Are AI Trading Bots Profitable?

Sometimes, in a particular market and over a particular period. But there is still no strong public evidence that AI trading bots can produce durable profits after fees, slippage, and risk are accounted for. A bot can sound convincing, analyze a market well, or even predict price direction often—and still lose money.

Trading bots are often judged by appealing numbers: prediction accuracy, a few profitable months, or a streak of winning trades. But the more important questions are whether the strategy survives a bad market, how much it loses at its worst point, and whether it beats a simple passive investment after all costs.

That is where many ambitious claims about AI trading still fall short.

Accurate predictions do not automatically make a profitable strategy

A bot needs more than the right answer to “Will this stock go up or down?” It also has to size the position, choose when to enter and exit, limit losses, and know when not to trade at all.

Imagine a strategy that makes $10 on each of nine trades, then loses $200 on the tenth. Its win rate is 90%, but it still loses money overall. The same problem appears when a bot reacts too late to news, trades too often, or keeps a risky position through a sharp market reversal.

That is why prediction accuracy is mostly a marketing metric. A serious evaluation needs returns after costs, maximum drawdown, risk-adjusted performance, and a comparison with a passive benchmark.

What stricter tests show

In StockBench, researchers placed LLM agents in a sequential trading environment. Each day, the agents received prices, fundamentals, and news headlines, then had to decide whether to buy, sell, or hold. Most of the tested agents did not outperform a simple buy-and-hold strategy. A few models showed stronger returns or better risk management, but the authors describe this as early potential—not proof that AI systems can reliably make money in live markets. Read the StockBench preprint.

Another study tested LLM-based investment strategies over longer periods and across more than 100 stocks. The apparent advantages reported in earlier work weakened substantially under this broader test. The strategies tended to be too cautious in bull markets, where they lagged passive benchmarks, and too aggressive in bear markets, where they suffered heavy losses. Read the FINSABER preprint.

That does not make AI useless in finance. It can help analyze filings, news, and anomalies in large datasets, or help a human trader think through scenarios. But turning a large volume of information into a plausible answer is not the same as creating a durable market advantage.

Why backtests can create false confidence

Many strategies look stronger in historical testing than in live trading. A system can be unintentionally fitted to the past, tested on a favorable period, or evaluated on companies that survived while weaker peers disappeared.

Language models add another complication: historical prices, dates, and market narratives may have appeared in their training data. A bot can then appear insightful because it partly recognizes the past rather than because it has learned a transferable investment process.

A recent preprint that controlled for this kind of leakage found that much of the apparent performance of LLM agents could be explained by broad market movement and style exposure, rather than persistent stock-selection skill. Read the KTD-Fin preprint.

AI Caver take

An AI trading bot may be a useful analysis or discipline tool. But “the bot predicted the market correctly” is not evidence that it is profitable.

A fair test should answer five questions: What is the result after all costs? What is the maximum drawdown? Does it beat a passive index? Does the effect persist in a new period? And can the profit be explained by something other than luck or a rising market?

For most public AI trading systems, the honest answer is still: interesting experiments, yes; proven and repeatable profits for ordinary users, no.

Read the full research

For a concrete example of why a 71% win rate can still produce a loss, read Accuracy Does Not Save You: Why an AI Trader Can Be Right and Still Lose Money.