The thesis
Mean-reversion is one of the oldest ideas in systematic trading, and this agent runs the textbook version: buy oversold names when RSI drops below 30, sell overbought ones when RSI climbs above 70. It operates over a 24-name large-cap universe spanning tech (AAPL, MSFT, NVDA, GOOGL), financials (JPM, BAC, V, MA), healthcare, staples, and energy. The bet is simple — that short-term price extremes in liquid blue chips tend to snap back toward the mean.
Backtest: the bull case
On paper, the numbers flatter the strategy. Over 451 days it returned 14.73%, ending at $11,473 on a $10k base — a 7.98% CAGR. The win rate is 70.6% across 38 trades, which is genuinely high and consistent with mean-reversion's known tendency to be right often while risking the occasional sharp loss. Trading costs were negligible ($38 in fees, no FX drag).
The caveats hide in the texture. The Sharpe ratio is a modest 0.58, max drawdown ran to 15.64%, and turnover was a hefty 879% — this is an active book, not a buy-and-hold. A high win rate paired with a middling Sharpe tells you the losers, when they come, are large enough to offset the frequent small wins.
Validation: where it breaks
This is the part that matters, and it is where I'd pump the brakes. The strategy failed cross-validation. Of four walk-forward folds, three were positive — a tidy 2.06%, a standout 11.1%, and 2.21% — but the most recent fold (Dec 2025 → May 2026) lost 2.84% with a negative Sharpe of −0.33 and a 14.96% drawdown. That the failure lands in the latest window is the opposite of reassuring.
The deflated statistics confirm the unease. After adjusting for the 6 trials explored, the Deflated Sharpe Ratio is just 0.30 — below the 0.5 line I'd want to see before trusting an edge is real rather than a product of selection. The Probabilistic Sharpe Ratio (0.785) is healthier but not decisive. In short: a respectable full-sample return built partly on one strong fold, with no durable out-of-sample confirmation.
Live activity
The live book echoes the warning. The agent is near fully invested — cash of just $271.67 — after a late-October buying burst (BAC, GOOGL, CAT) and a lone NVDA purchase on 6 November. Since then it has been quiet: six scheduled runs, five with zero executions. More tellingly, total equity has drifted between roughly $9,510 and $9,768, sitting at $9,662 as of 12 November — below the $10k starting line even as the backtest advertises a double-digit gain.
Verdict
Mean-reversion has a credible thesis, low costs, and an attractive hit rate. But a failed validation, a sub-0.5 DSR, a negative latest fold, and a live book underwater all point the same direction: treat the 14.73% headline with suspicion. This is a strategy to watch and stress-test, not to scale.