AI Stock Pickers: Can AI Pick Winning Stocks, and How to Test One
An AI stock picker ranks or scores stocks using data such as prices, fundamentals and news, then suggests what to buy or avoid. It can screen faster than a person, but no AI can predict markets. Treat picks as ideas to check, not orders.
What an AI stock picker does
An AI stock picker ranks or scores names. The inputs are the ones you would expect: prices, fundamentals, and news. The output is a list, a score, or a sentence that says buy, avoid, or hold. The speed is real. A person reading ten filings will lose a race to a model reading a hundred summaries. The race is not the question. The question is whether the list beats a simple index after costs, on names you can actually buy.
Treat the list as ideas to check, not as orders. No AI can predict markets. A high score is a sort key. It is not a claim about next month. The CFTC press release of January 25, 2024, Release 8854-24, says AI technology cannot predict the future. A picker that skips that sentence and prints a target price has changed jobs, from screening to fortune telling. Use it for the first job.
On this site the stock tools are AI stock trading and AI stock trading bot. Strategy pages that group names include tech stocks and broad market. Those pages describe baskets. They are not a promise that a model rank will beat the basket.
What the research says
The abstract of arXiv 2304.07619 says GPT-4 headline scores capture the initial market reaction, about 90% portfolio-day hit rates for the non-tradable initial reaction, and some later drift, especially for small stocks and negative news, and that strategy returns decline as LLM adoption rises. Read the limits that are already inside that sentence. The high hit rate is for the initial reaction, and the abstract calls that reaction non-tradable. Later drift is "some," and it is stronger in places that are harder to trade cleanly. Returns decline as more people use the same kind of tool.
The practical number this page computes is the cost of acting on ten picks instead of one index token.
A short-lived reaction to a headline is a different claim from "this picker beats the market for you." The first can be true in a study and still be gone by the time your order exists. The second is the ad. Does AI trading work stays on that gap. Can AI predict the stock market stays on how large a move must be before a hit rate pays the fee.
Worked example: ten picks versus one index token
Assumed numbers for the math, not a quote and not a price call. Example, not a recommendation. You have $1000. Path one buys ten tokenized stocks at $100 each. Path two buys $1000 of SPYx once. Both paths later sell everything. The question is only the fee gap, not which names were clever.
Each $100 buy is under the size where 0.75% exceeds the minimum, so each buy costs $0.90. Ten buys cost $9.00. Ten sells cost about $9.00. Together that is $18.00, or 1.8% of the $1000. The single SPYx buy costs $7.50, which is 0.75% of $1000. The sell costs about $7.50. Together that is $15.00, or 1.5%. The basket must beat the index by about 0.3 points just to match costs. That 0.3 is $3 on $1000, the gap between $18.00 and $15.00. It is not an edge. It is the cover charge for having ten tickets instead of one.
| Path | Buys | Sells | Round trip | Share of $1000 |
|---|---|---|---|---|
| Ten $100 tokenized-stock buys | $9.00 | about $9.00 | $18.00 | 1.8% |
| One $1000 SPYx | $7.50 | about $7.50 | $15.00 | 1.5% |
Whether you can buy a tokenized stock depends on where you live and on the issuer's rules.
If your picker is right by less than that gap, the index path wins on costs even when the picker "wins" on the story. Use the trading fee calculator when the tickets are not $100. A larger ticket can leave the $0.90 minimum and pay 0.75%, which shrinks the gap. A smaller ticket widens it. Swaps and agent market trades pay 0.75% with a $0.90 minimum per swap. Limit, take-profit, stop-loss and trailing orders pay a flat 0.75% with no minimum. Onchain fees apply too. A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees. Fees on your agent's trades earn up to 10% cashback in USDC. It is not a discount on the swap fee. Tokenized stocks, and whether the ticker you want exists here, are explained on tokenized stocks. The example is arithmetic. It is not a list to buy.
How to test a picker without marrying it
Save the list. Do not trade it the same day. Write down the index you are comparing with, in this example SPYx, and the fee gap, in this example about 0.3 points. A week or a month later, compare the list with that index using prices you look up yourself. If you cannot do it without the model narrating the result, you are not testing. You are listening.
Change nothing about the test after you see the result. Adding a name that worked, or deleting one that did not, and then calling the edited list the original test, is how a picker always looks smart. Keep the first list. The edited list can be the next test. It cannot rewrite the last one.
