How Automated Crypto Trading Actually Works

"AI trading" is one of the most abused phrases in this industry. Stripped of marketing, a systematic trading operation does four things in a loop: it observes the market, scores possible positions, sizes them against risk limits, and executes. Each stage has a specific job and a specific way of failing.

Understanding the four makes it much easier to tell a real operation from a story.

Stage one: observe

Before anything can be decided, the current state of the market has to be measured. That is more than price. A serious setup is also watching:

• Order-book depth — how much can actually be bought or sold before the price moves against you.
• Funding rates on perpetual futures, which reveal how crowded one side of a trade is.
• Realised volatility, because the same signal means different things in a calm market and a violent one.
• Spread and venue health, since a quote nobody will honour is not a price.

How it fails: bad data. A stale feed, a single venue printing a wick, or a thin market quoting nonsense will all produce confident, wrong decisions. Discarding suspicious data is unglamorous and it is most of the work.

Stage two: score

Candidate positions get ranked. This is where the modelling lives — and where the word "AI" usually gets attached. In reality the useful models are often unexciting: statistical relationships between instruments, momentum measured over several horizons at once, mean reversion in spreads that have historically converged.

The important part is that scores are conditional. A momentum signal that works in a trending market is often actively harmful in a choppy one, so a score is only meaningful alongside an assessment of the current regime.

How it fails: overfitting. Any model fitted to history can be made to look excellent on that history. The question is whether the relationship it found is structural or a coincidence in the sample, and you only find out afterwards.

Stage three: size and gate

This is the stage that separates operations that survive from ones that blow up, and it gets the least attention in marketing material. Before an order exists, the proposed position is checked against limits:

• Position size relative to total capital and to the instrument's liquidity.
• Correlation limits, so six positions that are secretly the same trade do not get taken as six independent ones.
• Exposure caps per instrument, per venue and in aggregate.
• Drawdown rules that cut risk automatically as losses accumulate, without anyone having to decide to.

A position that breaches a limit is reduced or skipped. The discipline is in never waiving the rule because this particular signal looks unusually good.

How it fails: correlation that only appears under stress. Assets that diversify each other in normal conditions tend to move together precisely when it matters, which is how a portfolio that looked diversified turns out to have been one bet.

Stage four: execute and reconcile

An order has to reach the market without giving away its own intention. Large orders are split and spread across venues to limit market impact, because a single order big enough to move the price pays for that movement itself.

Then — and this is the step amateurs skip — fills are reconciled against the strategy's intended book. The position you think you hold and the position you actually hold must match. Partial fills, rejected orders and venue outages all create drift, and drift that goes unnoticed compounds.

How it fails: slippage and liquidity withdrawal. The spread widens exactly when you most want out, so a forced exit in a dislocated market costs far more than a planned one.

What automation genuinely improves

It is worth being precise about the benefit, because it is real but narrow. Systematic execution removes a category of human error:

• The rule applies at 3am as consistently as at 3pm.
• A losing position does not get held longer because someone cannot face closing it.
• Position sizes do not quietly grow after a good week.
• The decision is reproducible — you can inspect afterwards why it was taken.

What it does not improve

Automation does not predict the future, and no amount of modelling removes market risk. A strategy can be well built, correctly executed and still lose money in a month the market has not done before. Any platform implying otherwise is selling certainty it does not have.

Our own version of this pipeline is described on the technology page, including the limits we cannot engineer away. If you want the risk side in full, understanding crypto risk is the companion to this guide.