Can an open-ended AI do well on a real portfolio? I asked my human adviser to replicate some trades I made in my holdings and assigned Gemini the same tasks, tracking the results of both. The AI-built sleeve is up slightly on returns after a few months — not quite 1 percentage point up — and is forcing clearer logic for decisions as well.
How I Put Gemini to Work on My Investment Portfolio
The experiment began with grounded goals, not generic prompts. I outlined an aggressive, long-term blueprint that would tolerate short-term volatility, tilt heavily toward equities, avoid overdoing exposure to any one area of the markets and seek to surpass a broad benchmark like the Nifty 50 over a decade. Without that context, any LLM has the tendency to offer empty platitudes.
I then fed Gemini structured data: the names of funds and their codes, the amount that was invested in each fund, and how much they are worth today, when investments started, expense ratios when provided and category tags. Raw text, properly labeled, proved more effective than spreadsheets. I combined this with fund house factsheets and portfolio disclosures to allow the model to figure out sector tilts, duplication risks and exit-load rules.
To broaden the lens, I requested deep scans for categories I own — large-cap, mid-cap, small-cap and thematic funds — so that Gemini could benchmark against the category average and historical drawdowns. Consider it a speedier alternative to the due diligence many investors plan to — but rarely get around to — doing.
How the A.I. Developed More Accurate Predictions
First, it attacked hidden overlap. By comparing underlying holdings and sector weights — the second-order denoising — Gemini flagged funds that were different but effectively hewing to the same factors. It suggested cutting back on duplicative exposure and increasing allocation to a smaller number of high-conviction funds that had continued to outdistance their category averages after fees.
Second, the explanations were specific. Instead of “stay invested,” I received rationale: why a small-cap fund might underperform in a risk-off quarter, what catalysts could reverse relative performance, how long historical recoveries lasted. That degree of transparency mirrored what professionals’ desks do with factor heat maps and attribution reports, filtered down for an individual investor.
Third, the model remained disciplined on core exposure. Even as it maintained a solid position in large-cap index exposure, it tilted into mid-caps and several thematics, including renewable energy and power infrastructure, where policy momentum and earnings revisions suggested an edge over a 5–7-year window. It eschewed “all-in small-cap” heroics and managed the risk budget.
Does the Performance Edge Still Hold Today?
Backtests are a cinch; live performance isn’t. Gemini can’t stream real-time NAVs or provide real backtesting like a quant platform, so I also used a clean carve-out of capital and tracked XIRR. The AI-driven sleeve is beating my adviser’s picks by just shy of 1% over the past few months. That’s just noise in the short run, but compounding magnifies small edges: An 8 percent annualized return versus a 7 percent return over 20 years turns $100,000 into around $466,000 versus $387,000 — an edge of roughly $79,000.
It also lines up with more general data: S&P Dow Jones Indices’ SPIVA scorecards tend to show a majority of active fund managers trailing their benchmarks over long stretches, which sets the hurdle higher for any adviser or AI. The destination is not alchemy stock picking but tighter process, lower overlap and fees that drag less.
Where Gemini Shines and Where It Fails Today
It’s the speed and synthesis that stick out. Gemini distills what would be hours of reading into minutes, exposes conflicts like style drift and creeping concentration, yet yields plain-English memos younger analysts can audit. It’s a robust second opinion when triangulated with literature from the fund and third-party research from places like Morningstar or Value Research.
Limitations are real. If your inputs are narrow, the model may hallucinate specifics anyway; it won’t have all the recent changes in mandate as context, and it lacks live data unless you feed it some. Regulators have also warned the industry about “AI washing” and bias in predictive analytics; the investor advisories from the S.E.C. and global watchdogs are a timely reminder to confirm, not take on faith.
Historically, what works best for me is doing the unglamorous work: clean data sets, clear benchmarks and rebalancing rules every once in a while, at least. The AI sped up my work; it did not substitute for judgment or responsibility.
Practical Takeaways for Everyday Investors
Define what “better” is — target volatility, some benchmark like the Nifty 50 or S&P 500, a time horizon and limits on sector concentration. Feed the model holdings and enhance with factsheets, expense ratios, and categories. Use it to find overlap, quantify costs and pressure-test the thesis for each fund.
Then validate. Step 2: Compare recommendations to category medians, and monitor persistence studies from good players, track XIRR in a standalone sleeve before revamping your core. A good human adviser is still going to give you context on tax, liquidity and behavior — areas where A.I. is not a fiduciary and can’t take responsibility from you.
Bottom line: In my own experience, Gemini’s structured analysis and reallocation nudges have helped me beat my adviser’s recommendations by a hair (so far).
Even if that edge diminishes, the clarity and time saved is worth it — and when that one extra point compounds, it can be a quiet difference-maker in long-term wealth.
