → return decomposition → portfolio evaluation → verification and backtesting. With consistent source labeling, sample re-checks and periodic reviews, AI can become a reliably working part of your investment research workflow.
