A correct forecast can still lose money
AI scalping gives a model some role in trades held for seconds or minutes. It might estimate the next move, decide whether current conditions are tradable or choose how to place an order. None of that makes the fill profitable. The captured move still has to beat the spread, fees, slippage and the cost of delay.
Win rate and prediction accuracy therefore tell only part of the story. The number that matters is net profit after execution costs, measured at the same order size, under the same venue rules and with the same latency the live bot will face.
The order book is part of the strategy
A centralized crypto exchange usually matches buyers and sellers in a limit order book. The best bid is the highest displayed buy price; the best ask is the lowest displayed sell price. The gap between them is the spread. A marketable order crosses it and takes liquidity. A passive limit order waits in the book and may earn a better price—or never trade at all.
The top quote applies only to the quantity shown there. A larger order may consume several levels at progressively worse prices. That difference between the screen price and the full-size fill is slippage.
| Microstructure variable | Question the system must answer | Why it changes the result |
|---|---|---|
| Spread | How far apart are the executable bid and ask? | Crossing the book creates an immediate cost before the market moves. |
| Depth | How much size is available near the best prices? | A larger order may fill across worse levels than the displayed quote. |
| Trade flow | Are aggressive buys or sells consuming one side of the book? | Recent pressure can contain information, but it can also reverse quickly. |
| Queue position | How many same-price orders are likely ahead of a resting limit order? | Being at the best price does not mean being next to fill. |
| Volatility | How far can price move before an order is accepted or canceled? | A signal and its safe order price can expire in milliseconds or seconds. |
Matching rules differ by venue and product. Coinbase Exchange, for example, uses price-time priority: a better price goes first, and an earlier order beats a later order at the same price. Candle data cannot tell whether a passive order reached the front of that queue. A useful simulation needs order-book and trade data.
Work backward from break-even
One basis point (bp) is 0.01%. Put the forecast and every cost in basis points so different trade sizes can be compared on the same scale. A first-pass estimate looks like this:
Expected net edge = expected gross move − spread cost − entry fee − exit fee − slippage − latency loss The spread term depends on how the gross move is measured and how orders execute. If a forecast measures a mid-price-to-mid-price move while the bot buys at the ask and later sells at the bid, the round trip pays approximately one full spread. Passive fills may avoid crossing, but they introduce fill uncertainty, queue risk and the chance that only unattractive orders execute.
A worked trade example
Consider an illustrative $10,000 spot trade. The model expects a 15 bp favorable mid-price move over its holding horizon. The observed spread is 3 bp. The hypothetical taker fee is 4 bp on entry and 4 bp on exit, and expected slippage is 1 bp per side.
| Item | Basis points | Dollar effect on $10,000 |
|---|---|---|
| Forecast gross move | +15 bp | +$15 |
| Round-trip spread | −3 bp | −$3 |
| Entry and exit fees | −8 bp | −$8 |
| Entry and exit slippage | −2 bp | −$2 |
| Expected net | +2 bp | +$2 |
Fifteen basis points sounds meaningful over a short holding period, yet only $2 survives before taxes, failed orders, infrastructure or extra delay. A 3 bp worse exit turns the trade into a $1 loss. These are example inputs, not current fees from any venue. Actual maker and taker charges vary by account tier and change over time, so the bot should load them from the venue and record them with every fill.
The price can change while the bot is still thinking
The bot receives market data, calculates features, runs the model, sends an order and waits for the venue. Every hop takes time. Measure market-data transport, feature calculation, model inference, order transmission and exchange acknowledgement separately. Quoting only the model’s inference time hides most of the delay.
For a marketable order, delay can turn an acceptable quote into a worse fill or a rejection under a price collar. For a passive order, delay can place the bot behind more orders at the same price. Canceling and reposting may reset queue priority. The system should therefore record exchange timestamps when available, local monotonic timestamps, sequence gaps, order acknowledgements, fills and cancels—not just the time at which the strategy produced a signal.
