Automated Crypto Trading: The Complete 2026 Guide

Almost every guide on this topic describes the first half of a trading bot: indicators, entry rules, backtests. That half is the easy half. This guide covers the whole loop, including the parts that only show up once real money is in the market: partial fills, rejected stop-loss orders, funding drift, and what happens when your process restarts while a position is open.

• A signal is the cheapest part of a bot. Execution, protection and recovery are where accounts are lost.
• Every entry must place its exit bracket in the same operation. A position without a live stop is the single most expensive bug class.
• Round-trip cost scales inversely with stop distance — a 0.5% stop pays roughly 4× the cost-in-R of a 2% stop on identical fees.
• Backtests that don't model fees, funding and intra-bar sequencing will overstate edge by a wide margin.
• Automation removes execution error, not strategy error. A losing strategy automated is a losing strategy that loses faster.

What is automated crypto trading?

Automated crypto trading means a program holds your exchange API key and places orders on your behalf when conditions you defined in advance become true. You are not delegating the decision — you are delegating the execution of a decision you already made, in writing, before the market moved.

That distinction matters more than it sounds. A discretionary trader decides at the moment of maximum emotional pressure. An automated system decides at the moment of maximum calm, then executes without renegotiating. The rules are identical; the psychology is not.

The mechanics are simple enough to state in one paragraph. A worker process wakes on a schedule, pulls current market data, evaluates your conditions, and if they pass, submits an order to the exchange over a signed REST call. It then places the protective orders, records what happened, and goes back to sleep. Everything else in this guide is detail on those five steps and the ways each one fails.

How does a crypto trading bot actually work?

A production bot is a loop, not a script. Five stages, each with its own failure mode:

Most published bot tutorials cover Ingest and Evaluate, occasionally Execute, and essentially never Reconcile. Reconcile is what separates a weekend project from something you leave running.

Evaluation cadence and why it matters

How often your bot checks conditions determines which strategies are even possible. A system evaluating once a minute cannot scalp a 15-second order-book imbalance; a system evaluating on every tick will burn API rate limits and fire on noise. On TradeFloor, indicator triggers evaluate on a 60-second cadence, offset from one another so the RSI pass and the order-book pass don't collide on the same second — enough resolution for 5-minute-and-above strategies, deliberately too coarse for latency arbitrage, which is a game retail infrastructure loses.

What signals do automated systems trade?

Signal families, roughly ordered by how much infrastructure they demand:

Indicator triggers

RSI, moving averages, MACD, Bollinger position. Cheap to compute, easy to reason about, heavily crowded. Their value in an automated system is less about edge and more about consistency: an RSI rule fires the same way at 3am as it does when you're watching. See the RSI indicator explained for how the calculation behaves in crypto's volatility regime.

Market structure

Swing highs and lows, change of character, break of structure, fair value gaps, order blocks. More expensive: you need a stable definition of a swing point and enough history to find one. A structure engine reading a 100-bar window and one reading 5,000 bars do not produce the same structure — they produce different structures, and the shorter one is often an artifact of where the window happened to start.

Order flow

Order book imbalance, tape aggression, liquidation clusters. Highest resolution, shortest shelf life. Order book imbalance is computed as (bid_volume − ask_volume) / (bid_volume + ask_volume), giving a bounded −1 to +1 reading. Detail in what order book imbalance is and the practitioner treatment in building an OBI strategy.

Positioning and flow

Open interest divergence, funding rate extremes, whale wallet movement. Slower, more contextual, better as a filter than a trigger. Whale tracking covers reading large-holder behaviour without over-reading it.

News and events

Classified headlines, sentiment scores, scheduled macro. Useful mostly as a veto: knowing not to open a new position in the ninety seconds after a high-impact headline is worth more than trying to trade the headline. See how to trade crypto news.

Position sizing: the part that decides whether you survive

Sizing is where most automated systems quietly go wrong, because the wrong method still produces plausible-looking trades for months.

Fixed notional — always $1,000 per trade — is the default in most tutorials and is structurally broken. A trade with a 0.4% stop and a trade with a 3% stop carry seven times different risk at the same notional. Your equity curve becomes a lottery on which setups happened to have wide stops.

Fixed fractional risk — always 1% of equity at the stop — is the correct default. Size is derived, not chosen: size = (equity × risk%) / |entry − stop|. Every trade now costs the same when it's wrong, which is the only way an expectancy calculation means anything.

Volatility targeting scales the fraction itself by realised volatility, so a quiet regime and a violent one produce comparable equity variance. This is the refinement, not the starting point.

