The thesis
Mean-reversion is one of the oldest ideas in systematic trading, and this strategy runs the textbook version: buy when a name is oversold (RSI below 30), sell when it is overbought (RSI above 70). It operates across a 24-stock universe of US large caps — the familiar megacaps (AAPL, MSFT, GOOGL, NVDA), financials (JPM, BAC, V, MA), healthcare (JNJ, UNH, PFE, ABBV) and staples/industrials (PG, KO, WMT, COST, CAT, XOM). It is currently flagged live.
The backtest looks good
On paper, the numbers are attractive. Over 451 days the strategy returned 14.73%, lifting a notional $10,000 book to $11,473, for a 7.98% CAGR. The headline strength is consistency of entries: a 70.59% win rate across 38 trades. Costs were negligible — $38 in fees, zero FX. Risk was moderate but real, with a 15.64% max drawdown and turnover of 879%, meaning the book churned itself roughly nine times over the period.
The catch is the Sharpe ratio of 0.58. A high win rate paired with a middling Sharpe tells a familiar mean-reversion story: many small wins offset by occasional larger losses. That is fine until the losses cluster.
Validation says wait
This is where the picture darkens. The walk-forward validation fails. Across four folds, three were positive — folds one through three returned +2.06%, +11.1% and +2.21% — but the most recent fold (2025-12-16 to 2026-05-29) returned -2.84% at a -0.33 Sharpe, with a 14.96% drawdown. The single strongest fold (+11.1%, Sharpe 1.32) is doing a lot of the heavy lifting for the full-sample result.
The deflated statistics reinforce caution. The probabilistic Sharpe (PSR) of 0.785 is respectable, but the deflated Sharpe ratio (DSR) of 0.304 — which penalises for the six trials run — sits well below the confidence threshold you would want before trusting the edge. In plain terms: after accounting for how many variants were tried, the evidence that this beats zero is weak, and the newest data is the weakest of all.
Live activity: a holding pattern
The live book mirrors that hesitation. The last executed trade was a buy of 21 WMT shares at $115.75 on 31 May. Every scheduled run since — six of them logged between 2 and 10 September — reports 0 executed, 0 rejected, with the book parked at $7,569 cash and a total value drifting between roughly $9,788 and $9,852. That total sits below the $10,000 starting line, so the strategy is modestly underwater live and has generated no fresh signals for over three months. No stock in the universe has recently been oversold or overbought enough to trip a threshold.
Verdict
The strengths are genuine — a clean, interpretable thesis, low costs, and a high hit rate. But two flags argue against sizing this up: validation fails on deflated metrics, and the deterioration lands squarely in the most recent fold. Combined with a live book that has been dormant and slightly negative since May, the honest read is promising but unproven. Keep it live at small size, watch whether the RSI triggers ever fire again, and treat the backtest's 14.73% as a ceiling, not an expectation.