What automated trading is
Automated trading means a program follows trading rules and places the orders. Instead of watching a chart and clicking buy, you write down what counts as an entry, where the stop goes, where the target goes and how big the position is. Software checks those rules all day and acts when they are met.
You will see the same idea under several names: algorithmic trading, systematic trading, trading software, a trading bot. In crypto, "bot" is the common word, and it usually means a program connected to an exchange account. The label matters less than the question underneath it: what are the rules, and has anyone shown they work after costs?
The appeal is real. Software does not get tired, does not freeze after three losses and does not move a stop because it hates being wrong. The risk is also real, and it is the part sellers like to skip. A bot enforces a plan perfectly. If the plan has no edge, you lose money perfectly and quickly.
Three kinds of system
Not all bots are built the same way, and the differences change how much you should trust a result.
Rule-based. A person writes fixed rules, for example "buy when the fast average crosses above the slow one, stop below the last swing low." Everything is readable and nothing changes unless the author changes it. This is the easiest kind to understand and audit. See technical analysis explained for the indicators these rules usually use.
Optimised. The structure is fixed, but the settings are searched automatically: which averages, which timeframe, which stop distance. A computer tries thousands of combinations on past data and keeps the winner. This is where the danger of fitting the past appears, because the more settings you search, the more likely one of them looks brilliant by chance.
Machine-learned. A model is trained on data to predict something, such as whether price will rise over the next few hours. It can pick up patterns a human would not write down. It can also pick up noise, and it is often hard to explain why it took a given trade. Models also go stale when the market they were trained on changes.
Some systems combine these. A genetic algorithm, for instance, optimises settings by evolution rather than by exhaustive search, which we come back to at the end.
The pipeline: signal, sizing, execution, risk controls
Almost every automated system, whatever its type, has the same four stages.
- Signal. The part that decides "now" and "long or short." It reads prices and indicators and outputs an idea. For what a complete idea contains, see what is a trading signal.
- Sizing. How much to put on. Good sizing starts from the stop distance and a fixed amount you accept losing, not from how confident the signal feels. Risk management in crypto trading covers the arithmetic.
- Execution. Turning the decision into orders: market or limit, split entries, attaching the stop and target at once. This is where fees and slippage are paid.
- Risk controls. The brakes: a maximum number of open positions, a daily loss limit, a kill switch, and checks that every position actually has a working stop.
Beginners tend to judge a bot by the first stage alone. In practice, the last three decide whether a decent signal survives contact with a real account.
How bots connect to exchanges
A bot talks to an exchange through an API key, a pair of strings that lets software act on your account. Exchanges let you set what each key may do, and that setting is your main protection.
- Trade-only permissions. Enable reading balances and placing orders. Never enable withdrawals. A leaked key without withdrawal rights can make bad trades but cannot move your coins to someone else's wallet.
- IP whitelist. Where the exchange offers it, restrict the key to the fixed address of the machine running the bot. A stolen key then fails from anywhere else.
- One key per bot. If you need to revoke access, you cut one program, not everything.
- Nothing in a chat, a screenshot or a spreadsheet. Anyone who sees the secret can use it.
Be cautious about handing a key to any third party you cannot inspect. Even trade-only access lets someone else open large leveraged positions on your account. See crypto leverage explained for how quickly that can go wrong.
What goes wrong in practice
Most bot failures are not dramatic hacks. They are quiet and technical.
Overfitting. The rules were tuned until they explained the past, which means they explain noise. A strategy that looks excellent on history and falls apart live is the classic story.
Look-ahead bias. A backtest accidentally uses information that was not available at the time, such as a candle's close used to decide an entry at its open, or an indicator array shifted by a few bars. The result looks like skill and is really peeking. Forward testing, covered in paper trading explained, removes this particular problem.
Fees and slippage. Each trade pays an entry fee and an exit fee, plus the gap between the expected price and the fill. A bot that trades often with small targets can have a real, tiny edge and still lose after costs.
Outages and silent failures. An exchange API times out, a connection drops, a rate limit trips, a process dies. The worst version is an order that fails without any message, so the bot believes it is protected by a stop that does not exist. Good systems verify that every position has its protection and alert loudly when something does not.
