The chart moved up. The real work starts here.
A 30-day BTCUSDT perpetual-futures simulation returned 2.37% with a maximum drawdown of 1.10%. The run also produced a 3.63 profit factor and a 3.80 Sharpe ratio.
These numbers look encouraging in a dashboard. They are also based on only 14 closed positions.
That gap—between an attractive result and sufficient evidence—is where most backtesting mistakes begin.
This edition breaks down what the result actually shows, what it leaves unanswered, and which tests should come next.
The experiment in one minute
The strategy used a simple RSI and moving-average structure:
• Market: Binance BTCUSDT perpetual
• Bar interval: 1 hour
• Test window: October 10–November 9, 2026
• Starting capital: 100,000 USDT
• Leverage: 1×
• Position size: approximately 30% of capital
• RSI: 14 periods, with 30/70 thresholds
• SMA: 60 periods
• Trading fees: 0.02% maker and 0.05% taker
• Funding payments: not modeled
• Additional slippage: not modeled
The basic mechanism looks for short-term RSI extremes while using the SMA as a lightweight regime filter. The intent is to avoid treating every oversold or overbought reading as equivalent.
What happened
Across the 30-day window, the simulation recorded:
• Return: 2.37%
• Realized P&L: 2,372.39 USDT
• Maximum drawdown: 1.10%
• Win rate: 50%
• Profit factor: 3.63
• Sharpe ratio: 3.80
• Closed positions: 14
• Orders: 28
The 50% win rate is a useful reminder that a strategy does not need to win most trades to finish positive. The relationship between average wins, average losses, and trade sequencing matters more than win rate alone.
But the small sample dominates every interpretation. Fourteen positions are not enough to determine whether the result came from a durable mechanism, a favorable month, or a few unusually helpful reversals.
The metric most likely to mislead
The Sharpe ratio of 3.80 is the easiest figure to overread.
Sharpe is often treated as a compact answer to the question, “Was the return efficient relative to risk?” In a short test with few trades, however, the estimate can move sharply because of one or two outcomes. It describes this sample; it does not establish the stability of the process.
The more responsible interpretation is not “high risk-adjusted performance.” It is “a result that deserves a larger sample and harsher assumptions.”
What the backtest still does not know
Two execution costs remain outside the model: funding payments and additional slippage.
For a perpetual-futures strategy, funding can accumulate differently across long and short exposure. Slippage can also become more important around fast price moves, exactly when an RSI-based strategy may be most active.
The run therefore answers a narrower question:
How did this fixed rule set behave over this historical window, using the modeled trading fees and the stated execution assumptions?
It does not answer whether the same behavior would survive live execution.
Six tests that matter more than another optimization
- Expand the market regimes
The same rules should be tested across trending, ranging, low-volatility, and high-volatility periods. A mean-reversion system can look convincing when the market repeatedly snaps back and fail when price moves persistently in one direction.
- Add funding and realistic slippage
The next simulation should include funding and a slippage model tied to volatility or order type. A result that disappears under modest execution friction was never robust enough to deploy.
- Reserve untouched data
Out-of-sample testing and walk-forward validation help separate discovery from confirmation. Parameters selected on one window should be evaluated on data that played no role in choosing them.
- Stress the parameters
RSI thresholds and SMA lengths should be varied across sensible neighborhoods. If performance collapses after a minor parameter change, the apparent edge may be a narrow fit rather than a stable mechanism.
- Compare adjacent markets
Testing the same logic on other liquid perpetual markets can show whether the behavior is broadly connected to the mechanism or unusually dependent on BTC during this period.
- Audit every fill
Summary statistics should not replace trade-level review. Entries, exits, reversals, fees, timestamps, and position transitions must match the intended rules. A clean equity curve cannot compensate for an implementation error.
The larger lesson
A positive backtest is not a verdict. It is a hypothesis that has survived one particular test.
The useful role of AI-assisted research is not to make that hypothesis sound more certain. It is to shorten the loop between idea, implementation, inspection, and falsification.
AlgoRover.ai was used to run and inspect this simulation. The tool can be installed in Codex, Cursor, or Claude Code to run and review the same type of backtest inside an AI coding workflow.
The next edition can compare the 30-day result with longer windows, realistic execution costs, and parameter-sensitivity results.
Which test should come first: market regimes, execution costs, or parameter stability?
Historical simulation only. Not a forecast, investment recommendation, or evidence of future performance.
Top comments (0)