- Reject a signal after its defined lifetime rather than trading a stale prediction.
- Stop or resynchronize when the market-data sequence has a gap.
- Model partial fills and the risk of holding only part of the intended position.
- Measure fill probability and time-to-fill for passive orders by venue, symbol and regime.
- Compare decision price, arrival price, fill price and exit price for every trade.
Where a model may earn its keep
| Possible model role | Useful output | Boundary |
|---|---|---|
| Short-horizon forecasting | A probability distribution for price movement over a named horizon. | Prediction accuracy is not net profitability after costs. |
| Regime classification | An estimate of whether spread, depth or volatility conditions suit the strategy. | Regimes can change faster than a retraining schedule. |
| Execution policy | A choice between waiting, posting, crossing or splitting an order. | Hard price, size and exposure limits still belong outside the model. |
| Anomaly detection | A warning about unusual data, fills, latency or model behavior. | The alert needs a deterministic response such as pause, cancel or reduce. |
| Feature extraction | A compact representation of order-book changes, trades and cross-market signals. | A model cannot recover information absent from a delayed or incomplete feed. |
A model cannot create liquidity, jump the queue, remove fees or make an unreliable venue reliable. It also should not have unlimited control over symbols, size or losses. NIST’s AI Risk Management Framework calls for testing before and during deployment, measuring performance in the real setting and providing a way to disengage systems that behave outside their intended use. That matters when model output can move money.
Why a beautiful scalping backtest falls apart
| Failure mode | What creates the illusion | What reveals it |
|---|---|---|
| Cost blindness | A backtest uses mid-prices or excludes one side of fees. | Reprice every simulated entry and exit at executable quotes with the correct fee tier. |
| Impossible passive fills | Every touched limit order is counted as filled. | Use trade flow, queue assumptions and conservative partial-fill rules. |
| Look-ahead leakage | Features or labels contain information not available at decision time. | Reconstruct the exact event timeline and delay features until they were observable. |
| Overfitting | Many symbols, horizons and parameters are tried until one looks exceptional. | Keep untouched test periods and report all trials, not only the winner. |
| Regime dependence | The sample contains unusually tight spreads or one-directional movement. | Segment results by volatility, liquidity, spread and market direction. |
| Capacity collapse | The test assumes size has no effect on fill price. | Repeat results across order sizes using depth-aware execution. |
| Stale market state | The bot keeps trading through feed gaps or clock drift. | Inject delays and disconnects, then confirm that new risk is blocked. |
| Adverse selection | Passive orders fill most often just before price moves against them. | Measure post-fill markouts at several horizons, not maker fees alone. |
Test at the speed and depth the bot will trade
Candles are usually too coarse when the result depends on spread, queue position and order flow inside the bar. Test data should match the bot’s decision speed and preserve the timestamps and market state needed to reconstruct what was actually knowable.
- Define the decision. State the asset, venue, holding horizon, allowed order types and exact conditions for entry, exit and no trade.
- Build an event timeline. Reconstruct book updates, trades and order events in their observable order; detect gaps instead of silently continuing.
- Use an execution model. Apply spread, tier-specific fees, latency, size-aware slippage, partial fills, cancels and conservative queue assumptions.
- Separate development from evaluation. Fit and tune on one period, validate on later untouched periods, then use walk-forward tests across different regimes.
- Run counterfactual costs. Increase fees, latency and slippage and reduce fill probability. A 2 bp edge that vanishes under a 1 bp change is fragile.
- Shadow the live feed. Generate decisions without sending orders and compare simulated arrival prices with real-time executable quotes.
- Use limited live exposure. If legally and financially appropriate, start below intended size and compare actual fills with the simulation before scaling.
Report trade count, average and median net return, dispersion, drawdown, turnover, fill and cancel rates, slippage, post-fill markouts and results by market regime. One rising equity curve hides many ways a short-horizon system can fail. The backtests versus live results guide covers the broader evidence gap between historical simulation and production trading.
