A trading bot is any software that trades for you by a fixed rule. The full taxonomy, how one reaches your account, why even a well-built bot can still lose money, and the pattern regulators keep warning about.
November 16, 2026
11
min to read
The short answer
A trading bot is software that buys and sells for you by a fixed rule, like a grid, a scheduled buy, or an alert-based signal, around the clock. It reaches your account through an API key or a wallet, not your full login or seed phrase. It runs that rule exactly, for better or worse, forever.
A trading bot is a program that watches a market and places orders when a condition it was given turns true. It doesn’t weigh a decision the way a person does. It checks a rule, such as “the price dropped 2%” or “the RSI, a momentum gauge from 0 to 100, closed under 30,” and if the rule matches, it buys or sells. That is the whole job, run without a coffee break, a bad night’s sleep or a change of mind.
Two different things both get called “trading bots.” One is a fixed rule, written once and run forever, which is what most of this article covers. The other borrows the name but works differently: an AI trading agent uses a language model, the kind of AI behind Claude and ChatGPT, to read your request and the market, then decide what to do next, which makes it more flexible and harder to predict than a rule. Some products marketed as “AI trading bots” are rule bots with an AI assistant bolted on for setup only; our comparison of 12 AI trading bots sorts out which is which. For the fuller definition of the agent side, see what an AI trading agent is, and for a job-by-job look at where a fixed rule beats an agent and where it doesn’t, see trading bots vs AI agents.
Under the hood, almost every trading bot is one of six jobs. Knowing which one tells you what market it suits and how it tends to lose money.
| Kind | What it does | What it needs to work | How it usually fails |
|---|---|---|---|
| Grid | Places a ladder of buy orders below the price and sell orders above it, re-placing each one after it fills. | A price that swings back and forth inside a range | The price breaks out of the range: it keeps buying all the way down, or sells out of a rally early. Our grid trading backtest shows both failure modes on real BTC data. |
| DCA (dollar-cost averaging) | Buys a fixed amount on a schedule, or adds bigger orders as the price falls, aiming to sell a little above the average cost. | A market that eventually comes back | A long decline with no recovery; on a leveraged position it can also run up an ongoing funding cost. Our DCA backtest tests exactly that against a lump sum. |
| Arbitrage | Buys where a price is lower and sells where it’s higher at the same moment, often across two venues. | A fast connection and low costs, since the gap is usually tiny and brief | A slow fill, a cost spike, or other bots racing for the same gap. |
| Market making | Quotes a buy price and a sell price at once, earning the small gap between them when both sides fill. | Capital, speed and code you trust | It gets filled on the wrong side right before a sharp move, and ends up holding the losing side of a trend. |
| Signal | Trades when an outside alert fires, often one a charting site sends to the bot’s web address. | A signal source you already trust | A late, missed or doubled alert, or a signal that only looked good once it had already happened. |
| AI or agent | A language model reads your request and the market data, then builds or runs the trade itself. | Clear limits, since it decides more than a fixed rule does | It misreads an instruction, acts on a stale or wrong read with confidence, or follows text it was never meant to treat as an instruction. |
Real products mix these constantly: a “grid bot” with an AI assistant that only helps you set the range, or an AI agent that writes a signal rule and then runs it like a plain bot. Our guide to Hyperliquid trading bots works through seven kinds of bot on one onchain exchange, with real market numbers. Ethy is one of those mixes: an AI agent that can also run scheduled buys and a dip-buying ladder, as our crypto trading bot page shows.
A bot has to reach your money without holding all of it. Two models cover almost every product.
| What you connect | What it can usually do | What it should never be able to do |
|---|---|---|
| An API key on an exchange or broker | Read your balance, and place and cancel orders, if you turned trading on | Withdraw. On Binance, for example, turning on withdrawal permission requires setting an IP address restriction first, and a standard key with no such restriction can only read. |
| An API wallet on an onchain exchange | Sign orders for your account | Transfer or withdraw funds. Hyperliquid’s own Python SDK describes this kind of wallet, which it calls an agent, as one that “does not have permission to transfer or withdraw funds.” |
| Your password or seed phrase | Everything, including moving your money out | This is never what a legitimate trading bot asks for. |
An API key’s permissions are something you choose when you create it, which is why giving a bot its own key, scoped to trading only, matters more than trusting the bot itself. Onchain, the equivalent is an API wallet, sometimes called an agent wallet: a separate key your main wallet approves, which Hyperliquid’s own docs say you can set to expire, at most 180 days out, after which it stops working. Either way, the same rule holds: a bot that only trades can still lose your money by trading badly, but it can’t walk away with it. Our comparison of Ethy and classic trading bots and our Hyperliquid bot guide go deeper on how specific products set this up.
