With futures, leverage is part of the product
A futures trading bot opens, adjusts and closes derivative positions under a set of rules. It can go long or short without owning the underlying asset, but every order comes with the contract’s margin, liquidation, funding, settlement and liquidity mechanics. Automating the orders makes them repeatable; it does not soften the leverage.
An AI futures bot simply gives a model one or more jobs, such as classifying the market, forecasting a return or scheduling an order. Hard limits should remain the same whether the signal comes from a moving average, a statistical model or an AI agent. A confidence score is not usable margin or available liquidity.
An expiry date changes the job of the bot
| Contract | How it stays connected to spot | Operating consequence |
|---|---|---|
| Dated future | Has an expiry or settlement date, creating convergence toward settlement. | The bot must manage expiry, settlement and any roll to a later contract. |
| Perpetual future or swap | Has no scheduled expiry; a funding mechanism helps align the contract with its index. | The bot must model recurring funding and changes in the contract's basis. |
Names, funding intervals and settlement mechanics vary by venue and location. The bot needs the specification for the exact contract in front of it; copying one exchange’s assumptions to another is a straightforward way to misprice risk.
What happens between signal and settlement
The signal model appears in only one step. The more dangerous failures often happen around it: a stale price enters the system, size is calculated from the wrong balance, one hedge leg fills alone, or the connection dies while the position stays open. See the broader automated trading systems guide for the full control path.
Four numbers decide whether the account stays alive
Notional exposure
Notional is the market value controlled by the position. For a simple linear contract, approximate it by multiplying quantity, contract size and price. Inverse and quanto contracts pay out differently, so use the venue’s formula rather than forcing them into the linear shortcut.
Initial and maintenance margin
Initial margin gets the position open. Maintenance margin is what keeps it there. The remaining buffer also moves with unrealized P&L, fees, funding, other positions, collateral haircuts and whether margin is isolated or shared.
Funding and basis
Perpetual funding is a periodic transfer between long and short positions. Its direction and size can change, and the interval differs by venue. Basis is the difference between the derivative and its spot or index reference. A directional trade can be correct and still underperform after funding; a hedged trade can lose when basis or leg sizes diverge.
Executable liquidity
Index price, mark price, last trade and executable price are not interchangeable. A backtest may value a position at a clean reference price even though a real exit would sweep several worse levels. Risk needs a stressed price for the full size, not the best quote on the screen.
A leverage example that does not pretend to know liquidation
Suppose an account commits $1,000 of equity and controls $5,000 of linear notional exposure. A 5% adverse price move represents a $250 mark-to-market loss before fees and funding, or 25% of starting equity. A 2% adverse move represents about $100, or 10% of starting equity.
Approximate equity impact = price move × position notional ÷ starting equity This shows leverage, not a liquidation price. Actual liquidation depends on the venue’s mark price, maintenance tiers, account mode, collateral, fees and every other open position. A calculator that reduces the answer to “entry price divided by leverage” is leaving out material account state.
Funding can switch from income to expense
For a simple linear position, an interval's funding payment is commonly based on position notional multiplied by the applicable funding rate. At $10,000 notional, a 0.01% interval rate corresponds to a $1 transfer before other costs. Whether that transfer is paid or received depends on the position side and the sign specified by the venue.
Illustrative funding transfer = position notional × interval funding rate Annualizing one favorable interval produces a tidy but fragile number. Rates can compress, reverse or fail to cover trading costs and basis movement. The dedicated funding-rate arbitrage guide explains the hedged structure and why “market neutral” does not mean risk free.
Useful jobs for a model—and decisions it should not own
| Task | Possible model role | Required guardrail |
|---|---|---|
| Market classification | Estimate regime, volatility or directional probability. | Reject stale inputs and cap the lifetime of every prediction. |
| Position proposal | Suggest direction, target exposure or hedge ratio. | A deterministic policy engine applies final size and leverage limits. |
| Order execution | Select timing or split a parent order. | Price collars, slippage limits and cancel conditions bound the action. |
| Monitoring | Flag unusual behavior or summarize account state. | Venue balances, fills and positions remain the source of truth. |
Do not let an unconstrained model invent symbols, raise leverage, bypass a loss limit or interpret silence as permission to continue. If the model is uncertain or offline, the fallback should be a known state such as “no new risk,” never an improvised order.
