Closed-Trade vs Equity Drawdown on MT5: A Portfolio Ratio Fell from 1.46 to 1.07 When I Counted Open Losses

This week I failed one of my own portfolio tests, and the fault was in the ruler, not the strategies. Early on Wednesday a five-part portfolio of my research legs cleared the return-to-drawdown line I had set for it. That evening I measured the same trades again, this time counting open losses at every daily close instead of only closed trades, and the ratio fell from 1.46 to 1.07. The line was 1.33. This entry is about why that happened, what I changed, why the changed version still fails, and what I deliberately did not do afterwards.

Nothing here is a product and nothing here is live money. It is research on historical data, plus a demo account. I am writing it up because the mistake is common, it is easy to miss, and a balance curve helps you miss it.

The target, written down first

For most of this year I tested strategies one by one and asked whether each had an edge after costs. This month I wrote the goal down as numbers instead, so that a candidate either meets it or does not:

• Holding period: a few days at most.
• At least 12 trades a month.
• Annual return of at least +40% with a maximum drawdown of no more than 30%. In R units, annual R divided by the deepest drawdown in R must be at least 1.33.
• If you start on a random day, the account should be positive 90 days later in at least 70% of starting days.
• Fully automated, one broker.

No single leg I have comes close. Each one is too thin, too rare, or too deep in drawdown on its own. So the question became whether several legs with different mechanics, simply added together at 1R per trade, could reach the line as a bundle.

The five legs

I will describe them by mechanism rather than by exact parameters:

• A gold breakout leg that can hold for up to about five days and has a fixed stop.
• A USDJPY narrow-range breakout (an NR7-style day followed by a stop order on either side), closed the same day.
• An intraday index leg that holds during New York hours only.
• An RSI(2) pullback buy on five index CFDs: buy at the close when price is above the 200-day average and RSI(2) is below 10, sell when the close moves back above the 5-day average. This is the well-known public rule. In the original research version it had no stop loss and no time limit.
• A pre-holiday index leg that holds over the session before a US market holiday.

Each leg's trade list came from an earlier, separate study. I added them together by date over a common window, July 2018 to July 2026, with each trade counted as 1R and costs already inside each trade.

First pass: it cleared the line

Counting closed trades only, the bundle had 2,549 trades (26.3 a month), +35.9R a year and a deepest closed-trade drawdown of 24.5R. That is a ratio of 1.46, above 1.33. Starting on a random day, 77% of starts were positive 90 days later. On the face of it, three of the numeric targets were met.

The weaknesses were already visible in that first report. The first half of the window, July 2018 to June 2022, had a ratio of 0.41. The second half had 3.96. Over a longer window starting in 2012 the ratio was 0.76. With every leg's average trade cut in half, the ratio was 0.38 and only 64% of starts were positive after 90 days. If you remove almost any one leg, the bundle falls below the line. And one item was flagged as a known gap: the drawdown counted closed trades only, and one of the legs had neither a stop nor a time limit.

Second pass: counting what was open

Before measuring again I wrote a short pre-registration file with the fixes, the exact way open trades would be valued, and the pass lines. It also says what I would do if the bundle failed: report it, and not rescue it by tuning. The file is timestamped before the calculation.

Open positions were valued once a day. The RSI(2) trades used each index's server-day close, minus the round-trip spread and the swap accrued so far. The gold leg used its own source data at 21:00 UTC. The pre-holiday leg holds only one night, so only its entry-day close was valued. The two intraday legs never hold overnight and were left on closed trades.

The equity curve then looks different. With the original RSI(2) leg, the deepest drawdown went from 24.5R to 33.5R, and the ratio from 1.46 to 1.07. The bundle that cleared the line that morning does not clear it once you look at what was open.

Almost all of the gap came from one leg. Taken alone over the same window, the original RSI(2) leg had a closed-trade drawdown of 29.4R and a mark-to-market drawdown of 46.9R. The gold breakout leg moved from 40.5R to 40.9R. The pre-holiday leg moved from 9.9R to 10.0R. The legs with a stop or a fixed short hold barely changed. The leg with neither grew by more than half.

Why the closed-trade curve hides it

A pullback buy without a stop does not take losses in the normal sense. It waits. When the market keeps falling after entry, the trade stays open, the loss grows, and nothing reaches the closed-trade history. Most of the time price eventually closes back above the 5-day average and the trade is closed for a smaller loss than its worst point, or even a win. The closed-trade equity curve shows a single step on the exit day. The account, during those days, was much deeper underwater.

This is the same reason closed-trade drawdown is a poor ruler for any strategy that closes losers late, including grid and averaging systems. The MT5 Strategy Tester does report equity drawdown alongside balance drawdown, and the difference between the two is worth looking at. My own Python bundle code had only the balance version until this week. Across the full 2012 to 2026 trade list of the original RSI(2) rule (625 trades), the longest hold was 24 calendar days, and the 143 trades (23%) that stayed open longer than four daily bars closed at an average of −2.02R. The closed curve shows each of them as one step on the exit day and nothing of the days before it.

The fix I chose, and why only one

The goal says a few days at most, and an unlimited hold does not fit that, so the RSI(2) leg had to change anyway. I wrote down one version before computing anything:

• A stop at 3 daily sigma below the entry price, active from the day after entry. If the day opens below the stop, the trade is closed at that open price.
• A hard exit at the close of the fifth trading day, counting the entry day as day one.
• The original 5-day-average exit stays, and whichever comes first wins.

