Despite a win rate of 95%, in the red. The pitfall lurking in GOLD EA's SELL
GOLD dedicated EA “Transition - TENRYU -” development log, Part 4.
When I look at the backtest of the EA, I can’t help but worry about the win rate.
If the win rate is 80%, that’s fairly high; if it exceeds 90%, it looks quite stable.
However, the diagnostic data left during TENRYU’s development showed results like these.
If you’re the first to see the numbers, you might find it a little strange.
Examining the performance of Resistance R1
In the diagnostic version created on September 5, 2026, the trade performance for each horizontal line was exported to CSV.
From which line did the entry occur.
Which entry number was it.
A system that aggregates profits and losses for each.
From that data, when verifying trades using Resistance R1, the results for 2024 were as follows.
The P/L here is the value recorded in the diagnostic CSV at that time.
Still, because PF is below 1, you can see that this group of trades was loss-heavy.
It’s a part that isn’t visible from win rate alone.
2025 was even tougher
When checking 2025 with the same diagnostic, R1’s results were as follows.
This is also above 90% win rate.
However, PF is even lower than in 2024.
In other words, there were many wins, but a few losses dwarfed overall profits.
With counter-trend trading using horizontal lines, you can seize many small rebounds, but if price continues moving without reversing, you can incur large losses.
This diagnostic result quantified that risk.
Why you can lose even with a high win rate
For example, suppose you pile up dozens of profitable trades.
But then one subsequent trade wipes out those profits all at once.
Even if the win rate stays high, the final P/L can be negative.
Conversely, even with a lower win rate, if you significantly extend profits and keep losses small, you can end up profitable overall.
That’s why PF and drawdown are important when you look at an EA.
Back then, TENRYU also adjusted the profit side, such as changing from a fixed 3-dollar take profit to a stepped take profit.
But that alone didn’t solve all the problems.
That is a question we needed to consider.
2026 was profitable even with the same R1
By the way, 2026 diagnostic results were somewhat different.
R1: 468 trades, win rate 96.37%, PF 1.079, aggregated P/L was positive.
2024 and 2025 were negative.
2026 was positive.
Even with the same type of entry, the results differ by year.
Looking at this, it’s not easy to simply say “stop SELL from R1 entirely.”
However, with only this data, it’s not possible to conclude which filters contributed to improving performance.
Market conditions vary by year, and confirming comparison conditions is also necessary.
What we understand here is that, at least among SELLs, there existed a group of trades that deserved careful examination.
Look for where you’re losing, not just overall win rate
When looking at backtest results, you may want to conclude evaluation from profit amount and win rate alone.
However, when you split trades by type, issues appear that the overall numbers can’t reveal.
In TENRYU, we could classify R1 and R2, standard reversals and Role Reversal, entry counts, etc.
As diagnostics accumulated, a mechanism to restrict entries according to trend direction also grew.
These figures are not the final version’s performance.
Nevertheless, they are an important record showing what needed to be investigated during EA development.
Capture the “turn” born from horizontal lines and chase the ensuing “flow.”
A GOLD (XAUUSD) dedicated MT5 EA.
- Automatically detects horizontal lines and supports wicks contact and roll reversal
- As long as the market trend continues, gradually follow profits
- Built-in fixed stop loss, time-based stop, and suppression of new entries (no martingale)
- Standard setting is 0.01 lot / Range boost OFF
Net profit +1,242,820 JPY / Maximum drawdown 24%
※Backtests are verification results based on historical data and do not guarantee future profits. We recommend testing on a demo account before live operation.
※The numbers in this article come from an early development versionv0.6.38DDiagnosed data. They do not represent the annual performance of the current retail version.