Risk warning: software only, not investment advice. Trading involves substantial risk, including loss of capital. Results shown are simulated and not reliable indicators of future performance. Full disclosure
← All articles

MT5 Strategy Tester Optimization: Fix Broken Runs

9 November 2026 · 10 min read

Key takeaways

  • Empty tester reports usually trace to missing history, a date range outside downloaded bars, or an EA that fails its OnInit check silently.
  • Real-tick mode is for validation passes; synthetic every-tick and control points are coarse filters unsuitable for intrabar EAs with tight stops.
  • Idle or crashed optimization agents are usually caused by zero agent count, DLL imports blocked in the sandbox, or RAM exhaustion in real-tick mode.
  • Forward testing with a chronological 30% holdout is the cheapest out-of-sample check; discard passes whose forward profit factor collapses.
  • Choose drawdown-weighted or custom OnTester criteria over max balance, and look for parameter plateaus rather than isolated peaks.

MT5 Strategy Tester Optimization: Fix Broken Runs and Get Results You Can Trust

MT5 strategy tester optimization fails for a handful of predictable reasons: empty reports, agents stuck idle, or crashes mid-run, and each symptom points to a specific fix. This guide walks through that decision tree first, then covers tester modes, optimization criteria, and forward testing so your backtests survive real capital rather than curve-fitting the past.

Why is the MT5 Strategy Tester not running or showing no results?

Nine times out of ten the cause is one of four things: no downloaded history, a mismatch between the EA's symbol/timeframe and the chart, a testing date range outside available data, or an EA that fails its OnInit check and exits silently. Match your symptom below before touching settings.

Symptom: report is empty or the tester stops instantly

Check the Journal tab first, not the Results tab. "There is no testing date range specified" or "no history" means the date window sits outside your downloaded bars. An instant stop with "EA does not trade" is often a licence key or account check blocking execution in tester mode. The MetaQuotes Strategy Tester documentation lists the required settings for a valid run.

Symptom: 'no history' or wrong symbol data

Journal messageLikely causeFix
no history for symbolBars not downloadedOpen the symbol's chart, scroll back to load data
wrong volume / invalid priceGaps in tick dataRe-download M1 history, then retest
test stopped, no dataDates before available historyMove start date forward

Symptom: EA loads but never trades in the tester

Verify the EA is attached to a symbol/timeframe combination it actually supports, that AutoTrading restrictions in its inputs are off, and that the modelling mode supplies enough data. If an EA only trades on M15-H4 symbols like indices and metals, testing it on an exotic pair produces a clean but meaningless empty report. This is the same discipline we apply to every bot we document at BotGenomix pricing: no result is published without a reproducible tester run behind it.

Which tester mode should I use: every tick, control points, or real ticks?

The three modes answer different questions. "Open prices only" simulates one price per bar, which is only valid for EAs that trade on bar close. "Every tick" generates synthetic ticks from M1 bars, and "control points" interpolates even more coarsely between them. For an EA that trades intrabar on M15 or faster, synthetic ticks are a guess about what the market did, and the guess gets worse the tighter your stops and take-profits are.

What 'every tick based on real ticks' does and costs

Real-tick mode replays recorded ticks from your broker's server, so spread, slippage windows and tick sequence match live conditions as closely as the tester allows. On US30 or XAUUSD with a 20-pip stop, that difference can flip a losing backtest into a winning one or the reverse. The costs: downloads are slow, history coverage depends on your broker, and optimisation over real ticks is impractical, so use it for validation passes, not parameter sweeps. At BotGenomix, every published result runs on real ticks first.

When control points are safe (and when they lie)

Control points are safe only when your EA evaluates prices once per bar and holds positions for many bars. If it reads the current tick price, uses tight stops, or scales position size off intrabar movement, control points and even synthetic every-tick mode will fabricate fills that never happened. Treat any result from these modes as a coarse filter, then re-run the surviving parameter sets on real ticks before you trust the profit factor or drawdown figures.

How do I fix optimization agents that crash or stay idle?

