The thesis
Bollinger-reversion is a classic mean-reversion play: buy when price closes below the lower Bollinger band, sell when it pushes above the upper band. The premise is that short-term deviations from a moving-average envelope tend to snap back. It runs live across a 24-name universe of large-cap US equities spanning tech (AAPL, MSFT, NVDA), financials (JPM, V, MA), healthcare (JNJ, UNH), and staples (PG, KO, WMT, COST).
The backtest looks respectable
Over 451 days the strategy returned 17.55% (final equity $11,755 on a $10k base), a 9.46% CAGR, with a 63.89% win rate across 76 trades. Those are numbers a trader would happily take. But the supporting metrics temper the enthusiasm: a Sharpe of 0.66 is modest, and a 20.57% max drawdown means the ride to that return was bumpy. Turnover of 1,720% also signals heavy churn — the strategy is trading constantly, and while fees here were a trivial $76, that level of activity is fragile to slippage in the real world.
Why validation failed
This is where the strategy earns its skeptical read. Walk-forward testing split the sample into 4 folds, and all 4 were positive — encouraging on the surface. The problem is the trajectory: fold returns decay steadily from 7.67% (Sharpe 1.78) to 1.42%, 0.77%, and finally 0.40% (Sharpe 0.14) in the most recent window. The out-of-sample Sharpe of 0.14 is a shadow of the in-sample figure.
The deflated metrics confirm it. With 6 trials searched, the probabilistic Sharpe ratio holds at 0.814, but the deflated Sharpe ratio drops to 0.342 — below the threshold — and the validation gate returns failed. In plain terms: much of the headline return came from one strong early period, and the edge has been eroding since. The recent-fold behaviour looks a lot more like noise than a durable signal.
Recent live activity
The live paper book tells a consistent story of stagnation. Across scheduled runs from July 6–13, the strategy executed zero trades, with a handful rejected each day, and total equity drifting around $9,700–$9,900 — below its starting base. The last executed trades were in late May and June: buys in DIS, GOOGL, PG, WMT, and COST, and a PG sell. Since then, few names have breached the bands enough to trigger entries.
Verdict
Strengths: a clean, interpretable thesis, a high win rate, and four positive out-of-sample folds. Risks: a failed validation gate, a decaying edge, a 20%+ drawdown, and a week of dormancy. Bollinger-reversion is a reasonable idea whose statistical support is thinner than its backtest suggests. It belongs in the watch-and-monitor bucket, not the deploy-with-confidence one.