Can AI give trading signals? What's really happening behind the "model's brain"

A language model, like Claude or ChatGPT, doesn't see the market, doesn't calculate indicators with mathematical precision higher than a calculator, and doesn't know what will happen tomorrow. It can explain, summarize, write code and formulate text. When an ad promises "the brain of the model combined with an oscillator," it's selling you the oscillator, which has existed since the 50s, with an AI company's logo on top of it. The signal itself, if it's even calculated, is a deterministic formula that every charting software generates for free.

On this page: what AI really contributes to trading, what it can't do, four signs that separate a service that can be checked from a signals group that sells hope, and how we use (and don't use) AI at TechAnalyzer Pro.

What a language model can do for you

• To explain: What a divergence is, why the stop is at that point, what a report says. Instead of jargon, simple language. This is an excellent use, and we do it without a model: the explanations in the report and in the Beginner's Mode are fixed, vetted text, not a model's guess.
• To write backtests: To describe rules to it and get code that tests them on data. This shortens weeks to hours. The testing itself still requires the protocoland the model will happily write a look-ahead bug too if you don't specifically ask otherwise.
• To summarize news and reports: Fast, but not always correct. A summary is a starting point for reading, not a substitute.
• To formulate a plan and challenge you on it: "Am I following my rules" is a question the model answers well, because it doesn't require knowing the future.

What it can't do, and what the ads hide

• Predict price. A language model has no access to information the market doesn't have, and the market has already priced in what's known. "Accuracy" is a marketing word, not a tested metric.
• "Combine" an indicator. A stochastic oscillator is a one-line formula. There's no brain to combine with it; there's only something to draw an ad on top of it. When we measured the stochastic signals on 184 stocks over three years, they earned +19% versus +95% the index, with no model at all, and with a model it was the same thing.
• Eliminate the costs. A WhatsApp signal arrives late, fills at a different price, and whoever sends it doesn't count the slippage. "Real-time" results without a fill price, stop and costs aren't results.
• Replace a license. In Israel, a personal recommendation for a trade is investment advice, and whoever gives it for payment without a license is breaking the law. An anonymous signals group is exactly that.

Four signs that separate a service from a group

• Written rules. A real service states exactly when to enter, when to exit and how much to risk. "The algorithm identifies divergences" is not a rule.
• A public track record, including the losses. A table of all the signals, not screenshots of the winners. If you can't see the failed trades, there's no track record.
• Fill price and costs. A result measured in R with a defined stop, not in percentages from a point chosen afterward.
• Who's behind it. Name, company, what the service sells and whether it's licensed. A sponsored ad with a profile with no history and a WhatsApp button answers none of these.

How TechAnalyzer Pro uses AI, and how it doesn't

The engine that generates our setups runs on deterministic rules: price structure, volume, 200 average, risk-reward ratio, closed candle. There is no language model deciding on a trade, and there never will be, because a model that decides is a model that can't be tested. The rules are written, detailed on their own pageand the track record of every signal is open on the site, including the losses. AI for us is a work tool for the developers: writing backtests, debugging issues, drafting explanations that are later checked and finalized. That's the difference between an "AI that alerts" and a system that anyone can take and verify for themselves.