It Wins 78% of Its Trades and Still Has No Edge

Engineering field note. Countertrend mean reversion is one of the most widely published strategy shapes in retail trading, and one of the most seductive: it wins the large majority of its trades, its drawdowns look tiny, and its equity curve climbs in small clean steps. We tested the family across 351 instruments in three arms. It wins 78% of its trades. It has no edge, and the reason is worth more than the verdict.

What the strategy is

It has no signals at all. Three moving-average lines are drawn; orders live on them and are dragged along every bar:

• a base average carries the closing order;
• a resting buy limit sits a fixed percentage below it;
• a resting sell limit sits a fixed percentage above it.

Price falls far enough, the buy fills, and the position closes when price returns to the average. Price rises far enough, the sell fills, same in reverse. Nothing decides to trade — price either comes to the order or it does not. The premise is simply that deviations from an average tend to be repaid.

The premise is broadly true, which is exactly the problem.

Why it looks so good

Run it as published and the report is genuinely attractive. Across our instrument set the median configuration won roughly seven trades in ten, drew down a fraction of a percent, and finished positive on four instruments in five.

A high win rate is not evidence of edge; it is a statement about the shape of the payoff. This strategy books many small gains and occasionally sits in a large loss while it waits for price to come back — and if price does not come back, it does not close at all. Win rate measures how often, never how much.

Three tests, and what each one removes

1. Compare against doing nothing

The first question for any strategy is not "did it make money" but "did it beat holding the same exposure". Our instruments were mostly in downtrends over the test window, so a short position made money by existing. Measured against that passive line, the published configuration's median result is negativa. It is harvesting a move that was free, and paying fees for the privilege.

2. Test it where it was not fitted

One instrument is one experiment. We ran the same fixed configuration across every instrument with enough history — no re-tuning, no cherry-picking. Passing means clearing the passive benchmark, the control below, and a minimum trade count.

3. Replace the offset with an arbitrary one

This is the control that matters, and it had to be designed specifically for this mechanism. Our usual null twin randomises the entry bars — meaningless here, because the strategy has no entry bars. So the twin keeps every part of the machinery and randomises only the distance to the average, between 0.4x and 1.6x the chosen offset. If an arbitrary distance earns what the chosen distance earns, then the parameter everyone tunes is decoration.

The campaign

Three arms per instrument: the strategy exactly as published in two variants, and a third where a genetic search re-optimises every parameter on that instrument's in-sample slice alone, judged on data the search never saw. The bar to read against is not zero — with a 5% threshold, a share of instruments passes on luck alone, and that share is printed next to every result.

Where we argue with our own gate

Seventeen instruments passed every gate, including the arbitrary-offset control. We still do not call that an edge, and the reason is a limitation of our own test that is worth stating plainly.

The control asks a ranking question: how often does an arbitrary offset beat the chosen one? On those seventeen it rarely did. But ranking says nothing about size. Comparing the actual numbers across the whole universe:

That is the finding. The strategy works in the sense that mean reversion is real; the parameter does not, and the parameter is the product. Anyone selling a tuned version of this is selling 0.35 percentage points of a move that was available for free — before fees, and before the cost of being wrong on the instruments that never came back.

We are adding effect size next to every pass count in our own reports because of this run. A rank-based control can be satisfied while the difference is economically meaningless, and a gate that says PASS on a 0.35-point difference is a gate that needs a second number beside it.

Two ways this family flatters itself in a backtest

Both are mechanical, and both were caught by testing the mechanism rather than the idea:

• Fills that could not have happened. A resting limit must be priced from the previous bar's average, or the same candle that moves the average also fills the order. And it must require price to trade through the level, not touch it — a wick reaching your price is not a queue position. Both details are worth more than any parameter.
• Positions that never close. The exit is a limit on the average with no stop. When price does not come back, the trade simply stays open and never appears in the statistics. A backtest that reports only closed trades will show a beautiful win rate and hide the inventory.

A related strategy family in our archive shows how expensive this class of error is: a trailing-stop implementation that folded the current bar's high into the peak before checking that bar's low turned an ordinary result into a headline "edge" across an entire board. After the fix, the live-eligible cells went from 119 to 1. The same book returned +720% at a 5 basis-point fee assumption and −599% at the venue's real spot fee — high turnover makes the fee assumption the whole result.

What would change the verdict

• an offset rule that beats an arbitrary offset by a margin large enough to survive fees and slippage, not by a rank;
• a closing rule that bounds how long a losing position can stay open, so the statistics include the inventory;
• results that hold on instruments where the passive direction was flat or against the trade, so the return is not the market's move wearing a strategy's name.

Until then the honest summary is short. It wins most of its trades, it loses to holding, and the tuning is worth a third of a percentage point.