How to Spot a Fake AI Trading Track Record (Denominator, Survivorship, Win-Rate Tricks)*

Three tricks turn a losing algorithm into a screenshot that sells — here is how the court reads them.

The Charge

A track record is a claim. Most claims presented to this court are true in the narrow and false in the whole.

Nobody prints a fabricated number. They print a real number computed on a rigged set. The number holds. The set does not.

Three devices do most of the work: the denominator trick, survivorship, and win-rate by design. Below is how each is built, and how it fails on examination.

Trick One: The Denominator

A percentage has two parts. The audience reads the top. The fraud lives on the bottom.

"+312% return" says nothing until you ask: percent of what, over what, held for how long.

Common substitutions:

• Peak, not start. Return measured from the best entry in hindsight, not the date capital went in.
• Best window. "+312%" over the one quarter that worked. The trailing twelve months: -18%.
• Notional, not capital. Percent of position size, not of the account. A $1,000 account trading $50,000 notional reports leverage as skill.
• Closed trades only. Winners are realized and counted. Losers sit open, unrealized, off the ledger. The account is down; the track record is up.

Concrete case. Screenshot: "+40% this month." Account start $10,000. One trade returned $4,000 on a $2,000 position. Six other trades lost $3,100, still open. Reported: +40%. Actual equity: $10,900, or +9%, and falling. The +40% holds as a sentence about one trade. As a claim about the account, it does not hold.

Test. Demand: starting equity, ending equity, same dates, all trades open and closed. One number, one denominator. If the denominator moves when you ask, the record is not proven.

Trick Two: Survivorship

You are shown the account that lived. You are not shown the graveyard.

Run 1,000 accounts on random or near-random signals. After a year, by chance alone, some are up 200%. Kill the losers. Photograph the winner. Sell the winner's screenshot as method.

This needs no lie. The winning account's numbers are real. The population it came from is hidden.

Where it hides:

• Signal groups. Ten Telegram channels launch. Nine post losses and delete. One posts wins and scales. The survivor was selection, not edge.
• Strategy resets. "Track record since March." The January and February accounts blew up and were not counted. The clock starts at the first survivor.
• Cherry-picked bots. A marketplace shows the top 5 of 4,000 bots. The top 5 of any 4,000 random walks look like genius.

Concrete case. A vendor advertises a bot at +180%/year, "live since 2024." Examination: the vendor runs 60 bot variants. This one is the best. The median variant returned -12%. The advertised bot is the maximum of a distribution, presented as its mean. Holds as "one bot did this." Does not hold as "this method does this."

Test. Ask for the full population. How many accounts, bots, or signal sets started? What happened to all of them? If you only ever see winners, you are reading survivorship. Absence of a graveyard is not absence of death.

Trick Three: Win-Rate by Design

"92% win rate" is the easiest true number in trading to manufacture. It measures how often, not how much. The two are unrelated.

Mechanics: take profit early and small, let losses run large. You will win often and lose everything.

• Set take-profit at +1%. Set stop-loss at -20%, or none at all.
• 92 trades close at +1% = +92%.
• 8 trades run to -20% = -160%.
• Win rate: 92%. Net: -68%.

The win rate is honest. It is also irrelevant. A high win rate with an unmentioned average-loss is a martingale wearing a suit.

Concrete case. "9 out of 10 trades profitable." Average win: $50. Average loss: $600. Ten-trade cycle: +$450 from nine wins, -$600 from one loss, net -$150. The 90% holds. The account bleeds. The court weighs expectancy — average win times win rate minus average loss times loss rate — not the win rate alone.

Test. Demand average win size and average loss size next to the win rate. Compute expectancy yourself. A win rate quoted without loss size is quoted precisely because the loss size is the story.

What the Court Requires

A track record is not proven by a screenshot. It is proven by a record that survives its own denominator.

Minimum exhibits:

• Third-party read-only verification. Broker-linked, tamper-proof, not a cropped image. A screenshot is a drawing.
• Every trade, open and closed. No selection. The graveyard included.
• One denominator. Starting equity to ending equity, fixed dates, capital not notional.
• Expectancy, not win rate. Average win, average loss, largest drawdown.
• The population. How many started; what happened to all of them.

Supply these and the record can hold. Withhold any one and the verdict is fixed in advance.

The Verdict Rule

We do not say scam. We say the claim is not proven — and name the missing exhibit.

• No verification link: not proven.
• Winners shown, population hidden: survivorship — not proven.
• Win rate without loss size: incomplete — not proven.
• Denominator that moves under questioning: does not hold.

The honest operator has the exhibits and hands them over. The one who cannot is telling you which trick he used by which document he will not produce.

A true track record fears no examination. That is the whole test.