The reasoning for 3 sigma was simple. Over four days, normal movement is about 2 sigma, so a 3-sigma stop is hit only occasionally by ordinary noise (I expected 10 to 20% of trades) but cuts the long tail of continued selling. It also matches the threshold the original study had already used to count large losses. I did not try 2, 2.5, 3.5 or 4 sigma. If I had tried five widths and kept the best one, the result would carry that selection, and I would have no way to tell how much.

Before building the new version I checked that my rebuild of the original rule reproduced the original trade list. It did: 625 trades against 625, the same entry and exit dates, and a per-trade difference no larger than floating-point rounding.

The fixed bundle still fails

Over the full 2012 to 2026 data, the changed RSI(2) leg has 669 trades instead of 625, because earlier exits free the next entry. The average trade fell from +0.325R to +0.231R and the win rate from 71% to 67%. The worst single trade improved from −9.57R to −3.47R, a gap-down on the German index in March 2025. The median hold went from 5 to 4 calendar days and the maximum from 24 to 8. Of the 669 trades, 93 were stopped out (three of them at a gap open below the stop) and 84 were closed by the time limit, the latter averaging −1.26R. Those two groups are the trades the original rule would have kept open and waited on.

Inside the bundle, over the common window:

• 26.6 trades a month: passes.
• +32.7R a year against a deepest drawdown of 27.8R with open trades counted: ratio 1.18, below 1.33. On closed trades only it would be 1.27, which also fails.
• Positive 90 days after a random start: 75%, passes.
• A discounted version, where every leg's average trade is cut in half to allow for having chosen these legs after seeing their results: ratio 0.30, positive after 90 days in 61% of starts, below the 65% line I had set for it.

Two of the four lines fail, so by the rule written beforehand the bundle does not pass. The half-window split also stayed lopsided: 0.25 in the first four years and 3.71 in the last four. Over the longer window from 2012 the fixed bundle is 0.86. In a block bootstrap of three-month blocks, only 39% of resamples reached 1.33 on the closed-trade measure.

The stop also made something visible that was not visible before: how much risk is open at the same time. The five index CFDs tend to give RSI(2) signals on the same days. At the worst moment in the window the bundle held seven positions at once, five of them RSI(2) trades, with 17.0R at risk (legs without a hard stop counted as 1R by convention). At the risk per trade that would scale the drawdown to 30%, that is about 18% of the account exposed in a single moment. With the original rule this number could not be computed at all, because there was no stop to measure distance to.

My estimate is still optimistic in several ways. Open trades were valued at the daily close, not at the intraday low. The stop assumes no slippage except on gap opens. The two intraday legs have intraday drawdown that I did not measure. A fuller mark-to-market would push the ratio down, not up. One item leans the other way: index swap was set at a flat 6% a year, and a later check against the broker's actual charges suggests that was on the heavy side.

What I did not do

One number in the output passed the line: if each RSI(2) trade is scaled to one third, so that its 3-sigma stop equals 1R like the other legs, the ratio is 1.53. That is a reasonable way to size a portfolio, and I may well adopt it in a future test. But the pre-registration listed it as a reference figure only, and promoting it now would mean choosing the weighting after seeing that the equal-weight version failed. Over the long window it is 1.15, and I never ran its discounted version. Reporting 1.53 as the result would be the same mistake as reporting 1.46: picking the ruler that gives the better number.

Removing the RSI(2) leg entirely does not help either. The four remaining legs give 1.07 (by coincidence the same figure as the original bundle marked to market). The bundle needs that leg's returns, and that leg's returns came with the open risk the first pass did not count.

What happened next

Later the same evening I kept the 1.33 line and looked for additional legs, using a further pre-registered test: nine candidate legs, with a multiple-testing correction. Two more legs passed that screen. The seven-leg bundle reaches 1.57 with open trades counted. I am not treating that as a pass in any meaningful sense yet. One of the two new legs is negative on an independent older data period. The other is a pair of markets I picked in August after looking at results. The first half of the window is still weak at 0.62, the discounted version is 0.35, and the window from 2012 is 1.21.

I then recomputed the bundle from price data for 24 July to 10 November, which lies after the window the bundle tests used. It is not a clean forward test: most of the original leg studies had already seen part of that period. The result was 140 trades and +4.71R. The two legs whose studies had never seen those months made −10.77R over 64 trades; the other five made +15.48R. A bundle with zero edge would do at least as well as +4.71R in about 38% of 110-day windows.

The bundle now runs on a demo account at minimum size, and its job there is to measure execution costs, not to prove an edge. By my own estimate for the five-leg version, telling it apart from zero would need about 23 months of forward data, and the seven-leg version would still need well over a year.

The checklist I am keeping

• Never compare a strategy with a stop and a strategy without one on closed-trade drawdown. Mark both to market, at least at the daily close.
• When a portfolio passes, ask which leg would change most if its open trades were counted. If one leg has neither a stop nor a time limit, that leg is the answer.
• Write the fix before measuring it, try exactly one version, and say how many versions you tried.
• When the fixed version fails, report the failure. A different weighting found after the fact belongs in a new test, not in the old result.
• Count how much risk is open at the same moment. Correlated legs can give you several positions at once even when each leg looks small on its own.

For readers who build or buy EAs: when a report shows a smooth balance curve from a strategy that rarely closes a loser, look at the equity drawdown, not the balance drawdown. In my case, that one change moved the result from passing to failing.

Previous entries in this journal are on my blog page. If you trade by hand under a daily loss rule, my free Prop Firm Risk Guardian is a display-only panel that tracks the day's loss on equity, open positions included, against a daily loss limit and a maximum drawdown limit.