How to Run a Backtest in MT5 (and Which of the Report's Two Drawdown Numbers to Believe)

You run a backtest in MT5 by opening the Strategy Tester with View > Strategy Tester (or Ctrl+R), choosing the Expert Advisor, the symbol, the timeframe, the date range and the modelling mode, then pressing Start. Five minutes later you have a report. The pressing question is not how to get there — it is which of the numbers on that report mean what you think they mean.

This guide does both halves: the procedure, and the reading. The reading is where the evidence on this page comes from, because we have 23 finished runs on file and their reports disagree with themselves in a way almost nobody checks.

What you get out of a finished run

The artifact is a report with two halves: a result block (trades, profit factor, drawdown, net) and the trade list behind it. Both matter, and only the first one gets quoted.

What you should be able to say at the end is not “it made $3,023”. It is: this rule set, on this symbol and timeframe, over this window, with this tick model, on a $10,000 account, produced this many trades, at this profit factor, having been this far underwater at its worst. Every clause in that sentence is a setting you chose, and changing any of them changes the answer. That is why the settings come before the procedure here.

Setup: the data and settings that decide the answer

Here is one real run we ran and published, quoted so you can hold the inputs and the output side by side. It is the record behind Stillwater, a long-only USD/JPY strategy on the 5-minute chart.

Setting	This run	Why it is set that way
Expert	Stillwater (19 inputs)	Every input the report was verified to carry — the surface you would have to re-enter to reproduce it
Symbol	USDJPY_DUKA	A dedicated tick-history symbol, kept separate from the live-quoted one so the run is repeatable
Period	M5	The chart the rules evaluate on; the trend filter reads M30 from inside the EA
Date range	2019-01-01 → 2026-07-01	7.5 years, long enough to contain more than one regime
Modelling	Every tick based on real ticks (Model 4)	Anything that trades inside the day needs the intra-bar path measured, not interpolated
Deposit	10,000 USD	Fixed across every run here, so drawdown percentages stay comparable
Leverage	Requested 1:500 — ran at 1:100	See below. This one is not what it appears to be
Optimisation	Disabled	A single pass. Searching parameters is a different job with different traps

Two of those rows deserve more than a table cell.

Data quality is the setting the result rests on. Every other figure in the report is computed on top of the price series the tester used, so the modelling quality figure bounds how much the rest of the report is about your strategy at all. The published record for the run above states 100% history quality over 229 million real ticks, read straight off the tester’s report. A run on interpolated bar data would have produced a different trade set, not merely different fills — the tester assumes a smooth path between prices it knows, and a strategy with a tight stop trades that assumption. If you have not read what tick data your terminal actually holds for the symbol, that is the first thing to check, before the strategy.

The leverage field is a request, not a setting. In all 23 runs on file the tester was asked for 1:500 and the test ran at 1:100 — 0 of 23 times did the requested figure take effect. That is not a guess from reading the field back: it was pinned by measuring the margin a position actually consumed inside the test, 126.39 USD, which matches the leverage-100 calculation and not the leverage-500 one. The practical consequence is that a margin-sensitive design — a grid, a martingale, anything that stacks positions — can pass a backtest at a leverage it will never see, or fail one at a leverage it would not have used. Check the margin figure in the report against your own arithmetic rather than trusting the box you typed in.

Step-by-step: run the test in MT5

• Check what history the symbol actually holds
• Choose the modelling mode to match the holding time
• Set symbol, period and an explicit date range
• Read every value on the Inputs tab
• Run one pass with optimisation disabled
• Save the report, not a screenshot
• Read both drawdown lines before the profit line

Reading the results: which numbers the report actually means

Take the same run’s output. This is what came back:

Result	Value	What it has to clear before you continue
Total trades	416	Enough that the average is describing a sample, not an anecdote
Profit factor	1.28	Above 1.0 by a margin that survives costs you did not model
Net profit	3,023.01 USD	On a 10,000 deposit — a return, not a score
Balance drawdown	5.82% (747.57)	The worst closed-trade dip
Equity drawdown	6.11% (787.60)	The worst dip including open positions — the one you would have watched
Win rate	60.82%	Meaningless alone; read it against the stop and target, 100 and 200 pips here

Test conditions
Experiment ID	EXP-BACKTEST-RUNBOOK-001
Run ID	stillwater__gate_v2_duka_realticks__2019-01-01_2026-07-01
MT5 build	6140
Contract specification	Exness-MT5Trial5
Symbol / timeframe	USDJPY_DUKA M5
Period	2019-01-01 – 2026-07-01
Model	Every tick based on real ticks (Model 4)
Deposit	10,000 USD
Leverage requested	1:500
Leverage effective	1:100
Measured	2026-08-23
Last verified	2026-08-25

Three things in that report are routinely misread.