Different strategies break in different ways
| Approach | Working hypothesis | Dominant failure |
|---|---|---|
| Directional long/short | A signal forecasts direction or relative strength. | Forecast error amplified by leverage and costs. |
| Trend following | Persistent movement continues long enough to exceed whipsaws and costs. | Repeated reversals, late exits and crowded positioning. |
| Grid or mean reversion | Price repeatedly crosses levels inside a bounded range. | A persistent trend accumulates exposure against the move. |
| Spot–perpetual carry | Funding or basis exceeds the cost of offsetting directional exposure. | Funding reversal, basis widening, liquidation or one-leg failure. |
| Cross-venue hedge | Related contracts can be kept economically offset. | Collateral fragmentation, venue failure and transfer constraints. |
Controls to prove before the first live order
- Instrument allowlist: exact venue, symbol, contract type, collateral and settlement asset.
- Exposure limits: maximum notional, leverage, margin utilization and asset concentration.
- Order limits: maximum size, slippage, price deviation, retries and outstanding orders.
- Liquidation monitoring: venue-specific calculation with warning and reduce-only thresholds.
- Funding controls: current and projected rate checks plus a maximum acceptable carry cost.
- Data freshness: reject delayed books, indices, positions, balances and model outputs.
- Execution recovery: reconcile partial fills, duplicate requests and uncertain acknowledgements.
- Account reconciliation: compare local state with venue orders, fills, positions and collateral.
- Independent stop path: cancel, reduce-only and emergency-close controls that do not depend on the signal model.
Make the bot fail on purpose before the venue does it for you
A profitable history says little about whether the bot can operate safely. Tests should deliberately feed it delayed data, rejected orders, partial fills, duplicate requests, changing funding, abrupt volatility, API downtime and a restart while exposure is still open.
- Recalculate results after commissions, bid–ask spread, slippage, funding and failed execution.
- Separate in-sample development from untouched validation and forward observation.
- Stress price gaps and reduced market depth instead of assuming every stop fills at its trigger.
- Confirm that restarting the bot cannot duplicate an order or forget an existing position.
- Verify that a breached risk limit prevents new exposure even if the signal still says “buy.”
Compare simulated evidence with the backtests versus live results research note before treating a performance chart as proof.
Six checks before handing a bot credentials
- Identify the product. Is it software you operate, a signal service, a managed strategy or a venue feature?
- Trace custody and permissions. Determine who holds assets and whether credentials allow withdrawals.
- Inspect contract support. Confirm every venue, margin mode, settlement asset and funding convention.
- Read the risk policy. Look for hard limits, stale-data behavior, liquidation warnings and an independent shutdown path.
- Demand results after every cost. Returns should include fees, spread, slippage, funding and periods when the system was unavailable.
- Verify claims. Treat guaranteed returns, unexplained win rates and “AI predicts the market” as warning signs.
When automation only makes the mistake faster
If the strategy cannot be explained without marketing language, it is not ready for automation. A bot is also a bad fit when the account cannot absorb loss, nobody can monitor it, contract rules are unclear or recovery has never been tested. Paper trading does not solve model risk, but it often exposes broken state, sizing and order handling before money is involved.
Primary references for contract mechanics
- CFTC — Understand the risks of virtual currency trading
- CFTC — Economic purpose of futures markets and how they work
- CFTC — AI won't turn trading bots into money machines
- Coinbase International Exchange — Funding-rate mechanics
- Onchain Off Emotion — Risk disclosure
This is an explanation of mechanics and operating risk, not a recommendation to trade derivatives. Contract access, margin, liquidation, tax and regulation vary by venue and location. Read the current documentation for the exact product rather than assuming these examples describe every contract.