Two very different things are called "AI trading" in the United States right now. One is a research operation with a colocated data center, a team of statistical-learning PhDs and a model that reclassifies the state of the market faster than you can blink. The other is a moving-average crossover with a landing page and an ad budget.
They share two letters and nothing else. Telling them apart is worth more than any prediction either one makes.
When somebody tells you a system "uses AI," they mean one of three things, and the difference matters far more than the word.
Rule-based automation. Fixed conditions written by a person: if this happens and that happens, buy, put the stop here and the target there. The computer executes. It does not decide anything that wasn't decided in advance. This is what almost every product sold as a "trading bot" actually is, including the one sold on this site. It is not artificial intelligence in any useful sense of the term, and calling it that would be a lie.
Machine learning. Here the rules aren't written, they're fitted. You hand a model a pile of historical inputs (prices, volumes, order book snapshots) plus something you want predicted, and an optimizer searches for the parameters that best map one onto the other. Nobody writes the logic, and nobody can fully read it back out afterward. That is the trade you're making.
Large language models. Text in, text out. Genuinely useful for getting through a 200-page prop firm rulebook or a Fed statement in a hurry. They have no price feed, no position, no idea what your account balance is, and no memory of what they told you yesterday.
AI is a category of tool, not a category of edge. The useful question is never "does it use AI." It's "what does this system do, on what data, and can I reproduce the result myself." A system that can't answer those three isn't sophisticated. It's unverifiable, which is a different thing entirely.
Serious quantitative firms use machine learning heavily, in places that would bore a sales page to death.
Execution is the biggest and least glamorous win. A fund buying 30,000 contracts cannot send one market order. It would run the price away from itself and pay for the privilege. The parent order gets sliced into children spread across venues and minutes, and how you slice it is a learnable problem with a measurable score. Notice what those models optimize: not whether to buy, but how to buy something a human already decided to buy.
Feature extraction comes next. A raw order book is a firehose, and compressing it into a few numbers that still carry the information (imbalance, pressure, the shape of resting liquidity) is exactly what neural networks are good at. The output isn't a trade, it's an input to something else. Regime classification works the same way: clustering and hidden Markov models label what kind of market today is, which tells a desk which of its existing models to trust this week and which to size down.
Text parsing over filings, central bank statements and headline feeds is real, it works, and it is a latency race measured in single-digit milliseconds. By the time a headline is legible on your screen, that trade has been over for a while. And anomaly detection, catching stale quotes and broken feeds and strategies that have quietly started behaving unlike themselves, saves more money than most alpha while never appearing on a poster.
Notice the pattern. Almost none of it is "the machine decides what to buy." Most of it is plumbing, measurement and cost control, at a scale where one basis point is somebody's salary.
| Area | What AI genuinely does | How it gets sold to retail |
|---|---|---|
| Execution | Slices a large order across time and venues to reduce market impact | "AI finds the perfect entry price" |
| Feature extraction | Compresses raw order book data into a few usable numbers | "Our neural network reads the market's mind" |
| Regime classification | Labels today's market state so the desk knows which models to trust | "It knows when a crash is coming" |
| Text parsing | Reads filings and statements in milliseconds, faster than any human | "It reads the news for you and trades the headlines" |
| Anomaly detection | Flags broken feeds, stale prices and orders that shouldn't exist | Almost never mentioned, because it isn't exciting |
| Pattern search | Finds structure in datasets far too large to eyeball | "It learns and gets better every single day" |
Every entry in the right-hand column describes an outcome, never a mechanism. That is the tell, and it's available to you before you spend a dollar.
Ask yourself why a genuinely predictive market model would ever be packaged as a consumer download.
Capacity kills it. Most statistical edges are small and hold only up to a certain amount of money. Ten thousand people running the same model on the same instrument at the same second do not each get the edge. They get worse fills and one crowded trade. An edge with limited capacity is destroyed by distribution, which is why nobody distributes one.
The cost structure doesn't fit, and decay is continuous. Data licenses, colocation, a research team and constant retraining are an operating budget, not a one-time purchase. A model that needs five people to stay alive does not become a $500 download. And these models rot: relationships shift, participants adapt, and a model trained on one regime quietly stops working in the next. A retail product pitched as "set it and forget it" describes the opposite of how they live.
The fair question, aimed at us: isn't that an argument against buying any system, including the two sold here? Partly, and it should be said out loud. The honest distinction is that a rule-based system trading one window a day on the most liquid index futures in the world, in micro contracts, is not a capacity-constrained arbitrage edge that dies when more people run it. But the defense isn't that story either. It's a backtest you can rerun and check line by line. Apply the checklist in how to choose a trading bot to us with the suspicion you'd use on anyone else.
Open the hood on a $199 "AI trading bot" and underneath is almost always one of three things: an indicator crossover, an RSI threshold, or a martingale grid that doubles down until it recovers or takes the account with it. None of those needs machine learning, and none is improved by two extra letters on the sales page.
The tell is what's missing. A description of a real model answers four things without being asked: what the inputs are, what it is predicting, what period it was trained on versus what was held out and never touched, and how often the copy you receive retrains. Nobody selling a black box at consumer prices answers those. Ask anyway. The reaction to the question is the answer.
If the explanation of how it works is a metaphor, there is nothing behind the metaphor.
What makes machine learning dangerous in retail hands has nothing to do with the models being too smart. A flexible model with enough parameters can explain any past perfectly, including a past made entirely of noise.
