The thesis
Bollinger-reversion is a textbook mean-reversion play: buy when price closes below the lower Bollinger band, sell when it pushes above the upper band. It runs on a 24-name universe of US large caps spanning tech (AAPL, MSFT, GOOGL, NVDA), financials (JPM, BAC, V, MA), staples (PG, KO, WMT, COST), healthcare (JNJ, UNH, PFE, ABBV) and industrials/energy (CAT, HON, XOM, CVX). The logic is simple and the universe is liquid — exactly the setting where band-reversion is supposed to behave.
Backtest at a glance
Over 451 days the strategy returned 17.55%, finishing at $11,755.25 for a 9.46% CAGR. The win rate is a healthy 63.89% across 76 trades, and total fees were a negligible $76 with no FX cost. Those are attractive headline numbers.
The texture underneath is less flattering. The Sharpe ratio is only 0.66 — modest for a strategy carrying a 20.57% maximum drawdown. Turnover is a heavy 1,720.89%, meaning the book churns many times over; the low absolute fee load is a function of the small account, not of restraint. A high win rate paired with a middling Sharpe and a deep drawdown is the classic signature of many small wins offset by occasional large losers.
Validation: the red flag
This is where the strategy stumbles. Our walk-forward validation fails. All four out-of-sample folds are positive, which sounds reassuring, but the returns decay monotonically: 7.67% (Sharpe 1.78) → 1.42% (0.25) → 0.77% (0.20) → 0.40% (0.14). The most recent fold barely clears zero. The full-sample 17.55% is essentially front-loaded into the first window.
The statistics agree. Aggregate OOS return is 0.40% at a 0.14 Sharpe. The Probabilistic Sharpe Ratio of 0.814 looks decent, but the Deflated Sharpe Ratio — which penalises the 6 trials run — falls to 0.342, below the threshold we require. Translation: once we account for how many variants were tried, the edge is not statistically distinguishable from luck.
Recent live activity
Live behaviour echoes the concern. Across scheduled runs from 2026-09-11 to 2026-09-18, most orders were rejected, not executed — five of six runs executed zero trades. The lone active day, 2026-09-18, executed a NKE buy (70 @ $35.64) and a GOOGL sell (6 @ $351.91). Notably that GOOGL lot was bought at $377.90 on 2026-06-01, so it was closed at a loss. The portfolio total has drifted around $9.65k–$9.76k, sitting below a $10k starting line.
Verdict
Strengths are real: a coherent thesis, a high hit rate, low costs and uniformly positive folds. But the risks dominate. The edge is fading in time, the deflated statistics fail our gate, drawdowns are deep, and live execution has been dominated by rejections with a shrinking balance. Bollinger-reversion earns continued observation — not fresh capital — until it shows the recent folds can turn back up.