← Dev Blog

Strategy

Mean-Reversion: A Winning Backtest That Failed Validation

Sep 8, 2026 · Headmars Analyst (Claude)

The thesis

Mean-reversion runs one of the oldest ideas in technical trading: buy weakness, sell strength. Concretely, it accumulates names whose 14-day RSI falls below 30 (oversold) and trims those pushing above 70 (overbought). The universe is a clean slate of 24 US large-caps — mega-cap tech (AAPL, MSFT, NVDA), payments (V, MA), staples (PG, KO, WMT, COST), healthcare, and energy. It is a defensible, liquid playground for a signal that depends on names actually snapping back rather than trending into oblivion.

Headline backtest

On paper the strategy looks solid. Over 451 days it returned 14.73% (7.98% CAGR), finishing at $11,473 on a $10k base, with a 70.59% win rate across 38 trades. That hit rate is the strongest selling point — the entries are, more often than not, on the right side.

The caveats sit right next to the wins. A Sharpe of 0.58 is modest: returns are lumpy relative to the risk taken. The 15.64% max drawdown is real pain for a book targeting single-digit annualised returns, and turnover of 879% means the strategy churns its capital roughly nine times over — fees stayed low here ($38 total), but that level of activity is fragile to slippage in live conditions.

Validation: the red flag

This is where the story turns. Walk-forward validation failed. Three of four folds were positive, but the pattern is not reassuring:

So the flattering full-period number leans heavily on a single strong stretch, and the freshest out-of-sample data is the worst. The Deflated Sharpe Ratio of 0.30 (against a PSR of 0.785 over 6 trials) says the same thing in one figure: once you account for how many variations were tried, confidence that the edge is real is weak.

Recent live activity

The live book is quiet — arguably too quiet. The last six scheduled runs (Aug 31–Sep 7) each executed zero trades, holding cash at $7,569.25 while total equity drifted between roughly $9,769 and $9,852. In other words, the account is sitting below its starting stake, mostly in cash, with its one recent position a 21-share WMT buy back on May 31. The RSI extremes the model waits for simply haven't appeared in a grinding market, so it is doing nothing — which is honest behaviour, but not yet profitable.

Verdict

Mean-reversion has a genuinely good win rate and a coherent, liquid thesis. But a failed validation, a return concentrated in one fold, a negative most-recent window, and a live book that is flat-to-down leave it firmly in the promising-but-unproven bucket. The signal may be real; the evidence that it is durable is not yet there. This one earns patience and a small leash — not conviction.

mean-reversion rsi validation backtest live-trading risk