The AI signals economy is a scam, and the data shows it

I have been collecting data on consumer AI trading signals services for about twelve months, partly out of professional curiosity and partly because every cohort that joins the Tradoki desk brings me at least one of them. The data is consis

I have been collecting data on consumer AI trading signals services for about twelve months, partly out of professional curiosity and partly because every cohort that joins the Tradoki desk brings me at least one of them. The data is consistent enough that I have stopped being polite about it. The AI signals economy is a scam in the technical sense — its marketed performance is not its real performance, its claimed methodology is not its real methodology, and the median customer ends a year on it worse off than if they had taken the subscription cost and put it in a savings account.

This is a contrarian piece, written in the first person, because the framing requires it. None of it is investment advice; some of it is a direct accusation against a category of products. The accusation is supported by the only data I have — the conversations, screenshots and trade exports that traders bring me — and acknowledged limitations are noted throughout.

What "AI signals" actually means in 2026

The retail AI signals market in early 2026 looks roughly like this. A vendor publishes a website with a backtest equity curve, a few testimonials, and a description of "proprietary AI" that combines transformer models with reinforcement learning and (in one case I have actually seen) "quantum-inspired optimisation." The vendor sells access to a signals feed for a monthly subscription, typically between $49 and $499. The signals feed publishes one to ten setups per day on a defined instrument list, with entry, stop and target.

The vast majority of these products have no AI in any meaningful sense. The signals are generated by rule-based systems that use indicators a junior trader could code in an afternoon. The "AI" in the marketing is, charitably, a thin classifier on top of the rules; uncharitably, it is a label.

Even when there is real machine learning underneath, the performance presented to the customer is almost never the live performance. It is some combination of cherry-picked historical periods, in-sample backtests, paper-trading approximations, and selection bias on which months get displayed.

How the marketed performance gets manufactured

Across the vendors I have looked at — and to be clear, the sample is dozens, not hundreds, so this is observational and not a market-wide audit — the same handful of techniques recur.

Cherry-picked equity curves. The published track record starts on a date that conveniently coincides with the start of a winning streak. The losing months from before that date are described as "the development phase" and excluded.

Survivorship bias on the vendor pool. Vendors that have a bad year quietly disappear. The vendors still selling have, by survival alone, a better-than-random recent track record. A new customer comparing the marketed performance of currently-selling vendors is choosing among a sample biased toward recent winners.

Retroactive trade reclassification. A losing trade gets re-labelled "educational example" or "demo trade" after the fact. The signals feed continues to show the same trade with no change in label; the equity curve published on the website silently excludes it.

Backtest assumptions that bear no resemblance to reality. No commissions. No slippage. Spread set to zero. Fills at the historical bar's mid-price. The backtest is not a reliable estimate of the live system.

Concierge customer service after a bad week. Not a performance hack, but a retention hack — a vendor who does outsized customer-service work after a losing week reduces churn that would otherwise communicate the truth.

Paid testimonials and screenshot fabrication. A category that I will not dwell on because everyone reading this knows the screenshots can be photoshopped. They are.

The cumulative effect is that the marketed performance is roughly indistinguishable from "the best 12 months of a strategy that has 50% win rate and standard volatility, displayed without context, with the bad months hidden."

Why customers stay

The interesting question is not why the products exist — they exist because they are profitable to operate — but why customers stay on them after the performance has visibly degraded.

A few mechanisms I see repeatedly:

The early hits. Most signals services, by chance alone, have a winning first month for any given new customer. Random variation makes some subset of customers join in a winning streak, and the early experience anchors expectations. The cumulative losing trades over the next eleven months get attributed to the customer's "execution" or to "abnormal market conditions," not to the system.

The "execution" excuse. When the customer's results diverge from the published results, the standard vendor response is that the customer entered late, sized incorrectly, or did not follow every signal. This shifts blame to the customer and preserves the vendor's marketing claim. It is also unfalsifiable — the vendor never publishes the precise execution required to match the equity curve, so the customer can never demonstrate they followed it correctly.

