Level 3 AI in trading · Part 1/9

AI Trading: How It Works and Where It Fails

AI trading is an umbrella term. It covers research teams that train machine learning models on decades of market data, retail apps that place orders based on a handful of rules, and schemes that use the word only as bait. The label says little about what happens inside. This article sorts the approaches, walks through how such a system is built, and shows with our own price data how high the bar is before any of it pays.

What AI trading means

Four kinds of product carry the label. Statistical models learn a relationship between inputs and future returns from past data, for example with gradient boosting or neural networks. Rule-based bots follow fixed conditions such as moving average crossovers; the “intelligence” is often just a parameter search on historical prices. Language models read news, filings or social media and turn the text into a sentiment score that feeds a trading rule. Signal and copy services sell the output of any of these to customers who execute the trades themselves or let an app do it.

The distinction matters, because each kind fails differently. A statistical model can overfit; a rule-based bot can simply stop working when the market changes character; a sentiment score can react to headlines that professional traders have already priced in.

How AI trading works: building a system

Every serious system goes through the same stages, whatever the model inside:

Data. Prices, volumes, fundamentals, text or order book data. The quality of this step decides most of the outcome. Survivorship bias, look-ahead errors and timestamps that do not match the moment the information was really available all produce results that cannot be repeated.
Features and labels. The raw data is turned into inputs, and a target is defined: the return over the next day, a move up or down, or whether a trade hits a profit target before a stop. How the target is set is a design decision with large consequences, as the article on labeling shows.
Training and validation. The model learns on one period and is tested on another. Because market data overlaps in time, ordinary cross-validation leaks information from the test set into training; cross-validation for financial data explains the fix.
Backtest. The strategy is simulated with realistic costs. Testing many variants and reporting the best one inflates the result, which backtest overfitting quantifies.
Execution and risk control. An interface to a broker sends orders, and rules limit position size and losses. Risk and position sizing covers the basics.

The model itself is often the smallest part. Firms that do this professionally spend most of their effort on data and on the question whether a result is real.

Why it is harder than it looks

Markets are not a stable environment. A pattern that held for five years can vanish once enough participants trade it, and the data that would reveal this arrives only after the losses. Price series also carry little signal compared with their noise, so a model with millions of parameters finds structure in chance with ease. The broader list of failure modes is in why machine learning fails in markets .

Costs make the problem sharper for short horizons. Our study on why most traders lose money measured it on index futures: over one trading day, the market itself closed higher in 54.5 % of cases. At costs of 0.2 % per round trip, a strategy trading daily would need a hit rate of 64.1 % just to break even. Over a year of 252 round trips, the costs alone add up to 39.6 %. An AI system that trades every day has to clear that gap before it earns anything. Over one week the required hit rate falls below what simply staying invested delivered, which is why the horizon matters more than the method.

AI trading bots and apps

Retail apps promise to do this work automatically. Their marketing rarely shows the numbers that would allow an assessment. A few things can be checked without any technical knowledge:

Authorization. Is the provider registered with a financial supervisor in its home country, and does the name in the register match the company behind the app? In the EU, that means a national regulator such as BaFin; in the US, the SEC, the CFTC or FINRA.
Track record. Are past results audited by an independent party, or only shown as a chart on the provider’s website? A backtest is not a track record.
Costs. Subscription fees, spreads, commissions and financing costs for leveraged positions. For a small edge, costs decide whether anything is left.
Custody. Does the money stay in an account at a regulated broker in the customer’s name, or is it transferred to the provider?
Withdrawals. Can funds be withdrawn at any time, and are the conditions written down?

A promised return is not one of these points. It says nothing about whether an offer is legitimate.

Offers that use AI as bait

Regulators have warned repeatedly about schemes that use artificial intelligence as a sales argument. In January 2024 the US Commodity Futures Trading Commission published the customer advisory AI Won’t Turn Trading Bots into Money Machines . It describes offers that promise high automated returns, often through “AI arbitrage” between crypto exchanges, collect deposits and then stop paying out. Typical signs are fixed daily or monthly returns, pressure to recruit new customers, and payouts that only work as long as new money comes in. Real arbitrage gaps between exchanges are small and close within seconds; they do not produce steady double-digit monthly returns.

What remains realistic

AI methods are useful in finance, mostly in places that attract less attention: cleaning data, detecting fraud, estimating risk, and executing large orders at lower cost. As a source of trading profits for private investors, the evidence is thin. Any system, with or without AI, can be held to the same test. After realistic costs, over a period the model never saw, and without choosing the best of many attempts, is there still an edge? If that cannot be answered with verifiable numbers, the label AI adds nothing.