Mean reversion EA: high win rate, bigger losers, and where the risk hides

What a mean reversion EA does

A mean reversion EA bets that a stretched price will snap back. When a pair has moved too far from some average, whether measured by an oscillator, a band or a simple distance from a moving average, the forex robot fades the move: it sells strength and buys weakness, aiming for a return to the middle. Targets are small and close. Stops, when they exist, are further away, because the EA is betting against the current direction and needs room.

It is the natural strategy for the quiet 80% of the time that currencies spend going nowhere. Range-bound pairs during Asian hours, EURCHF, EURGBP, AUDNZD: these are the classic homes of a counter-trend trading bot. Holding time is hours, sometimes a day, rarely longer. A well-built one trades a few times a day across several pairs.

The statistics you will see

Expect a win rate between 60% and 80%. The average loss is larger than the average win, usually 1.5× to 2.5×, which is the mirror image of a trend follower. Profit factor on an honest, stop-protected record sits around 1.2 to 1.6, and it is fragile: a few percentage points of win rate separate profit from loss. Expectancy is small per trade and relies on volume.

Holding time in the Activity distributions histogram concentrates in the 1–4 hours and 4–24 hours buckets, with some 15–60 minute winners when the snap-back is quick. Trades per week ranges from ten to forty depending on how many pairs it runs. The hour chart often shows a strong Asian-session bias, with almost nothing during the London open, because that is when ranges break.

The equity curve rises smoothly with frequent shallow dips: many 3–6% drawdowns recovered in days, punctuated by an occasional 15–20% fall when a range turns into a trend and the stop is hit repeatedly. The Z-score tends towards positive, meaning wins and losses alternate more than chance would suggest, since the strategy fades each direction in turn. Open positions are usually one or two per pair, often with a small floating loss while the trade waits for the reversion.

A worked example: why a few points of win rate decide everything

Take a mean reversion EA with a 72% win rate, a $40 average win and a $110 average loss. Expectancy is 0.72 × 40 − 0.28 × 110 = 28.8 − 30.8 = −$2 per trade. It loses money with a 72% win rate. Raise the win rate to 75% and expectancy becomes 30 − 27.5 = +$2.50. The entire edge lives inside three percentage points, and three points is well within the normal variation of a 200-trade sample.

Compare that with the break-even win rate formula from win rate is lying: 1 ÷ (1 + payoff ratio). With a payoff ratio of 40/110 = 0.36, break-even is 73.4%. Any EA whose observed win rate sits within a couple of points of its break-even is not proven, regardless of how green the curve looks. Put the numbers into the expectancy calculator and see how much margin there is.

Advantages

• Frequent small wins make it psychologically easy to run, and the frequent sample means the statistics settle within months rather than years.
• It works in the market condition that is most common, which is no condition at all.
• Short holds keep swap costs small and limit weekend gap exposure.

Disadvantages and failure modes

The failure mode is structural: mean reversion loses exactly when the market decides to trend, and it loses several times in a row because the EA keeps fading a move that keeps going. The 2022 dollar trend and the 2015 Swiss franc move are the textbook cases. A range EA on EURCHF with a wide stop lost years of gains in a minute in January 2015.

The other failure is the slow slide into not taking losses. Vendors learn that a visible stop-out ruins the curve, so they widen the stop, then remove it, then add a second entry to average down. At that point the mean reversion EA has become a grid or a martingale, and everything in those articles applies.

What the AI review and Score tend to flag

The review usually labels a mean reversion EA as Day trading, since average holds fall between an hour and a day. The flag to watch for is the high-severity tail flag, raised when the average loss exceeds 2.5× the average win. A disciplined range EA sits just below that line; a sloppy one crosses it. If the review also finds a win rate of 65% or more with losses at least twice the size of wins, the martingale signature fires even without lot progression, and the strategy type flips to Martingale. The stop-loss percentage reported in the review is the other key figure: under 15% raises the no-stop-loss flag.

The Score tends to do well on consistency, since most months are green, and is limited by a middling profit factor and whatever drawdown the worst trend inflicted.

Red flags when buying a mean reversion EA

• Win rate above 85%: almost certainly losses are being held rather than taken.
• Stop-loss on under half of trades with no explanation of a basket or equity stop.
• Maximum loss in the P/L distribution many times larger than the average loss.
• A short record that never included a strong trend. Check the calendar for the months when the traded pairs trended.
• Max DD (equity) in Glance far larger than Max DD (balance).

How to check on an account page

Start with Trade statistics: win rate, average win against average loss, and trades per week. Compute the break-even win rate and see how far above it the EA sits. Open the AI review for the stop-loss percentage and any tail flag. In Activity distributions, confirm the hour chart matches the quiet sessions the vendor claims. Performance depth's tail ratio should be close to 1 or above; a mean reversion EA with a tail ratio well under 1 is taking bigger bad days than good days. The deepest drawdowns table tells you how long the worst trend took to recover from. In Risk Lab, the chance of ruin ladder at −25% is the honest summary of the tail, and the VaR figures show what one bad day looks like in currency. Finally, look at Open positions: a single small position waiting for reversion is normal; three averaging entries on one pair are not. Filter Rankings by type at /rankings?type=ea and compare a few before you commit.