The thesis
Mean-reversion is one of the oldest ideas in systematic trading, and this agent runs the textbook version: buy when the 14-day RSI drops below 30 (oversold), sell when it climbs above 70 (overbought). It trades a universe of 24 large-cap US names — AAPL, MSFT, NVDA, JPM, JNJ, XOM, V, MA and the like. The wager is simple: short-term price extremes in liquid blue chips tend to snap back toward their recent average.
Recent activity
The strategy is live but notably patient. Of the six scheduled runs between Oct 12 and Oct 19, five executed nothing at all — the book simply held while no name reached an RSI extreme. The exception was Oct 16, which closed a Walmart position: 21 shares sold at $118.83 against a $115.75 entry from late May, a tidy +2.7% round trip. Earlier in the autumn the agent was accumulating — McDonald's, Home Depot and Bank of America all bought in September. The live paper book now sits at roughly $9,858 total with about $2,820 in cash, modestly below its $10k notional and a useful reminder that live results trail the backtest.
Backtest performance
Over 451 days the strategy returned 14.73% (7.98% CAGR), finishing at $11,473 on a $10k base. Sharpe came in at 0.58, with a 70.59% win rate across 38 trades and a 15.64% max drawdown. Fees were trivial ($38, no FX cost). A high win rate paired with a modest Sharpe is classic mean-reversion: many small winners offset by occasional larger losers. Turnover of 879% is elevated but unremarkable for a signal that round-trips frequently.
Validation: the red flag
Here is where I temper the enthusiasm. Walk-forward validation did not pass. Across four folds, three were positive — fold 2 was excellent (+11.1%, Sharpe 1.32) and the two shoulder windows added +2.06% and +2.21% — but the most recent out-of-sample period (Dec 2025 → May 2026) lost 2.84%, with a negative Sharpe (-0.33) and a near-15% drawdown. That -2.84% is the honest out-of-sample number; the headline 14.73% is a full-sample figure that includes the training data. A deflated Sharpe of 0.30 across six trials suggests the backtest flatters itself relative to what fresh data actually delivers.
Verdict
The character is genuinely attractive: high hit-rate, cheap to run, low activity, and economically defensible. But the gate failed for the right reason — performance was concentrated in one strong fold (early 2025) and decayed into the latest period. The current live stretch, slightly underwater and largely inactive, is consistent with a regime where oversold signals keep working against the trader. My call: keep it live on paper, watch whether the next fold reverts to positive, and resist sizing it up until out-of-sample performance stops deteriorating.