Winservers.NET

StrategyQuant X · Workflow

StrategyQuant X Workflow: A Real Automated Pipeline of 15 Projects

Most workflow guides stop at a diagram. This one shows a pipeline that actually ran: 15 custom projects handing over to each other on one server for five weeks, with the time of every step taken from the SQX logs.

Quick answer: each cycle builds until 400 strategies are accepted, filters them, runs OOS, Trade Manipulation, SPP, Randomize Data, Walk Forward and a tick retest, saves the survivors and restarts the Builder. The median cycle took 23 minutes. After about 10 cycles the project starts the next one. 100 of 106 handovers were automatic.

One cycle

The twelve steps of every cycle

The run lasted from August 4 to September 9, 2026 on an Intel Core i9-13900 server with 16 threads in use, 64 GB DDR5 and NVMe storage, StrategyQuant X build 144. The cycle below repeated 944 times. Task names are given as they appear in the SQX project logs.

StepTask in the SQX logWhat it doesAverage time per cycle
1Clear databanksStart every cycle from an empty databankunder 1 s
2BuilderGenetic generation until 400 strategies are accepted12.1 min
3Keep Quality (custom analysis)Keep the better part of the 400 accepted strategiesunder 1 s
4OOS retestCheck the out-of-sample part0.2 min
5Keep QualitySecond quality filter before the heavy checksunder 1 s
6Trade Manipulation retestMonte Carlo on the list of trades2 s
7SPP retestSystem Parameter Permutation4.8 min
8Randomize Data retestMonte Carlo on altered price history16.5 min
9Walk Forward retestRe-optimization on rolling windows9.5 min
10Tick RetestOne more backtest on tick precision0.8 min
11Keep QualityFinal quality filter, survivors go to final storageunder 1 s
12Clear databanks, Restart BuilderEmpty the working databanks and jump back to step 2under 1 s

The average cycle spent 12 minutes in the Builder and 32 minutes in checks. In 934 of 944 cycles the Builder stopped at exactly 400 accepted strategies, so the batch size was fixed and the cycle time depended on how many candidates survived the filters.

How long one cycle took

Fastest quarter of cycles, up toMedian cycleSlowest quarter, fromLongest cycle
6 min23 min51 min14.1 h

Why cycles differ so much: a cycle where few strategies pass OOS skips the heavy checks and ends in minutes. A cycle with many survivors waits for Randomize Data and Walk Forward. Time per check is in How long Walk Forward and Monte Carlo take.

The chain

How 15 projects take turns

The projects are the same pipeline with different templates: four timeframes, three trade directions and two exit types. They ran in a fixed order through the list and then started again. One turn of a project took about 5 hours and 10 Builder cycles (median). At the end of a turn the project filters its results, clears the databanks and a Stop & Start task launches the next project.

Project templateTurns in the chainBuilder cyclesHours in turnsStrategies generated
1. H1, long and short, end of day exit86056.32,117,405
2. H1, long and short, end of Friday exit76544.71,147,039
3. H1, long only, end of day exit75170.82,726,234
4. H1, long only, end of Friday exit87164.22,958,985
5. H1, short only, end of day exit87129.34,899,927
6. H1, short only, end of Friday exit76932.76,425,661
7. H2, long and short, end of Friday exit76546.61,366,187
8. H2, long only, end of Friday exit76351.74,504,659
9. H2, short only, end of Friday exit98635.312,967,267
10. H4, long and short, end of Friday exit86633.71,741,397
11. H4, long only, end of Friday exit75455.621,045,844
12. H4, short only, end of Friday exit88860.960,518,036
13. M30, long and short, end of day exit54343.51,865,980
14. M30, long only, end of day exit53851.43,121,928
15. M30, short only, end of day exit55434.43,340,508

In five weeks the chain went around the list about seven times. Turns are not perfectly even: the chain was restarted a few times and did not always resume from the same project, so the M30 projects at the end of the list got five turns each.

What splitting into projects gives

Each project keeps one timeframe, one direction and one exit rule. Results stay comparable, a slow template cannot starve the others, and you can see which template produces survivors.

What rotation gives

Short turns give every template fresh results every day or two. If the run is interrupted, no template is left without data.

Reality check

How unattended was it really?

The handover worked 100 times out of 106. The other 6 turns ended without a Stop & Start: in the middle of a Builder run or a retest. Between turns there were 12 pauses longer than an hour, 148 hours in total, the longest 63 hours. These pauses are most of the gap between 692 working hours and 860 calendar hours. The logs do not say why each pause happened.

What to prepare for

  • Check the chain once a day: is the current project the one you expect?
  • Decide who restarts the chain after a Windows update or a stopped session.
  • Keep final storage outside the databanks that the cycle clears.
  • Leave disk space for logs and results of weeks of work.

What the server must do

  • Stay on for weeks. A home PC that sleeps or restarts breaks the chain.
  • Keep the Remote Desktop session alive after you disconnect.
  • Give SQX all its threads. Do not run live trading terminals on the same machine.

Design lessons

What this pipeline gets right

A fixed batch from the Builder

Stopping at 400 accepted strategies keeps every cycle the same size. You can compare cycles and estimate the next day's output.

Quality filters between stages

Only 18% of accepted strategies went on to OOS, and only 14% of OOS survivors reached the heavy checks. The expensive tests see a short list.

Cheap checks before expensive ones

OOS and trade manipulation cost under a second per strategy and run first. SPP, Randomize Data and Walk Forward cost 39 to 136 seconds and run on what is left.

