I turned over part of my portfolio to artificial intelligence, not as a chauffeur but as a co-pilot. The idea was straightforward: see if an AI-assisted method might bring out better stock ideas, and do so more quickly, without giving up basic investing hygiene like diversification and cost control.
To date, they are neither a moonshot nor a disaster. Performance has been market-like, with somewhat sharper swings, but the workflow gains are real. It excels in compressing research chores formerly requiring hours into minutes. It still needs adult supervision.
How I Created an AI-Assisted Investing Workflow
I built a three-step system.
- First, I use a large language model to write summaries of 10-Ks, earnings call transcripts, and quarterly filings — flagging revenue mix, margin trends, and any risks hidden in footnotes.
- Second, I filter for fundamentals and momentum using off-the-shelf methodologies, starting with ratings on companies based on free cash flow yield, return on invested capital (ROIC), and 6–12 month relative strength vs. sector.
- Third, I use my risk rules for how much I allocate to a position and how I exit.
The AI’s job is synthesis. It writes a rapid thesis on each candidate, then I check every figure against primary data sources, such as SEC EDGAR and the database people at Capital IQ. If the model refers to a metric, it must be in the filing. This in itself has reduced my research time by easily half.
What the AI Really Picks for Stocks and ETFs
There’s cash generation and pricing power, so the machine will tug you toward large-cap quality — profitable software, semiconductor leaders, insurers — and broad ETFs when single-stock risk looks elevated.
It likes companies with improving gross margins and positive analyst revisions, and it also spotlights those spending heavily on buybacks without matching free cash flow (that’s a handy red flag).
The model also recommends ETFs to play diversifying themes — broad market exposure, energy, and investment-grade bonds — when it sees single-name volatility growing. That’s nudging me away from overconcentration and, in fact, is the exact type of discipline that people lose during hot runs.
Early Results and Their Meaning for My Portfolio
In the first few weeks, the AI-enhanced basket was also bouncier than my core index holdings. Daily swings grouped around ±1.5% for the basket versus roughly ±1% for a broad market ETF. A few trades played out fast — good tech on strong guidance, for example — though a cyclical end of things turned around as management moderated full-year margin targets.
Net-net, performance has been close to my benchmark with a slight positive edge that could disappear overnight if we had one big earnings miss. The bigger win so far is process: fewer “FOMO” trades, faster thesis checks, and better pre-trade rules. Hit rate isn’t as important as the average gain versus average loss; keeping losers small is where the AI’s alerts are most useful.
Risks, Guardrails, and Red Flags When Using AI
AI can hallucinate, particularly on esoteric accounting items. I need sourcing, and I refuse to countenance any claim that’s unverifiable. I also limit position size and spread my bets across different sectors while largely keeping the rest of it in low-cost index funds. Taxes and trading expenses still count; there is no direct relationship between more “smart” trades and better outcomes.
There’s also a systemic risk. Regulators have cautioned that pervasive use of similar AI models can cause herding. The SEC has cited conflicts of interest and the risk of “predictive data analytics” directing investors to inappropriate products. Translation: think of the model as an analyst, not an oracle.
The Data and the Experts on AI and Stock Selection
Evidence is mixed but intriguing. In 2023 tests, University of Florida researchers found that autoregressive word-based language models classifying news sentiment have some predictive power for short-term stock moves, but the edge fades as other investors adopt similar signals. Other research indicates that LLMs are perfectly capable of reading financial statements well, but can stumble over complex accounting nuance without guardrails.
Zooming out, SPIVA scorecards from S&P Dow Jones Indices consistently find the vast majority of active funds underperform their benchmarks over 10-year horizons. In research on thematic funds, Morningstar finds that many AI-labeled products trail broad indexes after fees despite big inflows. Quants at large banks still report that, once they account for quality, value, and momentum — old standbys — it helps explain a lot of the equity returns across cycles.
Or, in other words, AI could hone research and execution, but long-term drivers seem somewhat familiar. The costs, behavior, and risk management often claim more victims than clever stock picks.
My Playbook Going Forward for AI-Assisted Investing
I’m keeping the core portfolio in broad ETFs and using a small, clearly defined sleeve for AI-assisted picks. All ideas must meet the same tests: verifiable data, a recognizable catalyst, and an exit plan before entry. If you can’t describe how the model speaks to a thesis in plain English, it doesn’t belong in the portfolio.
Bottom line: AI is a great research assistant and an average risk cop, not a magical stock whisperer. The record has promised enough to continue — just not at the expense of first principles that have worked for decades.
Disclosure: This is about my experience, and how I research (not investment advice). Do your own due diligence, or consult a financial adviser or fiduciary, before making any investment decisions.
