AI Crypto Trading Bot

What the label covers

A crypto exchange trades around the clock, so a bot never has to sleep, and that is why bots are common there. The word AI is then added to many different things:

• A plain rule. A moving average cross or a momentum filter with a new name. See crypto trading bot for the usual types.
• A fitted model. A statistical or machine learning model trained on past prices, volume and other data. It learns patterns from the past. See machine learning trading.
• A language model. A chat assistant that explains an idea or writes bot code. It cannot see the future. See the ChatGPT trading bot page.
• A signal or sentiment feed. It reads news or posts and shows a score. The score is not a trade. What is done with it is the strategy.

In every case the bot still sends orders that pay a fee and cross a spread, and the market does not care what the bot is called. For the general questions to ask of any such tool, see AI trading software. This page is about the crypto case and one failure in particular.

Why a model trained in one regime fails in another

A regime is a stretch of market with a steady character: a strong uptrend, a steady drop, a tight sideways range, a period of wild swings. A fitted model learns whatever pattern paid off in the data it was shown. If that data was mostly one regime, the model has learned that regime, not the market.

Three things make crypto especially exposed. First, the history is short and dominated by a few large moves, so a model can fit those moves closely. Second, big changes arrive suddenly and from outside the price chart. Third, liquidity changes between busy hours and quiet ones, so the same signal fills at a worse price when the market is thin. The model has no way to know it has left the conditions it was built for. It keeps sending orders.

This is a form of overfitting: a rule that suits the past it was fitted to and little else. The test for it is a period the model never saw, ideally of a different character. See forward testing.

Worked example: one rule, two regimes

Everything here is made up. A model trades a coin pair with the same rule in two stretches of 40 trades each. The average win is 2.4% of the amount put into the trade, the average loss is 1.2%, and costs (fee plus spread) are 0.2% a trade. The first stretch is the trending months it was trained on. The second is sideways months it has not seen.

Calculated by us on 29 Oct 2026, exact. With these sizes the rule needs a win rate of 38.9% to cover its costs: the average loss plus the cost, 1.4, divided by the average win plus the average loss, 3.5999999999999996. In the trending months it won 55.0% and cleared that. In the sideways months it won 32.5% and did not, so the same logic that looked sound lost money. Nothing about the rule changed. The market did.

Notice what a single backtest over the first stretch would have shown: a 55.0% win rate and a rising line. The second stretch is the only part that tells you about robustness, and it is exactly the part a seller leaves out.

How to test any such bot on paper

• Use paper or a small demo run, and keep your own log. Time, pair, size, entry, exit, fees. Do not rely on the bot's own report. See paper trading.
• Cover more than one regime. A trend, a range and a sharp drop. If the run only covered one, say so, and treat the result as one regime's answer.
• Count every cost. Exchange fees, the spread, and for leveraged products the funding or borrowing charge. They differ by exchange, so check yours.
• Measure the loss side. Average loss, longest losing run and the worst fall from a peak. See drawdown.
• Compare with a plain baseline, such as holding nothing or a simple rule, over the same days.
• Expect paper fills to be kinder than live ones. Paper results are a best case. Thin markets and fast moves cost more in real trading.

How it hits bots

A bot sized for one regime keeps its size when the regime ends. A trend-following bot in a range buys the top of each swing and gets stopped out. A range-trading bot in a trend keeps adding to a losing side, and a grid or a martingale bot can add size as it does so. A model that "keeps learning" may retrain after losses, so its record is many different strategies in a row and cannot be judged as one. If the bot trades with leverage, a move against it can also end in forced closing, covered by liquidation price.

What this page does not tell you

It does not say any AI crypto bot works, or does not. It cannot tell you how common regime change is or how large the effect is. The only real answer comes from your own log over a long enough run.