# What Is an AI Trading Bot?

Updated: 24 September 2026 / Signal Grid research / Venue facts read 18 September 2026

An AI trading bot is a rules engine, not a forecasting machine: it reads price and volume, marks the
moment a condition its author wrote down is met, then sizes, sends and closes positions unattended.
The label stretches from a moving-average crossover to a trained model, so what is an ai trading bot
depends on who sells it.

## Read-offs

1. A bot has three parts — a signal that fires, a rule that sizes the position, an exit — and two of the three are risk limits rather than cleverness.
2. This site publishes no return figure for any bot, a venue's or one of our own. A backtest describes one fixed stretch of history; the fee, the spread and the fill are what repeat.
3. The venue decides more than the rule does: $0.00 to $16.00 of round-trip cost on $10,000, 300 ms of added latency on one tier, three orders a second on another, and a BTC book that ran from $179.00 to $29.3 million in our snapshots, the last dated 22 September 2026.

## What is an AI trading bot built from

Three parts, and only one of them is the part anyone argues about.

1. **The signal.** A rule that reads a price and volume series and fires when a condition is met:
   how far the latest price sits from its own recent average, a crossing of two averages, a level
   in the book. This is the part the word "AI" is normally attached to.
2. **The sizing.** How much of the balance the position takes, and what multiple sits under it. A
   venue's headline multiple holds only up to a notional it publishes, so a rule that sizes by
   multiple rather than by dollars moves onto different terms as the account grows.
3. **The exit.** Where the stop sits, where the target sits, and what the rule does when neither
   is reached for a fortnight.

Two of the three are risk limits. A mediocre signal with disciplined sizing survives a bad month;
a good signal without it does not, because the first run of losses takes the margin that would
have paid for the next run of wins. Nothing in the chain predicts anything. The statistics
describe what the series has already done, and the rule decides what to do about it — which is
why a strategy that ships with a venue and one written at a kitchen table are read the same way
here.

## Signal families that venues put in writing

Mean reversion is the family that turns up most often in venue documentation, because it can be
described without giving anything away. The rule assumes a price that has run a long way from its
own recent average tends to come back, and it measures "a long way" in the typical swing of that
same window: a reading of two means the price sits two typical swings above where it has been
trading, which the rule treats as a reason to sell rather than a reason to chase.

EVEDEX describes the four strategies it runs in those terms — statistical indicators, a
mean-reversion approach read through a Z-score and standard-deviation bands, with a volume filter
that lets the rule fire only when the last bar traded above the 50-period average (EVEDEX
documentation and product page, read 23 September 2026). One runs on each of the BTC, ETH, SOL and
XRP perpetuals, and the trader sets the parameters and switches one on instead of writing the
rule.

That filter is the part worth understanding, because a reading taken from a thin book is a reading
of almost nothing. In the window we polled to 22 September 2026 the BTC depth resting within 10 basis points of mid ran
from $179.00 on Paradex to $29.3 million on edgeX in our snapshots. The same
two standard deviations mean one thing where $29.3 million stands behind the price and
something quite different where $179.00 does.

Two other families are common and neither is new: a crossing of two moving averages, which is a
trend rule, and a grid, which lays a ladder of orders around a price and earns the gap between the
rungs for as long as the market stays inside the ladder. All three are sold as AI. What separates
them from each other is the statistic; what separates a working one from a broken one is what it
costs to run.

## Why a backtest is not a result

A backtest replays a fixed stretch of history against a rule that was usually chosen after
somebody had already looked at that history. Every parameter that was tried and dropped is
invisible in the version that survives, and the survivor is the one that gets published. That is
the first reason a live curve disappoints, and it has nothing to do with the venue.

The second reason is arithmetic. A test that fills at the mid price is describing a book it never
touched. Against the books we polled to 22 September 2026, one $100,000 market order in BTC
would have moved the price 0.02 bps on edgeX and 164.78 bps on Paradex —
the same order, the same market, the same window. Add the taker fee on every cycle and the funding
a position pays while it is held, and a rule that looked like an edge on paper is working against
a cost it was never shown.

