Backtests That Survive Live Crypto Markets

2026-03-16

Backtests are useful only when they are harder to pass than the live market. For BitBank's public metrics, that means reporting recent performance, separating forecast error from execution performance, and refusing to count a trade unless the historical candle could plausibly have filled it.

The Homepage Window

The headline return, win rate, drawdown, and 1-hour forecast MAE use a trailing 30-day window. The public summary endpoint computes the latest values and caches them for one hour, so the page stays responsive without turning stale historical numbers into permanent marketing copy.

How 1-Hour MAE Is Calculated

MAE is the mean absolute percentage error between the forecasted 1-hour price and the realized price at that target timestamp. If a forecast misses by 0.30% and another misses by 0.70%, their MAE contribution is 0.50%. Direction does not matter for this metric; distance from the realized price does.

How Simulated Fills Work

The backtest uses candle high and low data to decide whether a limit order could have executed. A buy order must be inside the bar by 5 basis points: the candle low must trade at least 5bps below the buy limit. A sell order must also be inside the bar by 5 basis points: the candle high must trade at least 5bps above the sell limit.

This is intentionally stricter than saying "the bar touched the price." Touch-only fills are often too generous because the visible high or low can be a single print with little available liquidity. The 5bp buffer does not model a full order book, but it removes a common source of overly clean simulated performance.

Forecast Skill vs. Trading Skill

The 1-hour MAE answers one question: how close were the forecasts to realized prices? The backtest answers a different question: did the strategy turn those forecasts and signals into executable trades after fees, drawdown, and fill constraints? Both numbers are needed because good forecasts can still produce poor trades if the execution assumptions are weak.

What to Watch

• 30d return: cumulative simulated strategy return over the trailing month
• Win rate: winning closed simulated trades divided by all closed simulated trades
• Max drawdown: largest peak-to-trough decline on the simulated equity path
• 1H MAE: average absolute forecast miss for 1-hour predictions over the same recent window
• Fill buffer: 5bps of required candle penetration before a simulated limit fill counts

No backtest guarantees live performance. The goal is narrower and more useful: make the simulator conservative enough that weak edges fail early.