The community lock-in. The Discord, the Telegram, the WhatsApp group. Subscribers form a community around the service, and the community itself produces value (some of which is real — peer learning is real even when the underlying signals are not). Leaving the service means leaving the community, and the cost of that is psychological as well as financial.

The sunk cost. Twelve months of subscription is a meaningful amount of money. Concluding "I have been wasting it" is harder than concluding "next month is when it turns around." The vendor knows this and produces just enough hopeful content per month to keep the framing alive.

The cost feels small. $99 a month is the subscription price most likely to clear the "this is fine" threshold. It is also, over a year, more than most of the customers I have spoken to have made trading.

What real AI in trading looks like

The argument is not that AI cannot help in trading. It is that the consumer signals economy is not the place where AI is actually being used productively.

Real AI in trading workflows, as we use it inside the desk and as it is used inside firms with serious capital, looks nothing like a signals feed. It looks like:

• A model that summarises a 200-page central-bank report with citations to the source pages, so the analyst can read the relevant sections faster.
• A classifier that tags news headlines for sentiment against a labelled corpus, used as one of many inputs to a discretionary trader.
• A code-generation assistant that drafts Pine Script or Python that a human reviews against a static checklist (see writing Pine Script with Claude and GPT).
• A research assistant that walks a learning trader through unfamiliar concepts (see LLMs as research assistants, not traders).
• An execution-quality analytics layer that flags slippage outliers and venue performance issues.

What unifies these is the absence of an autonomous order-routing decision. The model accelerates the work between decisions; the human takes the decisions. The signals economy reverses this — the model takes the decision and the human is asked to execute it on a platform where the slippage alone often eats the marketed edge.

For a deeper treatment of why autonomous AI in trading systematically fails, see why AI live-trading bots blow up.

What to do instead

I will be specific. If you have been on a signals service for six months and the live equity curve does not match the marketing, the right thing to do is cancel the subscription, take the trades you would have taken from the signals, run them through your own analysis using a structured framework like the top-down approach, and journal the outcomes against the post-mortem template.

The arithmetic is unforgiving. Education compounds. Subscriptions compound subscription fees. A trader who spends $1,200 a year on signals and breaks even is paying $1,200 a year for the privilege of executing someone else's analysis at retail platform fees. A trader who spends $1,200 a year on a structured education and ninety days of focused practice (see the deliberate-practice plan) ends the year with a transferable skill and a journal that explains how their account moved.

The signals economy sells the fastest path to looking like a trader. Education is the slowest path to actually being one. The slow path is the only one I have ever seen pay.

The signals service is profitable. It is just not profitable for you. The vendor's customer is the next subscriber, not you.

— Author's note

● FAQ

Are all AI signals services scams?

Almost all the consumer-marketed ones, in our observation, are. There is a small and uninteresting category of legitimate institutional research products that uses machine learning to support analyst workflow; that is a different industry from the consumer signals economy.

How do they fake the performance?

Selection bias on the published track record (only the winning months are shown), retroactive label changes on losing trades (calling them 'educational examples'), unrealistic backtest assumptions (no slippage, no commissions, no spread), and survival bias (failed services disappear, the surviving ones look like the average).

Why do people keep paying for them?

Because the marketing is sophisticated, the early-month signals sometimes work, and the cumulative cost of a $99-a-month subscription is small enough to feel low-stakes — until you realise you have been on it for a year and broken even at best.

What is the difference between a signals service and a research service?

A research service publishes analysis and lets you decide. A signals service publishes 'enter long here' and asks you to follow. The first sometimes adds value; the second almost never does at the consumer price point.

What should I do instead?

Learn to read the markets you trade well enough that no one's signal can outperform your own analysis. Education compounds; subscriptions compound subscription fees. The arithmetic is unflattering to the subscription side.