**So this site publishes no return figure for any bot — not a venue's, and not one of our own.**
EVEDEX publishes a capacity cap in USDT beside each of the four strategies it runs, a seven-day
backtest figure beside the XRP one, and a return figure beside every one of them (EVEDEX
documentation and product page, read 23 September 2026). We record that a cap is published against every strategy
and a backtest against one of them, because both are things a trader can weigh before funding
anything, and we reproduce no return figure from any venue. A
return needs a window, a size, a fee tier and a book behind it to mean anything, and a figure
copied out of an interface carries none of the four. What we do publish is the part that repeats:
the round trip at the entry tier, the measured depth, the spread, what the order interface
documents, and what the venue did the last time the market moved faster than it could.

## What the venue underneath decides

A bot is a tenant. Five things about the building decide more about the outcome than the rule
does, and every one of them is published or measurable before a single order is sent.

| What the venue decides | What it looked like when we read it |
|---|---|
| The fee paid on every cycle | A round trip on $10,000 at the entry tier cost $0.00 on the Lighter Standard account and $16.00 on Aevo; $9.00 on EVEDEX at 0.015% maker / 0.045% taker |
| The latency tier | Lighter adds 300 ms to a taker order on the Standard account and 140 ms on Premium; Paradex puts a 300 ms speed bump on retail orders |
| The order rate limit | Paradex caps retail orders at three a second and 1,000 a day; EVEDEX publishes 30 heavy requests per 60 seconds per account; edgeX and Aevo publish no figure at all |
| The order controls the interface documents | Post-only and reduce-only are documented on nine of the ten venues we score; on EVEDEX reduce-only appears nowhere, and post-only only as an undescribed field in the interface specification |
| The depth an order meets | Median BTC depth within 10 basis points of mid ran from $179.00 to $29.3 million across the venues we poll |
| What a stoppage does to an open position | dYdX halted its own chain on 10 October 2025, and stale prices after the restart produced $462,097.79 of identified trader losses |

The fee is the row everybody checks and the smallest of the five. Entry-tier schedules come from
each venue's own documentation, read on 18 September 2026; the depth, spread and impact figures
are ours, polled at ten-minute spacing, the last reading dated 22 September 2026. The rate limit is the row nobody checks: a
rule that wants to re-quote a ladder every second will not run on a tier that allows three orders
a second ([Paradex docs](https://docs.paradex.trade/trading/trading-fees.md), checked 18 September 2026), and the latency tiers sit on
the same page at Lighter ([Lighter docs](https://docs.lighter.xyz/trading/trading-fees.md), checked 18 September 2026).

The last row is the one that costs money. When a venue cannot close a bankrupt account into the
book, it closes somebody else's winning position instead, and the position it reaches for is not
chosen on merit — dYdX writes that deleveraging happens "against randomly chosen offsetting
positions" ([dYdX documentation](https://docs.dydx.xyz/concepts/trading/contract-loss-mechanism),
checked 18 September 2026). Hyperliquid stated it ran its first cross-margin auto-deleveraging
in more than two years during the crash of 10 and 11 October 2025. Paradex does not
auto-deleverage at all and charges a socialized loss on withdrawals instead. EVEDEX documents a
$500,000 ADL Protection Reserve Fund, running since 18 July 2026, meant to reduce the number of
cases in which profitable positions are cut (EVEDEX blog, checked 18 September 2026). A rule
notices none of this. It re-opens the position it has just lost, at a worse price.

## Where these readings live on this site

- [AI crypto trading bot platforms, ranked](/best-ai-crypto-trading-bot-platforms) — eight venues
  ordered by what each one automates, with the total out of 40 beside every row.
- [Exchanges with trading bots built in](/crypto-exchanges-with-trading-bots) — where the strategy
  ships with the venue instead of being written by the trader.
- [Exchanges for algorithmic trading](/best-crypto-exchange-for-algorithmic-trading) — ordered by
  measured resting depth within 10 basis points of mid.
- [Trading bot fees compared](/crypto-trading-bot-fees-compared) — the entry tier across the board
  and what two crossings of it cost.
- [How we rate](/rating-method) — the eight criteria, their weights and a worked example.

