The In-Sample Trap: A Bollinger-Band Mean-Reversion That Dies Out-of-Sample (BTCUSDT 1h)

Hypothesis

Price closing outside its Bollinger Bands over-extends and reverts toward the mean. So we fade the break: go long when price closes below the lower band, short when it closes above the upper band.

Strategy (RLXBT JSON rules)

{
  "entry_rules": [
    {"condition": "close < BB_Lower", "direction": 1,  "signal": "fade_lower"},
    {"condition": "close > BB_Upper", "direction": -1, "signal": "fade_upper"}
  ],
  "exit_rules": [{"condition": "close > BB_Upper", "reason": "reverted"}],
  "max_hold_bars": 72, "stop_loss_pct": 3.0, "take_profit_pct": 5.0
}

Data: BTCUSDT 1h, ~33.5k bars. Commission 4 bps, slippage 1 bp.

In-sample result (the seductive part)

A positive Sharpe over more than a thousand trades feels like signal, not noise. This is exactly where most backtests stop — and exactly where they mislead.

Out-of-sample: walk-forward analysis

Re-fitting across rolling train/test windows tells a different story:

A negative WFE means the strategy performs worse out-of-sample than in-sample — the edge is an artifact of the fitting window, not a durable market inefficiency.

Verdict: Rejected (as-is)

The in-sample numbers are a mirage. On BTCUSDT 1h, naive band-fading has no out-of-sample edge — likely because in a trending crypto regime, a close beyond the band is often continuation, not exhaustion.

The takeaway

A believable win rate and a positive Sharpe over 1,000+ trades are not evidence of an edge. Walk-forward (or Monte-Carlo) is the cheapest insurance in quant research. Next iterations worth testing: a volatility/regime filter (only fade in low-BB_Width ranges), a mean exit at BB_Middle, and a higher timeframe.

Generated and validated with RLXBT (backtest + walk-forward).