A discretionary method with an 84.7% win rate turned into an EA, revealing the "reproducibility barrier" — measuring the distance between the person and the EA with six months of data
“If you turn your own winning discretionary method into an EA, you can make money even while you sleep.” — This is something discretionary traders think about at least once. I was the same. Six months ago, when I analyzed my GOLD discretionary history, I found statistically almost guaranteed edge numbers: win rate 84.7%, expected value +$4.95/oz, cumulative +5.01 million yen, and maximum drawdown 116,000 yen.
I thought, “If I make this into an EA,” and started turning it into an EA with Claude and MQL4.
In short,from v1 to v6, all failed, and only in v7 did I barely surpass break-even. Backtests did not come close to the actual performance.
In this article, without disclosing the concrete method, I will write in numbers how far the “personal discretion” and the “EA-ed copy” were apart in the data. This is the reality you should know before considering discretionary EA conversion.
In this article,① evidence that the person’s discretionary edge was statistically real, ② the progression of reproducibility for 7 EA versions, ③ the numerical view of the difference between the person and the EA, ④ the nature of the parts that cannot be reproduced, I will write only with actual numbers. It is meant to help those trying/considering discretionary EA conversion avoid detours.
? Table of contents
▼ Free公開
- Evidence that the person’s discretionary edge was statistically real
- Reproducibility progression of EA versions 7
▼ Read more for公開
- Numerical view of the difference between the person and the EA (win rate, expected value, PF, DD)
- Statistical backing confirming the person’s edge is real through three directions of falsification
- From the 19 demo details, the nature of the non-articulated exit structure
- Cause of not filling residual 0.35R: you cannot train a selection rule with only positive examples
Evidence that the person’s discretionary edge was statistically real
Before starting to EAize, I confirmed whether there truly was edge in my own discretion.
This is the result of analyzing all 85 actual trades (June–October 2025, GOLD) + 19 demo details (July–August 2026).
All numbers for 85 actual trades
- Win rate: 84.7%
- Payoff ratio: 2.68
- Expected value: +$4.95/oz
- Cumulative profit (yen): +5,010,000 yen
- Maximum drawdown (yen): 116,000 yen
- Period profit / maximum DD: 43.3 times
- Buy/Sell breakdown: 56 buys, 29 sells
- Sell trades win rate: 89.7% (GOLD rose 18.4% during the period)
How statistically genuine is this "real"
A win rate of 84.7% × payoff of 2.68 yields an expected value dozens of times higher than an average retail trader.
Moreover, in a market where GOLD rose +18.4% during the period,29 short trades had an 89.7% win rate, higher than the buys, so the explanation that “edge came from market direction β” does not hold. Since selling against the market and still winning is possible, there is a clear alpha in direction selection and exit handling.
In other words, the claim that the “person’s discretionary edge exists” is statistically confirmed statistically confirmed. If so, the idea was that turning this into an EA would yield a profitable EA—this was the expectation at the start.
Reproducibility progression of EA versions 7
I worked with Claude and MQL4 to create 7 versions in sequence. Each version added one new element to the person’s verbalized rule and tested its effect with a backtest.
I will not disclose the method’s inner workings, but I will present the numbers for each version.
v1–v7 win rate and expected value progression
- v1: Win rate 43%, expected value -0.421R
- v2: Win rate 47%, expected value -0.298R
- v3: Win rate 49%, expected value -0.256R
- v4: Win rate 51%, expected value -0.201R
- v5: Win rate 53%, expected value -0.187R
- v6: Win rate 58%, expected value -0.108R
- v7: Win rate 64.3%, expected value+0.021R(first positive)
Distance from the person’s actual performance
- Person: Win rate 84.7%, expected value +$4.95/oz (roughly +0.35R in yen)
- EA (v7): Win rate 64.3%, expected value +0.021R
- Difference: Win rate −20.4%, expected value about −0.35R
Even though v7 surpassed break-even for the first time, the person lost about 0.35R per trade.
About 35R across 100 trades—a gap where more than half of the winning trades by the EA disappear.
This was the practical milestone for “discretionary edge to EA conversion.”
How to interpret the distance between “Person vs EA”
This gap means“less than half of the edge the person uses could be translated into code”.
The other half largely depends on implicit judgments in the person’s head (something not verbalized).
From here, I will write aboutquantifying the differences in PF, DD, and holding time between Person vs EA, the process by which the person edge was statistically confirmed as real through three-direction falsification, the non-verbalized exit structure found in the demo details, and the reason residual 0.35R cannot be filled (only positive examples cannot teach a selection rule), in detail.
This is especially impactful for those currently trying or considering discretionary EA conversion.
From here (Read more) onward, this will be published:
- Quantifying the differences between Person and EA in PF, maximum DD, and holding time
- In three directions of falsification, how the person’s edge was proven real (denying β explanations, zero monthly losses, excluding the top 5 wins)
- From the 19 demo details, facts such as “TP reached 0/19” and “the price moved beyond the set level stop 12/19”
- Why residual 0.35R cannot be filled: not only positive examples can train a selection rule
- A one-month preparation procedure for those who will pursue discretionary EA conversion