Same EA, two platforms: the drawdown gap we measured on gold

We wanted to stop paying for a tick data subscription. The plan was simple: check that the native MT5 tester gives us the same answer as our MT4 setup, then switch and save the money.

It does not give the same answer. And the difference is not random noise, which is what makes it worth a post.

On six gold strategies tested under aligned conditions, the two platforms agreed on how many trades happened and on how much gross profit those trades made. They disagreed on the losses. The maximum drawdown reported by MT5 was lower in 6 cases out of 6, between 22 and 55 percent lower, roughly 35 percent lower on average.

If you size your risk from a backtest, that is the single number you cannot afford to have flattered.

Correction added on 24 August 2026, and it cuts against this article. A later review found that the comparison above put the MT4 drawdown against MT5's balance drawdown. MT5 reports two counters, balance and equity, and the equity one is the closer match to what MT4 reports. Redone against MT5's equity drawdown, the gap shrinks from roughly 35 percent to roughly 15 percent on average, and it is no longer 6 out of 6: three of the six land within about one percent of the MT4 figure, and one comes out 5 percent higher. That three of them match the equity counter almost exactly is itself evidence that the equity counter is the right one to compare. We are leaving the original numbers below so the mistake stays visible, but the honest summary is narrower than the one we published: the two platforms do disagree on drawdown, the disagreement is smaller than we first reported, and part of what we measured was our own choice of counter. The rest of the article, trades and gross profit agreeing closely, is unaffected.

The setup, so you can attack the method

• Six of our own XAUUSD strategies: one M30, one H1, four H4. They are six of the 279 strategies that had both an MT4 report and a compiled MT5 version. Having both was the requirement to be eligible, not the rule that selected these six, and we did not record a selection rule: read the "6 out of 6" counts below with that in mind.
• Window 1 September 2021 to 10 April 2026, identical on both sides. Same symbol, same timeframe, 0.1 lots, 10,000 USD deposit. Our comparison script refuses to interpret the deltas if lot size, deposit or window do not match, because that mistake alone can produce a factor of ten.
• MT4: every tick, with real tick data from a commercial tick source, 262 million ticks, 99.90 percent modelling quality.
• MT5: model 0, every tick generated from the M1 bars, 156 million ticks. Broker data from the same broker group, but two different demo servers.
• Delta below means MT5 relative to MT4. For gross loss we report the change in the size of the loss, so a negative number means MT5 lost less money on its losing trades.

The measurements

Read the columns from left to right and the story tells itself:

• The strategies did the same things. Trade counts land within 12.5 percent, and gross profit within 8.5 percent in 6 cases out of 6. The entry signals agree. This is not a data problem at the bar level, not an indicator problem, not a different strategy running.
• The losing side is where they split. Gross loss is smaller on MT5 in 5 cases out of 6, and the maximum drawdown is smaller on MT5 in 6 cases out of 6, with no exception and no case going the other way.
• Net profit and profit factor inherit the difference. Profit factor rose on MT5 in 5 cases out of 6, and in two of them the net profit came out more than 50 percent higher than the MT4 figure of the same strategy.

All of these are backtest figures, on historical data, and past performance does not guarantee future results.

The obvious explanation, and why we had to drop it

The first guess is tick resolution: 156 million generated ticks against 262 million real ones. Inside a coarse bar the price path is invented, so a stop loss can get skipped or hit at a friendlier price, while a take profit sitting at a fixed target does not care. That would explain the whole pattern.

Instead of downloading four gigabytes of tick data per symbol to find out, we ran a cheaper test. On our disk we already had real ticks for a recent three month window. Same strategy, same window, same everything, changing only the tester model: generated ticks against real ticks.

The gaps here are net -4.2 and -2.1 percent, max drawdown +2.0 and +4.4 percent, profit factor -2.4 and -3.9 percent. The direction is the expected one, real ticks are slightly harsher, but the size is 2 to 4 percent, not 35. On that window the MT5 generator produced 43.2 million ticks against 44.2 million real ones, about 98 percent of them.

So on that window tick resolution does not account for the gap, which was worth knowing before spending a night on downloads. We are careful with how far this carries: it is a strong indication, not proof across the whole test period. It rests on three months, one symbol, one volatility regime and about 27 trades in total.

What we think is left, without pretending we know

Three candidates remain, and we have verified none of them yet:

• Different price history. Two different demo servers, two different builds, two different sets of stored bars and spreads. Nothing forces them to agree.
• Cost modelling. The MT4 run used historical variable spread from the tick source, the MT5 run used its own. A smaller gross loss on MT5 is consistent with cheaper simulated friction.
• Order execution. Netting against hedging, and how each tester resolves stop orders inside a bar. These six strategies enter with stop orders, which is exactly the order type most sensitive to that choice.

We are not claiming one platform is broken. We are reporting that two reasonable setups, given the same strategies, disagree systematically about the worst part of the equity curve.

What we do with it

We did not cancel the tick data subscription. We now use the two testers for different jobs: the more optimistic one as a fast, free pre-filter (if a strategy fails there, it will not pass the stricter one), and the stricter one as the source of truth for any number that ends up in a product description or in a risk calculation.

For anyone evaluating an EA, the practical version is shorter:

• A drawdown figure without its platform, broker, data source, spread model and date range is not a figure, it is a decoration.
• Never compare your own test to a seller's screenshot unless those five things match. Different numbers are the normal outcome, not evidence of a lie.
• When two tests disagree, size your account on the worse one. It costs nothing to be wrong in that direction.

You can reproduce the whole idea for free. Take Wonder 8 (free on MT5, also on MT4, USDJPY H1, fixed 0.01 lot, stop loss and take profit on every trade), run the same window on both platforms with the same lot and deposit, and compare the drawdown line. Whatever you find, that number is about your data and your broker, which is the point.

All figures in this article are backtest results measured on historical data, under the conditions described above. Backtests are hypothetical, past performance does not guarantee future results, and trading involves risk of loss.