You set your agent's per-trade and daily spending limits and market-cap range; Moby blocks red-rated tokens. Limits are only as tight as you set them. Sells and exits skip the market-cap and safety checks, and tokens with no safety rating are not blocked. Turn on Ask before every trade and your AI asks you first. Moby does not hold or approve trades. A picker that can send orders is an agent, not a screener. Leave Ask on until you have seen the fee gap with your own eyes. One agent per user, with its own wallet. Agents trade Solana tokens only. The agent cannot withdraw, transfer or send funds. It can still buy ten names you meant only to read. Reading is the default. Trading is a second switch.
Using scores without turning them into a basket by accident
AI Trading and copy trading currently need an access pass in the app. Supported hosts: Codex, Claude Code (CLI), Gemini CLI (with a paid Gemini API key or enterprise access) and Grok CLI. ChatGPT: works on Business, Enterprise and Edu workspaces that allow custom MCP servers (web). Moby does not run a schedule for your agent; the schedule lives in your AI client. Your resting orders (limit, take profit, stop loss, trailing stop, DCA) keep working on Moby even when your app is closed. A morning score can live in your AI client. Moby will not run that clock. If you place a limit for a name you already checked, that resting order keeps working when the app is closed. A fresh ranking does not place itself. Download Moby if you want the app. Then decide how many names are worth the extra tickets. Ten is not a magic diversified number if the extra cost is the thing the ten must overcome.
One index token is a boring plan. Boring is allowed. A picker earns its keep only when the checked list, after the 0.3 point gap and after your own delay, still looks better than the boring plan for a reason you can say without the score. "The model said so" is not that reason. A filing you opened, a holding you understand, or a decision to do nothing: those are reasons.
Why the delay is part of the test
The abstract's high hit rate sits on the initial reaction, and it says that reaction is non-tradable. Your order is later. Later is where the edge the abstract describes is already smaller, and where more users of similar tools shrink it further. A picker that shows you a headline after you have read it in three places is not early. It is a summary. Summaries are useful. They are not the 90% sentence.
So the honest use is a queue. Headlines in, your own check next, an order only if the name still makes sense at the price you can actually get, after the fee in the table. If the queue is empty most mornings, the tool is working. A queue that is always full is a machine for the $18 path. You do not have to walk that path to have used the picker. You used it when you declined nine of the ten.
Frequently asked questions
Can AI pick stocks that beat the market?
Not as a rule you can rely on. A model can rank names faster than you can read, but no AI can predict markets. Treat a rank as an idea to check against a simple index. The fee gap in the example can erase a small win before skill shows up.
What does an AI stock picker actually do?
It scores stocks from prices, fundamentals, and news, then suggests what to buy or avoid. The suggestion is a sort, not an order, unless you arm a bot to trade it. Check every figure yourself. A confident paragraph can still name a token you cannot buy.
What does the headline-score study say?
The abstract says GPT-4 headline scores capture the initial market reaction, with about 90% portfolio-day hit rates for the non-tradable initial reaction, some later drift, and returns that decline as LLM adoption rises. The high hit rate is on the part the authors call non-tradable, so it is not a forecast you can trade.
Is an AI stock picker worth paying for?
Only if its checked picks beat a simple index after fees, on names you can buy. In the example, ten $100 buys and sells cost about $18.00 against $15.00 for one $1000 index token. Save the list, compare it later with prices you look up yourself, and drop a tool that never clears that gap.
What do ten picks cost versus one index token?
On $1000, ten $100 buys and ten sells are about $18.00, or 1.8%. One SPYx buy and sell is about $15.00, or 1.5%. The basket must beat the index by about 0.3 points just to match costs. That gap is fees, not skill. Onchain fees are extra.
Can I buy these tokenized stocks where I live?
Whether you can buy a tokenized stock depends on where you live and on the issuer's rules. They are tokens on Solana, not brokerage shares. The example's tickers are for the arithmetic. They are not a shopping list, and a model score does not change the residence rules.
How should I test an AI stock picker?
Save the first list, name the index you are comparing with, and include the fee gap. Look up prices later yourself. Do not edit the list and call it the original test. Leave Ask on if anything can send an order. Declining a name is a valid result.