The thesis
bollinger-reversion runs one of the oldest ideas in technical trading: buy when price closes below the lower Bollinger band, sell when it pushes above the upper band. It's a pure mean-reversion bet — that stretched moves snap back — applied to a 24-name universe of US large caps spanning tech (AAPL, MSFT, NVDA, GOOGL), financials (JPM, BAC, V, MA), staples (PG, KO, WMT, COST) and healthcare (JNJ, UNH, PFE, ABBV). It's marked live.
What the backtest says
On paper, the headline numbers flatter the strategy. Over 451 trading days it returned 17.55% (a 9.46% CAGR), turning a $10,000 book into $11,755. The win rate is 63.89% across 76 trades — high, and consistent with a mean-reversion profile that clips many small winners. Fees were a negligible $76.
But two figures deserve a raised eyebrow. The Sharpe of 0.66 is mediocre for a return that size, implying a bumpy ride — confirmed by a 20.57% max drawdown. And turnover of 1,720% is heavy churn for so few net trades, the kind of activity that punishes any real-world spread or slippage the paper model ignores.
Why validation fails
This is where the story turns. The walk-forward test splits history into four folds, and all four are positive — a genuine strength. The problem is the trajectory: returns decay monotonically, 7.67% → 1.42% → 0.77% → 0.40%, with Sharpe collapsing from 1.78 in the first fold to 0.14 in the most recent. Nearly all the edge lived in one early window (Aug 2024–Jan 2025); everything since is a rounding error.
The statistics agree. The Probabilistic Sharpe Ratio is a respectable 0.814, but the Deflated Sharpe Ratio — which penalises the 6 trials run to find this configuration — is just 0.342. That gap is the fingerprint of selection bias, and it's why the gate returns passed: false. An out-of-sample return of 0.4% is not a strategy you deploy with conviction.
Live reality
The live tape reinforces the caution. The last executed trades were back in late May and early June (buys in DIS, WMT, COST, GOOGL, PG). Since then, every scheduled run from July 17 to July 24 reads the same: 0 executed, 1 rejected. Cash has sat frozen at $1,001.98 while total equity drifted down from $9,787 to $9,467 — the book is underwater versus a $10,000 start, and the engine is effectively stuck, unable to act on its signals.
Verdict
bollinger-reversion is a clean, interpretable strategy with an honest edge in one regime and a high hit rate. But the evidence points to overfitting: the out-of-sample edge has bled away, the deflated statistics fail, and live performance is flat-to-negative with orders piling up rejected. Strengths are real; the risk is that they're historical. This one belongs in the lab for a re-fit — ideally with wider bands, a regime filter, or a hard look at why live orders keep bouncing — not in a funded book.