Phase I: We tested 220 volatility-envelope breakouts. Zero survived.

We added a new family of candidates to the research lab this week. We hypothesized that Bollinger Band envelope breakouts might survive our graduation gates where simpler trend-following entries had not, because the BB envelope expands with recent volatility. A close that breaks above the upper band only happens when the move is large relative to recent realized volatility, which implicitly carries a "volatility just expanded" confirmation that fixed-window breakouts (Donchian high crosses) do not.

We ran 220 candidates across 7 distinct entry shapes, 2 timeframes, 4 profit targets, and randomized bar caps. Every candidate was tested on 10 coins across 3 years of Binance data with real fees and per-coin L2 spread sampling. None of them graduated.

What we tested

Seven entry shapes covering raw breakouts, ADX-filtered breakouts, trend-filtered breakouts, multi-window confluence breakouts, and the short-side mirror of each:

• Raw Bollinger upper break: close crosses_above bb_upper_20
• BB upper break + ADX above 25 (trend confirmation)
• BB upper break + ADX above 30 (stricter trend confirmation)
• BB upper break + close above 200 EMA (long-term trend filter)
• BB upper break AND close above 50-bar Donchian high (multi-window confluence)
• BB lower break + ADX above 25 (short side, trend confirmation)
• BB lower break + close below 200 EMA (short side, trend filter)

Tested on 1-hour and 4-hour bars. Profit targets between 1.5% and 3.0%. Bar caps between 60 and 400. Total: 220 unique configurations.

What "failed our gates" means

Each candidate had to pass all six of the following to graduate to the public survivor registry:

1. Walk-forward validation P&L positive (out-of-sample test on data the optimizer never saw)
2. Win rate at or above 30%
3. Sharpe ratio above 0.5
4. Deflated Sharpe fluke probability below 50% (adjusted for the number of candidates tested, the Bailey-López de Prado correction)
5. Edge consistency across at least half the qualifying coins (not just a one-coin fluke)
6. Frozen holdout window survival (a reserved tail of data the walk-forward train and validation never saw)

220 of 220 failed at least one gate. Most failed multiple. The single most common failure mode was signal density: BB envelope breakouts fire too rarely on calmer regimes to produce statistically significant trade counts in our walk-forward windows, even when the edge per trade was positive.

What this does and does not mean

It does not mean Bollinger Band breakouts never work. It means our specific encoding of them, on the specific 3-year window of 10-coin Binance spot data we tested, did not survive a deliberately strict graduation stack designed to weed out overfitting.

The result is also constrained by what our primitive system can express. Our condition format compares an indicator to a value or another indicator. We cannot express stateful patterns like "BB width was in the bottom 30th percentile of the last 100 bars and is now expanding" without a richer condition language. The "squeeze then expansion" pattern that traders commonly associate with volatility breakouts is closer to that stateful form than to what our 220 candidates tested.

If anything in this family does eventually graduate, it will likely come from the multi-timeframe Phase L family, which adds a higher-timeframe trend filter that the single-timeframe Phase I candidates couldn't carry. Phase L is currently queued behind J and K in the lab.

Where the survivors are

So far the lab's graduates remain concentrated in mean-reversion patterns within ranging, low-volatility regimes. RSI-based oversold entries with SMA-200 trend filters and ADX chop filters are the shape that has survived the gates most often. You can browse the profiles at /survivors or read the long-form writeups from the /research hub.

If you want to test a volatility-breakout idea of your own with different parameters or a different envelope, the same 6-gate backtest is available free at /prove. Verdict in under 2 minutes.