Home › Guides › AI & trading

How do you check an AI stock picker's track record?

Ask for every pick, with the date it was made, after fees, from the day it went live. Most "AI picker" records fall apart on one of those five checks. Here's how to run each one.

Beginner guide · 8-minute read · Updated 2026-10-28

The short answer

A real track record is every pick, dated before it moved, after costs, since the day the tool went live. If any of those is missing, you're looking at a sales page, not a record.

That's the whole test. The rest of this page shows how to run it before you pay for anything.

Why the ad always looks good

An ad has one job: get you to sign up. So it shows the best part of the story.

That isn't always a lie. The winners shown may be real. The trouble is what's left out: the losers, the fees, and the months it went badly. A record with the bad parts cut out can make a losing tool look brilliant.

So don't argue with the winners. Go looking for the missing parts.

Check 1: Is it the full record, or the best picks?

Ask for every pick the tool has ever made. Not "recent highlights". Not "top calls". Every one, winners and losers, in one list.

Then count them. A handful of picks tells you almost nothing. Five or ten trades can win on luck alone. My page on Sample size explains why small numbers fool us.

Look at the Win rate too, but don't stop there. A tool can win most of its picks and still lose money, if the losers are bigger than the winners. What matters is the Expectancy: the average result per pick, losers included. You can work it out from the full list with our expectancy calculator.

Check 2: Were the picks dated before they moved?

A pick only counts if it was public before the price moved. Anyone can "call" yesterday's winner today.

Ask:

• Where was each pick posted, and when?
• Can the post be edited or deleted later?
• Is there a record you can check yourself, not just a screenshot?

Screenshots are easy to fake and easy to choose. A dated public post, an email you received, or a log the seller can't edit is far stronger.

Check 3: What's left after fees?

The record should show results after everything you'd pay:

• the subscription itself;
• trading costs and the gap between the buy and sell price;
• taxes on short-term gains, which can be higher than on long-term ones.

A tool that's slightly ahead before costs can be well behind after them. Ask the seller for the after-fee number. If they don't have one, work it out.

Check 4: What happened to the models that didn't make it?

This one is sneaky. Say a company starts three "AI models". Two do badly and are quietly shut down. The third has a good year. Now the ad says: "Our AI beat the market."

Every word is true. But you're only seeing the one that survived. That's called Survivorship bias.

Ask: "How many models, funds or strategies have you launched, and how many did you close?" Then ask to see the closed ones too.

Check 5: Is it a backtest, or real picks since launch?

Backtesting means running a rule on old prices to see how it would have done. It's useful for testing an idea. It's not a track record.

Backtests are easy to tune until they look perfect. The builder can tweak the rules until they fit the past. That's called Overfitting, and it falls apart on new data. My guide on why past backtests don't guarantee results goes deeper.

So find the launch date. Only picks made after that date count. If the chart starts years before the product existed, the early part is a backtest.

The five checks on one page

What following the ad can look like (hypothetical)

Here's the story we see over and over. Everything here is made up, an illustration, not a trade idea.

A service, call it Service X, runs an ad: "Our AI's last 5 picks were all winners." You'd have made +$1,100 following them with $500 each. You sign up and follow every pick for a year, $500 each time.

At the end of the year you ask for the full list. Here it is:

Look at what each check would have caught:

• Full record: the ad showed 5 of 40 picks. The other 35 lost $1,900 between them.
• Win rate: 15 of 40 won. The losers were bigger and more common.
• Fees: the subscription added another $480 to the loss.
• Dates and survivors: you never found out if Service X had other models it closed. You didn't ask.

The ad wasn't false. Those 5 winners were real. It just wasn't the record.

What about AI tools that don't pick at all?

Many AI tools don't claim to pick winners. They explain, sum up the news or check your math. Those are easier to judge: you can test them on things you already know.

For option sellers, that's the healthier pattern. In the wheel strategy, the decision rests on rules you can check yourself: would you own the stock at that strike, and is earnings inside the trade? Our AI for options trading guide shows where AI helps and where it shouldn't decide.

If someone in a group chat is selling the picks, read my page on AI signal groups on Telegram and WhatsApp first.

Questions