If agents sit idle, the usual culprit is the tester itself: open the Journal tab, check "Optimization" is ticked, and confirm the number of agents isn't set to zero or capped below your core count. Crashed agents show up in the journal as "tester agent stopped" or "no connection" entries.

Reading the agent journal and tester logs

Each local agent keeps its own log under %APPDATA%\MetaQuotes\Terminal\<hash>\Tester\<agent-name>\logs. Read that file first; the terminal journal only shows the summary, while the agent log names the actual fault, usually a DLL call, an array out of range, or a memory failure. The MQL5 documentation covers agent logs and custom OnTester criteria in detail.

Common crash causes: DLLs, out-of-memory, bad inputs

Three causes cover most failures. First, DLL imports: agents run in a sandbox, so an EA calling DLLs without "Allow DLL imports" dies on every pass. Second, real-tick mode plus many agents exhausts RAM; reduce agents or test fewer symbols per run. Third, bad input ranges, like a stop-loss of zero or a negative lot, throw division errors that kill passes silently. Log the offending inputs inside OnTester so failing passes are identifiable, the same discipline we document for every BotGenomix EA at BotGenomix MT5 Expert Advisors.

What optimization criterion should I choose instead of default balance?

Max balance finds parameter sets that happened to sit on a lucky equity path. Two runs on adjacent inputs can differ by thousands in profit while sharing the same trade logic, so the winner is usually curve-fitted. Balance + max profit factor and the built-in complex criterion both penalise that fragility by weighing trade quality, not just the final number.

Custom criterion via OnTester

Write your own metric in OnTester and MT5 ranks every pass by it. A common starting point: profit factor multiplied by the square root of trade count, discounted by recovery factor below a threshold. This lets you encode what matters on M15-H4 indices and metals, where a few large trades can distort raw balance rankings.

Why drawdown-weighted criteria survive live trading better

Dividing profit by max drawdown, or rejecting passes above a fixed relative drawdown, biases selection toward parameter sets that degrade gracefully rather than explode. We publish every bot's month-by-month drawdowns and losing months at the BotGenomix demo track record for exactly this reason: a backtest that hides its worst stretch tells you nothing about live behaviour.

How do I use forward testing to avoid curve-fitted results?

Choosing a forward fraction and date split

Set the forward period in the tester's Optimization tab as a fraction of your backtest: 0.3 means the last 30% of your date range is held out and never touched during optimization. For M15-H4 index and FX strategies, aim for at least six months of forward data; one month proves almost nothing. The split must be chronological, not random, so the forward window contains market conditions the optimizer never saw.

Reading the forward column in the optimization table

The forward test is the cheapest out-of-sample check you own, and it works only if you treat it as a verdict rather than a bonus. Every pass in the optimization table gets two sets of numbers: backward results from the data the optimizer trained on, and forward results from the held-out window it never saw. A parameter set that is genuinely fitted to market behaviour, not to noise, should show a forward profit factor and drawdown in the same ballpark as its backward numbers. When the forward column collapses, a profit factor dropping from 2.1 to 0.8, or a drawdown tripling, you are looking at curve-fitting, and that pass should be discarded no matter how good the backward half looks.

Sort by the forward column instead of balance, and be suspicious of any pass where the two halves diverge widely. MetaQuotes documents how forward mode runs as part of optimization in the MQL5 documentation. At BotGenomix, every published bot includes a separate 2025+ unseen-data result alongside its month-by-month backtest, clearly labelled simulated, because a forward split alone is still one sample; the licence terms at BotGenomix legal terms spell out what those results do and do not claim.

How do local vs cloud (MQL5 Cloud) optimization agents differ in cost and speed?

Local agents use your own CPU cores, so speed is capped by your hardware. A 12-core machine runs 12 agents, and a multi-thousand-pass real-tick optimization on M15-H4 symbols can take many hours or days. Cloud agents rent additional cores from MetaQuotes' network, cutting wall-clock time roughly in proportion to how many agents you add, but each agent bills per minute of use, so a full real-tick run across thousands of passes can cost tens of pounds or more. Exact rates are shown in the terminal's Agents tab before you start.