The starker failures are mechanical. On August 1, 2012, old code that was meant to be retired activated inside the order-routing system of Knight Capital, a US broker-dealer. In the first 45 minutes of trading, it sent more than 4 million orders while trying to fill 212 customer orders, and the firm ended up holding positions it never meant to take. The SEC’s own account of the case, in its press release on the matter, puts the result plainly: Knight Capital “acquired several billion dollars in unwanted positions, and eventually suffered a loss of more than $460 million.” That was an institutional system, not a retail bot, but the failure mode is the same at any size: a trading bot does not know its own code is wrong. A dropped price feed, an exchange’s rate limit queuing your cancels behind a fast move, or two scripts sharing one key and racing each other’s order numbers, the nonce problem Hyperliquid’s own API docs warn about, can all turn a small rule into a large loss while nobody is watching.
The more common failure has nothing to do with bugs. A grid bot built for a range keeps buying as the market falls straight through the bottom of it. A DCA plan built to catch a dip keeps adding to a position that never comes back. A signal rule that looked sharp on a trending month gets chopped up by a sideways one. And a rule whose settings were tuned to fit the last year of a backtest can be tuned to noise that will not repeat, a trap our backtesting guide and the research summarized in our guide to AI trading scams both cover. A bot cannot tell you the market has changed character. It can only tell you that the rule still says what it always said.
Most trading-bot products are ordinary software: your account stays where it was, the bot places orders in it, and you can watch every fill. Some products are not software at all, and the US Commodity Futures Trading Commission (CFTC), the Securities and Exchange Commission (SEC) and FINRA, the US brokerage regulator, have each warned about the same pattern dressed up as a bot.
The CFTC’s 2024 advisory on artificial intelligence is direct that AI technology can’t predict the future or a sudden swing in the market, whatever a seller claims about its bot. The SEC has brought cases that show what an empty claim looks like in practice: in October 2024 it settled charges against Rimar Capital and its owner, who the SEC says “lured investors and clients with multiple fabrications, including with buzzwords about the latest AI technology,” behind a platform that did not do what it claimed, after raising close to $4 million from 45 investors. FINRA’s own 2025 article on unregistered auto-trading services adds the narrower warning that matters most once you’re actually looking at a bot: these services are often promoted on social media with talk of consistent returns and AI, and you should be wary of any that asks for the username and password to your brokerage account, something no legitimate bot needs.
Our guide to AI trading scams goes through the full playbook regulators describe, the red flags, and the checklist to run any offer through before you connect anything.
A good tool answers these in writing before you connect anything:
A trade-only API key or an API wallet is normal. A request for your password or your seed phrase, the words that restore a wallet, is not.
Place orders only, or also move funds, change its own settings, or add leverage you didn’t set.
A backtest on past prices, a demo or paper mode on live prices, or a venue’s own testnet. If none exist, treat the first small run as the test.
A subscription, a fee per trade, a cut of profits, or a mix, on top of whatever the exchange or broker itself charges.
A named company, or public code with recent activity and answered issues, not an anonymous “proprietary algorithm.”
One action that cancels its open orders, and a clear answer for what happens to any position still open.
Once you know what to look for, the products themselves are compared in our list of the best AI trading bots and, for onchain markets, the best Hyperliquid trading bots. If you want to start without paying, see what a free AI trading bot really gets you. If the real question is a fixed rule against a tool that reads the market and decides more on its own, that split is covered in trading bots vs AI agents.
Using software, AI included, to place your own trades isn’t illegal in itself; what regulators go after is fraud and unregistered firms. A firm that manages your money, gives advice, or runs a trading platform for others needs to be registered, so check the SEC’s and FINRA’s registries before you trust one with that job.
Some do, in some markets and periods, and lose money in others. No bot removes the two costs that eat into returns: trading fees and a rule tuned to the past rather than the future. Treat a backtest as a test, not a promise, and read our guide on what the research says about bot returns before you trust one.
Not by withdrawing, if it only holds a trade-only API key or an API wallet, which a well-built exchange or onchain venue won’t let withdraw funds. It can still lose money by trading badly, so the bigger risk is usually a bad rule, not theft, unless you hand over your password or your wallet’s seed phrase. Never do that.
A bot follows fixed conditions someone set in advance: the same inputs always produce the same order. An AI trading agent uses a language model to read data and plain English and decide what to do next, which makes it more flexible and harder to predict. Our explainer on AI trading agents and our task-by-task comparison with bots go deeper.
It depends on what you trade and how much you want automated. Our comparison of AI trading bots covers crypto and stock tools side by side, and our list for Hyperliquid covers onchain perpetual futures specifically.
Not in the classic sense. Ethy is an AI trading agent: you ask it about a market, it reads the chart and tests an idea you describe in plain words on past data, and then you can turn that exact rule into an automation that trades by itself with stops around it. The automation behaves like a bot, which is what makes it testable first.
Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.
Deploy your AI trading agent in one click.
