Custom Squeeze-Release Short Strategy — Marginal / Overfit Warning

Verdict: marginal · Asset/TF: BTCUSDT 1h · Sample: ~1 year of 1h data (2025 - 2026)

Hypothesis

Short-only strategy exploiting sudden volatility squeeze releases. We enter short positions when the custom indicator squeeze_release_short exceeds 0.5, indicating a high-probability downward break. We exit when the return z-score of the last 24h exceeds 0.2 (ret_z_24 > 0.2) or when the local entropy of the past 24 bars is high (entropy_24 > 0.72), indicating that the downward momentum has decayed into noisy consolidation. Tight risk controls are used: 0.75% Stop Loss, 1.75% Take Profit, and a maximum holding period of 10 bars.

Strategy

{
  "entry_rules": [
    {
      "condition": "squeeze_release_short > 0.5",
      "direction": -1,
      "signal": "squeeze_release_short"
    }
  ],
  "exit_rules": [
    {
      "condition": "ret_z_24 > 0.2 || entropy_24 > 0.72",
      "reason": "release_failed_or_noise"
    }
  ],
  "max_hold_bars": 10,
  "position_size": 0.3,
  "stop_loss_pct": 0.75,
  "take_profit_pct": 1.75
}

Backtest

Robustness (the proof — do not skip)

• Walk-Forward: WFE = 0.35 → Overfitting / Parameter Decay Warning. A WFE of 0.35 indicates that the out-of-sample performance is only 35% of the in-sample performance. While 6 out of 7 out-of-sample windows were positive, the drop in performance suggests severe overfitting to the historical parameters.
• Monte-Carlo: risk-of-ruin = 0.0%; probability of loss = 5.8%; median return = 10.37% (500 iterations).
• Sensitivity: The entry threshold for squeeze_release_short is highly sensitive, and small changes lead to rapid decay in Sharpe.

Research trail

Tools called: load_dataset → ai_run_backtest → walk_forward → monte_carlo

• What I tried: We implemented custom squeeze-release indicators with varying exit thresholds (ret_z_24 and entropy_24) and holding periods.
• What failed: Attempting to extend the holding period beyond 10 bars led to rapid drawdown increases as short positions got caught in macro trend reversals.
• What I learned: While custom feature models can produce very high backtest Sharpe ratios (1.64), they are highly prone to parameter decay. A WFE of 0.35 means this strategy should NOT be traded in production without dynamic parameter re-optimization or additional macro trend filters.

Reproduce

Dataset: /Users/serg/projects/trading/rlxbt_custom_features_BTCUSDT_1h.csv. Strategy JSON above. Re-run the same tools in the RLXBT app.