The cost asymmetry almost nobody accounts for

Trading costs are usually quoted as a percentage of notional — say 7.5 basis points per side, so roughly 0.15% round trip on a taker-taker fill. That number looks small and constant. In R-multiple terms it is neither:

cost_in_R = round_trip_fraction × entry_price / |entry_price − stop_price|

At a 2% stop, 0.15% round trip costs about 0.075R per trade. At a 0.5% stop it costs about 0.30R — four times as much, for identical fees. A scalping strategy showing +0.2R gross expectancy is net negative; a swing strategy showing the same number is comfortably positive. Any backtest that reports gross R and any strategy comparison that ignores stop distance is producing a number that cannot be acted on.

Protective orders: the bracket must exist before you look away

The invariant: an entry and its stop-loss are one operation, not two. If the entry fills and the stop is rejected — wrong tick size, notional below the exchange minimum, a transient rate limit — you now hold a live leveraged position with no protection, and nothing in the system knows.

Concretely, Hyperliquid rejects reduce-only limit orders below roughly $10 notional. A five-leg take-profit ladder on a small position can have its last two legs silently fail that check while the first three place fine. Your bot reports "entry successful" and your actual exposure is not what your config says.

Two defences, and you want both:

• Validate before submitting. Compute every ladder leg's notional at config time and refuse a configuration whose legs fall under the venue minimum.
• Reconcile continuously. A separate process compares open positions against live protective orders on a short cycle and re-places anything missing. This has to run independently of the process that opened the trade, because the most common reason a stop is missing is that the opening process crashed.

Backtesting without fooling yourself

A backtest is a claim about a strategy's behaviour on data it has already seen. Almost every way that claim goes wrong has a name:

The practical test: run the sweep, take the best configuration, then run it untouched on a period the sweep never saw. If performance collapses, you fitted noise. It usually does at least partly — the question is how much, and whether what survives is still worth trading after costs.

Risk controls that operate above the strategy

Strategy-level rules manage one trade. Account-level rules manage the thing that actually ends accounts: correlated exposure and behavioural spirals.

• Correlated cluster caps. Eight long positions across eight majors is one long position with extra fees. Cap exposure per correlation cluster, not per symbol.
• Account drawdown gradient. Reduce size as account drawdown deepens, before any hard stop triggers.
• Liquidation distance monitoring. Watch margin ratio continuously; escalate through warning, de-risk recommendation and — only if you explicitly arm it — forced close. Liquidation explained covers the calculation.
• Behavioural circuit breakers. Consecutive-loss counts, re-entry speed after a loss, daily loss caps. These target the human, not the market. Anti-tilt makes the case that this is the highest-yield control most traders never install.

Choosing an exchange for automated trading

The three venues most retail automation runs on differ in ways that matter more to a bot than to a manual trader — custody model, API authentication, rate limits, and how the WebSocket behaves when it degrades. The general trade-offs are in Binance vs Bybit vs Hyperliquid; the developer-level comparison, including auth schemes and reconnection behaviour, is in the API comparison.

Measuring whether it works

Win rate is the most quoted and least useful metric in retail trading. A 30% win rate at 4R and a 70% win rate at 0.3R are wildly different businesses, and win rate cannot tell them apart. Track expectancy instead:

expectancy_R = (win_rate × avg_win_R) − (loss_rate × avg_loss_R)

Then track it net of cost, and track it per strategy, per symbol and per market regime — because an aggregate number hides the case where one strategy funds another's losses. A journal that records the exit reason on every close makes this trivial; without one you are guessing. See building a trading journal.

What automation does not fix

Automation removes execution error: the missed entry, the moved stop, the position sized by feeling. It does not create edge. If the underlying rules have negative expectancy after costs, automating them converts a slow loss into a fast, consistent one.

It also introduces failure modes manual trading doesn't have — silent config drift, stale data, unhandled exchange errors, a process that dies holding a position. Those are engineering problems with engineering solutions, but they are real and they need the reconcile loop described above.

The honest summary: automation is a discipline multiplier and an execution-quality upgrade. It multiplies whatever sign your expectancy already has.

How TradeFloor implements this

TradeFloor runs the loop above across Hyperliquid, Binance and Bybit from one interface. Indicator and order-flow triggers evaluate on a 60-second cadence with up to five AND/OR conditions per rule. Every entry places its bracket in the same operation, and a separate healing process reconciles open positions against live protective orders continuously, re-placing anything the exchange rejected. Liquidation monitoring escalates through four levels, the last of which does nothing unless you explicitly arm that wallet. Backtests charge fees and funding by default. Keys are stored encrypted and the platform never takes custody of funds.

Start with the short definitional overview if this guide is your first exposure, or go straight to setting up a bot on Hyperliquid.

// questions or corrections · [email protected] · more essays · /blog