The two drawdown lines are different questions. Balance drawdown measures the closed-trade equity curve; equity drawdown includes the float on positions that were still open. Across our 23 runs the equity line was deeper every time, and the size of the gap is not predictable from the strategy’s description: the narrowest was 0.04 points (0.67% against 0.71%) and the widest was 11.10 points (14.52% against 25.62%, a ratio of 1.76). A strategy that closes fast has a small gap. One that sits through adverse movement has a large one, and that is precisely the strategy whose drawdown you most need to know before you fund it.

MT5 prints each drawdown twice, under two definitions. Maximal drawdown is measured at the moment the amount was largest; Relative is measured at the moment the percentage was largest. They are different moments in the same run, so the two figures are not two views of one event. Of our runs, 16 are recorded at the Maximal definition and 5 at Relative — and the pipeline that writes those records refuses to pair a Maximal balance figure with a Relative equity one, because a page showing both would be comparing two different instants. When you quote a drawdown, name the definition with it.

The report’s own totals are rounded, and its window may not be the window you mean. Recomputed from the published trade list, the same run’s profit factor is 1.2843 rather than the 1.28 in the report — harmless here, and not harmless when you are comparing two candidates three decimal places apart. More importantly, that 416-trade run splits into 404 trades up to the 2026-04-01 publication cutoff and 12 after it. The later 12 returned a profit factor of 2.3317. Neither number is wrong; they answer different questions, and the cutoff is a date a human chose, not something the tester printed. If you are going to divide a run into “the part I fitted on” and “the part I did not”, write down where the line is before you look at the second half. Reading an MT5 backtest report does that division across all 23 runs and shows how far the headline moves.

Pitfalls that quietly invalidate a run

Improve it: change one thing, rerun, compare

The loop is deliberately boring. Change one variable, rerun, and write down both results. Two variables at once and you cannot attribute the difference to either.

The order that gets you the most information per run is:

1. Vary the window before you vary the strategy. Run 2019–2022 and 2022–2026 separately. If the two halves disagree, the parameters are describing one period, and no amount of tuning fixes that.
2. Vary the symbol. A rule that only works on one pair is a fact about that pair. Our set includes the same broad approaches across USD/JPY, EUR/GBP, AUD/CAD, NZD/USD, an index and a crypto pair for exactly this reason.
3. Vary the modelling mode. Rerun the surviving candidate on a coarser model. If the result collapses, the edge lived inside the bar, and the intra-bar path is the least trustworthy part of any simulation.
4. Only then vary parameters — and treat any setting you had to search for as a weaker claim than one you chose in advance.

Stop iterating when a change stops moving the result outside the noise of the previous runs, or when you notice you are choosing settings because of what the last report said. That second one is the real stopping condition, and it arrives earlier than people expect. The backtest analyzer is useful here because it reads a saved report’s headline figures back out, which makes comparing two runs an act of arithmetic rather than memory.

Next steps: forward test before you go live

A single clean backtest is the first of four steps, not the finish line.

Optimise, if the rule has parameters worth searching. Record the whole search range, not the winning pass — a result you had to look for is evidence about the search, and the range is what tells a reader how hard you looked. Parameter optimisation walks that pass and reads our own eight recorded searches: 116 combinations, and not one setting we were willing to ship.

Then walk forward. Fit on one segment, measure on the next, and repeat. This is the only one of the four steps that directly attacks overfitting, which is why it is worth more than a longer backtest.

Then forward test on a demo account. This is where latency, requotes and real spread widening finally enter the experiment. Run it long enough to watch the strategy lose and recover, because the number you actually need is not the return — it is whether you can sit through the drawdown you measured above.

Then decide. Write down, in advance, the loss that would make you switch it off. A stop you set before funding is a decision; one you set during a drawdown is a reaction.

If you want a worked reference before running your own, every listing in the published EA catalogue carries its complete closed-trade list, both drawdown lines and the run manifest behind them, and how we test states the standard those records are held to. The rest of the backtesting guides go deeper into the optimisation and walk-forward steps this page only points at. Nothing here promises a result: every figure on this page describes runs that have already happened.