Live controls should work without the model
FINRA wrote its algorithmic-trading guidance for regulated securities firms, not every crypto account. The engineering lessons still travel well: test before production, control changes, watch live behavior, reconcile the account and keep a fast way to stop the algorithm. The exact controls depend on the venue, location and permissions involved.
- Instrument allowlist: permit only reviewed venues, markets and order types.
- Pre-trade limits: cap order size, price deviation, open orders, position and daily turnover.
- Loss limits: block new risk after per-trade, rolling and daily loss thresholds.
- Freshness gates: stop on stale books, missing sequences, clock drift or delayed model output.
- Rate-limit handling: back off safely without duplicating orders or assuming a timeout means rejection.
- State reconciliation: compare local orders and positions with the venue's acknowledged state.
- Self-trade protection: prevent the system's own orders from trading with one another.
- Independent kill switch: cancel outstanding orders and prevent new ones without relying on the model.
- Credential controls: use the least permissions required and disable withdrawals where the venue supports it.
- Change management: version code, models and configuration; require a new validation whenever any of them changes.
Seven questions for an AI scalping service
- Ask what “AI” does. A provider should identify whether the model predicts, classifies, executes or monitors—not use AI as a substitute for an explanation.
- Request results after every trading cost. Look for spread, maker and taker fees, slippage, rejected orders and downtime in the reported return.
- Inspect the sample. Check the time period, venues, assets, trade count and whether results include quiet, volatile and falling markets.
- Separate backtest, paper and live evidence. Each answers a different question; simulated fills should never be presented as live executions.
- Check custody and permissions. Determine where assets sit, who can move them and what happens if the service disappears.
- Review controls. Confirm position limits, loss limits, stale-data behavior, monitoring and a shutdown procedure.
- Reject guaranteed returns. The CFTC warns that there is no guaranteed trading strategy and advises understanding how a strategy can lose money.
Questions people ask about AI scalping
Is AI scalping profitable?
Profitability only makes sense for a named model, market, venue, period and trade size. A good historical result does not prove a lasting live edge. It should survive realistic costs, later untouched periods and comparison with actual live fills.
Is a crypto scalping bot the same as high-frequency trading?
No. Both can trade over short horizons, but professional high-frequency trading commonly depends on specialized infrastructure, direct market connectivity and strict latency engineering. A retail bot that checks a public API every second may be automated and short-term without being competitive HFT.
Are maker orders always cheaper?
Maker and taker fees often differ, but a lower fee does not make a passive order better by itself. The order may not fill, may fill only partly or may fill just before the market moves against it. Compare total execution quality—including missed trades and post-fill price movement—not the fee label alone.
What is the best time frame for AI scalping?
There is no universal best interval. The horizon must be long enough for the expected move to exceed costs and short enough for the signal to remain useful. Test horizons using event-time data and executable prices rather than choosing a candle interval because it produces the most attractive backtest.
When to choose a slower horizon
Live scalping is a bad fit when fills cannot be reconstructed, market data is unreliable, end-to-end latency is unknown or nobody can supervise the system. It is also a warning sign when fees are large next to the target move or leverage is needed to make a negligible edge look meaningful. A slower trend-following approach or a broader algorithmic trading system may reduce execution pressure, but each still needs its own evidence and controls.
Technical and regulatory references
- Coinbase Developer Documentation — Exchange matching engine and price-time priority
- Coinbase Help — Maker, taker and partially matched order fees
- FINRA Regulatory Notice 15-09 — Supervision and control practices for algorithmic trading
- NIST AI RMF Core — Testing, monitoring and disengagement of AI systems
- CFTC — Understand the risks of virtual currency trading
- Onchain Off Emotion — Risk disclosure
This guide explains mechanics, testing and operating risk. It does not claim that AI scalping is profitable or recommend a bot. Fees, matching rules, product access and regulation vary by venue and location, so use the current documentation for the exact market and account.