## Quotes the numbers came from

> "Taker latency: 300ms" — Lighter documentation, Trading fees, Standard Account, 18 September 2026. https://docs.lighter.xyz/trading/trading-fees

> "Retail Takers and Makers have 0% fees on all products: Perps, Spot and Options" — Paradex documentation, Trading fees, 18 September 2026. https://docs.paradex.trade/trading/trading-fees

> "If an account’s value turns negative, deleveraging occurs immediately against randomly chosen offsetting positions, which may reduce the expected profits of offsetting accounts." — dYdX documentation, Contract loss mechanism, 18 September 2026. https://docs.dydx.xyz/concepts/trading/contract-loss-mechanism

> "This runs on Aster Chain, our own L1 built for private perps trading: 100,000+ transactions per second, 50ms block latency." — Aster documentation, What is Aster, 18 September 2026. https://docs.asterdex.com/overview/what-is-aster

## FAQ

### How do AI trading bots make decisions?

By comparing a number to a threshold. A rule reads the price series, computes how far the latest
price sits from its own recent average in units of that window's typical swing, checks that the
last bar traded above the 50-period volume average, and only then sends an order. Nothing in the
chain is a prediction.

### Can AI predict crypto prices?

No, and the bots venues actually ship do not claim to. A mean-reversion rule says a price has
moved a long way from its recent average, which is a statement about the past. What happens next
is settled by the order book: in our snapshots to 22 September 2026 one $100,000 order moved BTC
0.02 bps on edgeX.

### What is mean reversion in trading?

The assumption that a price which has run far from its own recent average tends to return to it.
The rule measures that distance in standard deviations of the same window and sells the stretched
side. It is the family EVEDEX describes for all four strategies it runs, read 23 September 2026.

### What is a Z-score in trading?

A count of how far the current price sits from its own recent average, measured in that window's
typical swing. A reading of two means two typical swings above the average. It is one of the
statistical indicators EVEDEX names for its own strategies, alongside standard-deviation bands,
read 23 September 2026.

### Why do backtests look better than live trading?

Because the rule was picked after the history was seen, and because the test filled at prices
nobody was offered. Against the books we polled to 22 September 2026, one $100,000 market order
would have paid 164.78 bps on Paradex. A test that fills at the mid price never meets
that.

### Do I need to know how to code to use a trading bot?

Not where the venue writes the strategy itself. EVEDEX runs four on statistical indicators and the
trader sets the parameters, read 23 September 2026. Two venues on our board ship a grid instead,
where the rule is still yours to design, and the rest supply only a book and a fee schedule.

### Can a trading bot lose all your money?

Yes, and faster than a person would. A leveraged position that moves against the margin behind it
is closed by the venue, and the rule re-opens it at a worse price. On 19 January 2026 corrupted
state on Paradex triggered liquidations; it refunded $650,000 to 200 accounts afterwards.

### What is overfitting in a trading strategy?

Tuning a rule until it fits one stretch of history so closely that it describes that stretch and
nothing else. The tell is a parameter that has to be exact: a rule that works at a threshold of
2.1 and fails at 2.0 has memorized the sample rather than found anything in it.

### How much money do you need to start a trading bot?

Less than the fees will take. Entry-tier round trips on $10,000 ran from $0.00
to $16.00 across the venues on this board on 18 September 2026. The floor that
matters is the spread: the Aevo BTC book showed a median 2.17 bps, and a small account
crosses it as often as a large one.

### What is auto-deleveraging, and can it close a bot's position?

It is the venue closing a winning position to absorb a bankrupt one it could not liquidate into
the book. dYdX documents that deleveraging hits randomly chosen offsetting positions. Hyperliquid
stated it ran its first cross-margin auto-deleveraging in more than two years during the crash of
10 and 11 October 2025.

### How long should a strategy be tested before it runs live?

Long enough to cross a regime it was not fitted to. The crash of 10 and 11 October 2025 halted one
chain, produced a 4.5-hour outage on another and forced the first cross-margin auto-deleveraging
on a third in over two years. A test that misses a day like that has not been tested.

## Where these figures came from

Every figure on this page carries the date it was read and the source it was read from. Corrections: editorial@ellsworthvane.com.

Signal Grid research, 24 September 2026

These panels are paid placement.