Clean databanks every cycle

Each cycle starts from empty working databanks, so results of one cycle do not mix with the next. The run never ran out of memory with 56 GB given to SQX.

The full path from 130.7 million generated strategies to 669 final ones is in How many strategies do you need to generate?

After StrategyQuant X

Treat generation, validation and MT5 testing as different jobs

A common mistake is to let one workstation do everything: generate thousands of strategies, validate them, export them, and run final MT5 tests on the same box. That is convenient at first, but it makes logs harder to track, increases queueing delays, and blurs the difference between discovery metrics and final acceptance metrics.

StageMain goalBest place to run itWhat to avoid
GenerationCreate many candidates quickly and remove obviously weak ideas.Dedicated CPU server or larger research box with enough RAM and NVMe.Judging final production quality from broad search output alone.
ValidationApply stricter filters and verify robustness on the shortlist.Separate project folder or second machine so exported results stay clean.Reusing loose generation thresholds as final pass rules.
MT5 testingConfirm behaviour in MetaTrader Strategy Tester with your target execution setup.Local MT5 master terminal plus remote agents or dedicated MT5 testing infrastructure.Skipping MT5 just because StrategyQuant metrics looked good.

StrategyQuant X can reduce the candidate set, but final MT5 testing is still useful because it checks the shortlisted strategies inside your MetaTrader Strategy Tester environment. If MT5 results diverge from StrategyQuant expectations, inspect symbol settings, spread assumptions, data scope and execution model.

Checklist before you trust a StrategyQuant X export

  • Check that the strategy came from the intended data range and template.
  • Confirm your robustness filters were stricter than the generation filters.
  • Remove duplicate logic variants that differ only by minor parameter noise.
  • Keep notes on why a strategy passed, not only that it passed.
  • Run MT5 tests with the target symbol settings and clear report naming.
  • Archive accepted reports separately from research noise.

Troubleshooting

Common workflow problems and what they usually mean

Generation is fast but validation becomes a backlog

That usually means your shortlist is still too wide, your filters are too loose, or your validation machine is sharing too many other tasks.

MT5 results do not match the StrategyQuant X shortlist

Check data scope, symbol properties, spread assumptions, session handling and whether the export rules changed between runs.

One server feels overloaded

Separate research from final MT5 testing first. If the bottleneck stays, move generation onto a stronger dedicated machine.

Reports are impossible to audit later

Use batch labels, consistent folder structure and saved filter presets. Without that discipline, repeated runs become hard to compare.

Hardware

When a VPS is not enough for this workflow

A standard VPS is often fine for light StrategyQuant X work, one or two MT5 terminals, or basic review tasks. It stops being ideal when you turn it into a heavy research node and a final testing node at the same time. A pipeline like the one above needs a machine that works for weeks with all its threads: a dedicated server. A small MetaTrader VPS is still the right place for live terminals. If your final MT5 stage becomes the main bottleneck, move to the MT5 backtest farm model with remote agents or a dedicated MT5 optimization server.

Sizing by the numbers: StrategyQuant X system requirements and the StrategyQuant X benchmark.

Ask which server fits my pipeline

FAQ

Frequently asked questions

What does an automated StrategyQuant X workflow look like?

In the pipeline we measured, every cycle was: clear databanks, Builder until 400 strategies are accepted, a quality filter, OOS retest, a second quality filter, then Trade Manipulation, SPP, Randomize Data, Walk Forward and tick retests, a final quality filter, and a jump back to the Builder. The median cycle took 23 minutes.

How do you chain several StrategyQuant X projects?

Each project ends with a Stop & Start task that launches the next project. In our run 15 projects took turns in a fixed order, about 10 Builder cycles and 5 hours per turn, and went around the full list about seven times in five weeks.

Can a StrategyQuant X pipeline run unattended for weeks?

Mostly. In our run 100 of 106 project turns ended with an automatic handover. 6 ended without one, and there were 12 pauses longer than an hour, 148 idle hours in total. Plan for monitoring and for restarting the chain after an interruption.

Why separate generation, validation and MT5 testing in StrategyQuant X?

Separating generation, validation and MT5 testing keeps discovery work, robustness review and final MetaTrader checks from contaminating each other. It makes results easier to audit and helps you see where compute bottlenecks actually are.

Should final testing stay in MT5 if StrategyQuant X already validates strategies?

Yes. StrategyQuant X can reduce the candidate set, but final MT5 testing is still useful because it checks the shortlisted strategies inside your MetaTrader Strategy Tester environment and remote-agent workflow.

When should I move from a VPS to a dedicated server for StrategyQuant X?

Move from a VPS to a dedicated server when large generation batches, validation jobs and MT5 tests start competing for the same CPU, RAM or NVMe resources. Dedicated hardware gives you a cleaner split between research stages.

When does an MT5 backtest farm make sense in this workflow?

An MT5 backtest farm makes sense when final MetaTrader testing becomes the main bottleneck and you want your local MT5 terminal to stay the master while remote agents process larger optimization or validation workloads in parallel.

Need help splitting StrategyQuant X research from MT5 testing?

Describe your current batch size, number of MT5 terminals and whether you already use remote agents. We can point you to the right path: VPS, dedicated server or MT5 backtest farm.

Ask in WhatsApp or Telegram

K.S.M. Trade Sp. z.o.o. | NIP/VAT PL5273128546 | REGON 52957272900000 | Warsaw, Poland