How the backtest simulator models friction
What the Coin AI engine does and doesn't model when it runs your strategy against historical data. Read this before treating a backtest result as a prediction.
1. Execution and fills
Every trade in a backtest executes at the closing price of the bar that triggered the rule. There is no partial-fill modeling, no queue priority, no order-book depth. If your strategy fires a buy on a 15-minute candle that closed at $97.42, the engine fills you at $97.42 for the full quantity, instantly.
Real-world execution adds delay between the signal and the fill, price movement during that delay, and the possibility of partial fills on illiquid names. None of that appears in the backtest.
2. Commissions and fees
The engine charges zero commission per trade and zero exchange or regulatory fees. Most modern retail brokerages charge $0 commission on US equities, so the gap on stock trades is small. Crypto exchanges and options brokers still charge real fees that will compound across a frequent-trading strategy. If your backtest is a high-turnover system, the commission drag in production may be material.
3. Slippage and spread
Slippage is the gap between the price you expected and the price you got. It comes from bid-ask spreads, order-book impact, and latency. The Coin AI engine assumes a zero spread and zero impact. For a liquid mega-cap like AAPL during regular hours, this is a tolerable simplification. For thinly traded names, small-caps, or after-hours fills, real slippage can be the difference between a profitable strategy and a losing one.
4. Short selling
The engine supports four trade sides: buy (open long), sell (close long), short (open short / sell-to-open), and cover(close short / buy-to-close). When a short trade opens, the engine credits cash by the short proceeds (quantity × price) and records a negative position quantity. When a cover closes the short, the engine debits cash by the cover cost. Realized P&L is (entry price − exit price) × quantity, the mirror of a long.
The engine assumes you can always locate shares to borrow. There is no hard-to-borrow restriction, no fail-to-deliver modeling, and no preborrow process. In real markets, some tickers are uncovered at any price, or available only at punishing borrow rates, especially small-caps, meme stocks, and recent IPOs. A backtest that profits by shorting names with no real shortable supply is fiction.
5. Borrow costs and locate
Brokers charge interest on borrowed shares for the duration of a short position. The rate varies from a few basis points annually on liquid large-caps to 50%+ APR on hard-to-borrow names. The Coin AI engine charges zero borrow fee. A short position that runs for months in the backtest pays nothing for the privilege.
The practical effect: short strategies look more profitable in the backtest than they will be in production. For a strategy that holds shorts for an average of 10 trading days, a 5% APR borrow rate would shave roughly 5% × (10 / 252)≈ 0.2 percentage points off each trade, small per-trade but compounding across the run.
6. Margin and the cash floor
Real brokers monitor your account's equity against maintenance margin requirements (typically around 30% of the short position's notional). When equity falls below that threshold, the broker issues a margin call and may force-close positions at the market.
The Coin AI engine does not model maintenance margin. It applies a single guardrail: if total portfolio equity (cash plus mark-to-market value of holdings) would fall below zero, the engine force-closes every open short position at the bar's closing price. These force-closes appear in the trade log as margin events (distinct from rule-fired covers and stop-loss force-closes). In real markets, the call would have come earlier and at a worse price.
7. Crypto and options
Crypto shorts in the backtest assume access to a perpetual futures venue (Hyperliquid-style) where you can short spot-equivalent exposure at no funding cost. Most retail spot exchanges (Coinbase, Kraken, Binance.US) do not allow shorting. If you build a short-crypto strategy on Coin AI and the result looks promising, you need a venue that actually supports the position, and you need to model the funding rate yourself.
Options shorts in the backtest are naked; the engine treats a short call or short put as a single-leg position with unlimited (calls) or strike-bounded (puts) downside. There is no assignment risk modeling, no early-exercise logic on American options, and no margin requirement tracking specific to options. Use the option_pricer output as a directional approximation, not a position that you should run un-hedged.
8. Historical data
Equity prices come from Polygon.io, adjusted for splits at load time. Dividend reinvestment is not modeled. Crypto prices come from CoinGecko. Intraday backtests fetch 1-minute source bars and resample; long intraday ranges can hit the 120-second engine timeout (Coin AI will warn you in chat).
Survivorship bias is real: the engine only knows about tickers that exist today. A 2018 backtest of a small-cap strategy will not include companies that delisted before 2026. Treat results on delisted names as unknowable.
9. Bottom line
The Coin AI backtest engine is a research tool. Its purpose is to tell you whether a rule-based strategy would have been profitable in the past, ignoring friction. It is not a guarantee of future performance, and the friction it ignores is real money in production.
Use backtest results to compare strategies against each other, to size up the magnitude of the edge, and to stress-test the rules. Do not use them as a forecast of live trading P&L.
Questions or want a specific friction modeled? Email [email protected].