"A high-win-rate martingale EA is still scary—the story of 0.2% that you don't see in averages"

When looking at EA backtest results, the first thing that catches the eye is often the win rate, isn’t it?

In particular, nanpin-type EAs tend to show high win rates. The performance ratio (PF) exceeds 2, and the graph is upward-sloping. Visually, it looks excellent.

This time, based on that experience, I will write about “why it can be scary even when the win rate is high.”

Let’s consider an extreme example (purely as an analogy).

• Out of 100 runs, 99 win by +1,000 yen

• Only 1 loss of -100,000 yen

The win rate is 99%. However, the total is 99,000 yen − 100,000 yen =−1,000 yen, resulting in a loss.

Nanpin EAs have a structure close to this pattern.

Small adverse moves add positions to improve the average price, and you can take profits on a rebound. That’s why “wins” accumulate easily. On the other hand, if the market continues in one direction, positions increase and unrealized losses can explode.

In other words,the size of losses tends to concentrate in a few moments.

Behind “average” numbers like win rate, PF, and average profit, this big loss in a few moments often hides.

Therefore, I personally think thatfor nanpin-type EAs, win rate is hardly a reliable metricbecause it tends to be high for any EA.

When choosing an EA, I believe there are two things I really want to confirm.

1. How large can a rare, large unrealized loss be?

2. How often does that occur?

So I actually investigated with my own developed EA “Senju” (XAUUSD, M5).

Backtests from 2023 to August 2026 completed7,071 basketsand recorded the lifetime maximum unrealized loss per basket tick by tick.

As a result, most baskets ended with small unrealized losses. Then, how big were the rest?

This article will cover four things:

• The distribution of the “maximum unrealized loss” for 7,071 baskets (table)and the contents of cases that swelled to about 585,000 yen

• Common market characteristics in positions that became large unrealized losses

• Three numbers to check when looking at an EA (checklist from the sales page)

• Why you should not blindly trust an EA with long backtest periods and low DD

99.80% of baskets had maximum unrealized losses under 100,000 yen. 99.93% under 200,000 yen, and 99.97% under 300,000 yen.

At first glance, it looks like a very stable EA. But what I want you to look at is the bottom of the table.

Only 14 baskets exceeded 100,000 yen, but among them wasa case that expanded to about 585,000 yen.

On a 1,000,000 yen account, this unrealized loss alone would be about 60% of the funds.

I interpret this result as follows:

Most end with small losses, while a very small fraction occasionally expands greatly.This is the so-called “long-tail” risk structure.

If you only look at averages or win rates, you won’t see this “0.2% remaining.” What matters is to decide whether to use it after seeing what happens in that 0.2%.Make the decision only after understanding what happens in the remaining 0.2%..

Looking by year, risk manifestation differed greatly.

In 2023, as volatility remained low, the number of stages increased, followed by a large trend that recorded unrealized losses of about 585,000 yen.

Also, in many of the baskets where the maximum unrealized loss exceeded 100,000 yen, important fundamental factors such as FOMC, US payrolls, statements by Fed officials, US elections, and geopolitical news overlapped,indicating a confluence with major fundamentals.

The danger is not the moment of a data release itself. It is the pattern of reversing into the trend that lasts for hours to days after the release, causing deepening nanpin losses.

Before looking at win rate or PF, please check whether the following three items are stated in the sales page or report.

If these three are not written, or if you cannot get an answer, it is safe to assume there is not enough information to judge the EA’s risk.

Another point I watch carefully is the backtest period.

An EA with a long period but a low maximum drawdown may look excellent at first glance. However,there is a large potential for post-hoc adjustments. It is easy to craft a filter to avoid entries that caused large DD.

Even for Senju, a long-term stress test back to 2015 showed three major SELL incidents in 2019, 2020, and 2022.

If you exclude these later by conditions, you could create a “backtest surviving for 10+ years.” But that tends to be optimization aligned with past markets and does not guarantee future effectiveness.Forward performance is a big factor. However, if, as in my own experience, the developer can turn on or off during operation, you cannot read performance as solely the EA’s result.

With Senju, I learned that simply stopping for 30 minutes before and after indicators is not enough. Therefore,I added ON/OFF control for starting new baskets.

Even if turned OFF, any existing basket’s nanpin, profit-taking management, and emergency cuts continue. It only stops the new first-level entry.

However, this does not mean zero accidents. Since nanpin is used, if caught by a large trend, there remains a possibility of large unrealized losses.

• In nanpin systems, win rate is hardly a reliable metric(high across all EAs)

• Averages and win rates conceal“a few big losses”.

• Check maximum basket loss, maximum number of stages, and longest holding period (of course also consider profits)

• Do not blindly trust an EA with long backtests and low DD, as this could indicatepost-filter adjustments.

• Decide in advance how far to dig into a move and what to do when you dig before using.

I believe nanpin-type EAs are realistically something you engage with on the premise that a big loss can occur at the end.

If so, the question becomes how far to dig, what to do when you dig, and thento rotate with limited funds and take profits out of the account, which is another viable strategy.

However, please operate with funds you can afford to lose without affecting your living.

I will describe next time, with actual simulation results, how a 1,000,000 yen account would behave with about 60% of that balance, and for a 300,000 yen account…