The arithmetic is unforgiving. Test a thousand rule variants against history, keep the ones that clear a 5% significance bar, and roughly fifty clear it by pure chance. Publish the best of those fifty and you have a beautiful chart describing something that never existed. Nobody had to lie. The search itself manufactured the result.
Which is why an equity curve that runs as a smooth diagonal is a warning, not a selling point. Real edges are lumpy: flat stretches, ugly months, drawdowns that make you question the whole thing. The published Algomyth engine backtest covers 1,164 trades with the target at three times the stop, which puts break-even at a 25% win rate. It survives on the size of the winners, not their frequency. A curve built out of that is not smooth, and if it were, you should assume somebody polished it.
A backtest is a claim about the future written in the past tense. The more knobs were turned to produce it, the less of a claim it is. How those numbers hang together is worked out in risk-reward and win rate.
Every system has a bad stretch. That's not a possibility to plan for, it's a certainty on a schedule you don't get to see. The published Algomyth engine backtest has a worst drawdown of $5,273 on the account it models, and that valley happened in a simulation where the ending was already known. Sitting through the same thing with real money is a different experience.
Inside a drawdown, exactly one question matters: is this normal, or is this broken?
With a rule-based system you can answer it. Check whether the setup conditions still occur at the usual frequency, whether the losses are the usual size, whether entries are landing where they historically landed. You compare current behavior against documented behavior and you get a verdict.
With a black box there is no test to run. You can see that it's losing. You cannot see whether the model is doing what it always did in a market that turned against it, or whether it stopped working three weeks ago. Both look identical from outside, so you're choosing between faith and a coin flip with money on the outcome. Most people choose wrong, because most people switch systems off at the bottom. More on that failure mode on the risk page.
None of this makes the tools useless to a retail trader. It puts their use somewhere other than the trade decision.
What it should not do: decide whether to enter, choose your size, or talk you out of a stop. A language model is an excellent research assistant and a terrible portfolio manager for exactly the same reason: it is fluent, and fluency is indistinguishable from confidence right up until the money is gone.
The full report, the maximum drawdown, the losing stretches and the live signal log that archives every new signal with its outcome, on one page. No model to trust. Just numbers to check.
Worth stating plainly, because claiming otherwise would sell better. The engine contains no a neural network, learning in production or retraining. Take the Algomyth engine: a fixed set of conditions, evaluated in a single window once a day at 9:30 AM ET, when the US cash market opens and the day's volume arrives all at once. When the conditions line up it enters, and in the same instant it places a stop and a target in the market, the target set at 3× the risk. It closes on one or the other, and it never holds a position overnight. The mechanism is written out in how it works.
That isn't a limitation to apologize for. It's the feature. Deterministic means the same data plus the same rules produce the same trades, which is the only reason a backtest is worth anything. You can load the strategy into the Strategy Analyzer in NinjaTrader 8, set the same period, press Run and see whether the numbers you were shown are the numbers that come out. A model with a random seed and a rolling retrain schedule cannot offer you that, however good it is.
The published simulation covers January 2022 to September 2026, day by day, on a mid-size account: $87,359 net across 1,164 trades, a profit factor of 1.46 and a worst drawdown of $5,273. That works out to an average of $18,814 a year. There is no gross figure alongside it, on purpose: commissions of $1.90 a contract round turn and a tick of slippage come out inside the simulation, trade by trade, so what you read is what would have been left.
The ugly part belongs in the same breath as the good one. A drawdown deep enough to trip a funded account's limit ends that account, and the evaluation fee goes with it; the deepest one in this record was $5,273. Nothing about a simulated result makes that fee less real. That arithmetic is laid out in funded capital.
Hypothetical performance. Every figure above comes from a backtest over historical data, not from a live account. Simulated results are prepared with the benefit of hindsight, carry no financial risk, and cannot fully account for real execution, slippage or liquidity. Past performance, real or simulated, does not guarantee future results.
Run the same five at everything else you're weighing, free options included. There's a survey of those in free trading bots, and the broader fraud patterns in how to spot a trading bot scam.
Neither is better as a category, because neither one is a strategy. A machine-learning model can find structure a human would never spot, and it can fit pure noise with the same confidence. A rule-based system is only as good as the person who wrote the rules, but you can read it, test it and tell when it breaks. For a retail trader with no research team, the auditable option beats the theoretically stronger one.
Yes, extensively, and mostly in places that never reach a marketing page. Execution algorithms, feature engineering, regime detection, text parsing and anomaly detection are all standard equipment. What is far less common is one model deciding on its own what to buy and how much, with no human structure around it. The interesting work is the plumbing, not the prophecy.
Not usefully. It has no market data feed, no broker connection, no knowledge of your positions and no consistent behavior between two identical questions. Ask it the same thing twice and you can get two different answers, which is fatal for a trading rule. Use it to read documents, write code and check arithmetic, and keep it away from the order ticket.
Ask what the inputs are, what the target variable is, what period it was trained on and what was held out for validation. A real model has specific answers to all four and the seller gives them without being pushed. A repackaged crossover produces vague language about proprietary algorithms and neural technology. The vagueness is itself the finding.
Possibly at the institutional level, where the data, infrastructure and staffing exist to maintain models continuously. At the retail level the binding constraint isn't intelligence, it's verifiability: you have to be able to check what you bought and tell a normal losing stretch from a broken one. Until a retail-priced model can be reproduced and audited by the person paying for it, a documented rule set is the more defensible purchase.
Everything we write comes back to one idea: verify, then decide. The live terminal exists for exactly that.
Unlock Algomyth Live