When cloud agents are worth paying for

Pay for cloud agents when a run is a one-off and time-sensitive: validating a final parameter set before deploying a licence key, or testing a bot across several symbols overnight. For iterative work, prune the parameter space first, run every-tick or control-points mode locally, and reserve real-tick cloud runs for the shortlist. Repeated cloud runs on unpruned grids are the fastest way to burn money on a curve-fitted result.

How do I interpret the optimization results table and 2D/3D graphs?

Plateaus beat peaks: reading parameter neighbourhoods

Sort the results table, then ignore the top row. The top result is often a lucky spike; what you want is a cluster of nearby parameter sets with similar profit factors and drawdowns. If a 10% shift in a stop or period halves the profit factor, that single pass is curve-fitted noise, not an edge. A plateau means the logic works across a range of conditions, which is what survives live spreads and slippage.

Red flags in the 3D surface

Switch to the 3D graph and look at the shape, not the height. Isolated sharp peaks, knife-edge ridges, and surfaces that collapse to a cliff one grid step away all signal overfitting. Flat or gently rolling regions are your candidate zones. Also check the forward-test column: passes that only win in-sample are the classic trap, and the MetaTrader 5 platform site notes that modelling quality and real tick availability materially change how trustworthy any of these passes are.

How does BotGenomix run its own tester optimization, and where does it fall short?

Every BotGenomix bot publishes its full month-by-month backtest: drawdowns, losing months, profit factor and win rate, plus robustness checks across parameter ranges rather than a single cherry-picked set. Each bot also carries a separate 2025+ unseen-data result, labelled simulated, so you can see how the parameter set behaved outside the optimisation window. Everything is sold as a flat monthly licence (Super £149/mo, Expert £249/mo) with no performance fees, and the .ex5 files run on your own terminal and VPS.

Where it falls short: the out-of-sample check is a single split, not a rolling walk-forward. A 2025+ window is one unseen period, so it can't rule out regime-specific luck the way rolling forward testing across multiple windows can. Results are also simulated, not live-forwarded, and backtests are only as good as the tick data behind them. If your standard is walk-forward validation before capital goes at risk, run that yourself in the Strategy Tester using the settings from the sections above before buying either tier.

FAQ

Why does my MT5 Strategy Tester show no trading results even though the EA compiles?

Compilation only proves the code runs. The tester still needs downloaded history matching the EA's symbol and timeframe, a date range within that data, and an OnInit check that passes. An EA with a licence or account check can exit silently, and an EA attached to an unsupported symbol produces an empty report with no errors.

How many optimization passes are too many for curve fitting?

There's no fixed number; the risk comes from passes relative to independent price behaviour. A few thousand passes over a narrow date range on one symbol will find lucky equity paths even with sensible criteria. Watch the neighbourhood instead: if adjacent parameter sets diverge wildly in profit factor or drawdown, your grid has found noise rather than an edge.

Is genetic optimization better than full slow complete optimization in MT5?

Genetic search is faster because it skips most of the grid, which matters on large parameter spaces or real-tick runs where complete optimization is impractical. The trade-off is that it can miss isolated strong parameter sets. Use genetic mode to prune the space, then run complete optimization or real-tick validation on the surviving candidates before trusting any ranking.

Can I trust a backtest with 99% modelling quality?

Not on its own. Modelling quality measures how closely the tester reproduced tick data, so 99% under every-tick based on real ticks is meaningful, while a high percentage under synthetic modes can still fabricate fills for intrabar EAs with tight stops. Check which mode produced the figure, then verify the parameter set survives a held-out forward window before risking capital.

How much MQL5 Cloud computing cost is a typical optimization run?

Cloud agents bill per minute, so cost scales with passes, mode, and agents rented. A pruned every-tick run costs little; a multi-thousand-pass real-tick optimization across several symbols can run to tens of pounds or more. Exact rates appear in the terminal's Agents tab before you start, so estimate there rather than treating any quoted figure as universal.

Important: BotGenomix provides downloadable automated trading software for independent use. It does not provide personalised investment advice or manage customer trading accounts. Trading involves substantial risk, including the potential loss of capital. Past performance and backtested results are not reliable indicators of future performance. Full risk information