Regime change. A rule that thrived in a steady trend can bleed in a sideways market. The bot cannot tell by itself that the conditions changed. Market regimes explained describes why a few good months prove little.
A worked example (hypothetical numbers)
This is an invented illustration, not a real result. Suppose a bot trades one pair with a stop 1% from entry and a target 1.5% away. Over a backtest of 200 trades it wins 110 and loses 90.
Gross, that looks fine: 110 x 1.5% = 165%, minus 90 x 1.0% = 90%, a net 75% of position size over 200 trades, about 0.375% per trade.
Now charge costs. Say a round trip of fees plus slippage is 0.15% per trade. Over 200 trades that is 30%, leaving 45%, or 0.225% per trade. Still positive, but 40% of the gross edge is gone.
Then ask about the backtest itself. If the settings were chosen after searching many combinations, or any look-ahead leaked in, the true win rate going forward might be 50% rather than 55%. At 100 wins and 100 losses: 100 x 1.5% = 150%, minus 100%, equals 50%, then minus 30% in costs equals 20% over 200 trades, about 0.1% per trade. A little more slippage or a slightly worse win rate, and it is zero or negative.
The point is how thin the margin is. A bot does not need to be broken to lose money. It only needs to be slightly less good than its backtest, which is the normal case.
How to evaluate a bot or a bot seller honestly
You can test almost any claim with five questions.
- Is the record forward or backtested? Live or forward paper results count for more than history. Ask when the record started and whether the rules changed since.
- Are fees included? A record without a fee line is a gross number. Ask whether every trade is charged a full round trip.
- Are the losses shown? A complete list of closed trades, with the bad ones, is the minimum. Screenshots of winners are marketing.
- Is there a baseline? Compare against random entries with the same stops and targets. If chance does as well, the signal adds nothing.
- Is the sample big enough and varied enough? A few dozen trades in one kind of market says little. Hundreds across rising, falling and flat conditions says more.
Also notice what the seller asks of you. Requests for withdrawal-enabled keys, promises of fixed returns, or pressure to deposit with a specific broker are warning signs.
Execution bots versus signal sources
There are two different products hiding under the word "automation."
An execution bot places orders on your account. It is fast and does not need you to be awake. The costs are key security, exposure to its software bugs, and the fact that its mistakes happen in your account at machine speed.
A signal source publishes trade ideas and you decide what to do. It is slower and needs your discipline, but your keys and funds never leave your hands, and you can skip anything that looks wrong.
Neither is safer by definition. The right choice depends on whether you would rather carry technical risk or behavioural risk, and on whether the underlying rules have any proven edge in the first place.
The biggest beginner mistake
The most common mistake is trusting a backtest chart and switching on a bot with real money and leverage. The chart is smooth because it was built after the fact. The bot goes live, meets fees, slippage and a different market, and the first losing streak arrives at a size nobody rehearsed.
A safer order of operations is to run any bot on paper first with fees charged, then at a very small size with a hard cap on what the account can lose, and to scale only after a forward record exists. Treat the backtest as a reason to test further, never as a result.
What running a self-evolving bot in public teaches
Darwin Lab is a self-evolving genetic-algorithm bot for Binance Futures. Each strategy carries about 30 genetic parameters covering indicators, timeframes and risk rules. Strategies paper-trade around the clock, and every 4 hours the winners breed and mutate while the losers are eliminated. Promotion depends on forward paper results, not backtests, precisely because of the overfitting and look-ahead problems above.
It traded real money from April to June 2026. Funds were then pulled to fix fee drag, and the signals have run on paper since June 2026. Measured net of fees, its published paper signals have not yet shown an edge over random entries with the same stops and targets. We say that plainly because it is what the data shows so far, and a bot seller who cannot say the equivalent about their own product should make you suspicious.
Three lessons stand out. First, fees are a design constraint, not a footnote, since they decided the move to paper. Second, evolution is a search method, not a source of edge: breeding strategies on noise produces well-bred noise. Third, publishing the full fee-inclusive record at /track-record/ is uncomfortable and also the only way a claim becomes checkable. No API keys are ever shared with users, so nobody has to trust us with their account.
How Darwin Lab works | Join the free channel
This article is educational and not financial advice. Automated trading and futures carry a high risk of loss, and past or simulated results do not predict future ones.