From 4,720 MT5 Reports to Five Robust Candidates: How I Build Multi-Strategy EAs

Most traders show the final equity curve. I am more interested in everything that had to fail before that curve was allowed to survive.

My current MT5 research archive contains 8,404 candidate files. After parsing and classification, it produced 4,720 performance reports, 3,399 unique result signatures and 2,464 reports based on real-tick modelling. Only 1,027 results remained in my internal “publication candidate” category.

Those numbers are not proof of an edge. They show the size of the filtering problem.

The objective is not to find the backtest with the highest net profit. It is to build a portfolio of strategies that are individually understandable, sufficiently tested and different enough to avoid becoming the same trade under another name.

The real problem is selection

Generating a profitable backtest is easy when enough parameters, symbols and date ranges are tested. The difficult part is deciding what deserves another hour of research.

My first filter removes results that cannot be interpreted safely:

• duplicate reports and repeated result signatures;
• unclear modelling quality;
• incomplete or unusually short test periods;
• variants with too few trades;
• results dominated by one exceptional trade or one short market regime;
• configurations where equity drawdown is materially worse than balance drawdown;
• versions whose attractive return comes mainly from aggressive sizing.

This is why I maintain both a research archive and a smaller publication shortlist. A report can be interesting for development without being suitable for a public performance claim.

Five representative research candidates

The following results are selected historical MT5 simulations. They are not live performance, and they are not presented as a forecast.

All five reports used real-tick modelling in MT5. Returns are calculated relative to the initial deposit in each report. Costs and assumptions still need to be read from the original test configuration before attempting any independent reproduction.

These are some of the strongest headline results in the archive, but headline strength was not the only reason for selecting them. Each candidate also offered a useful combination of trade count, profit factor and controlled equity drawdown.

Why I do not choose one winner

A single strategy can look excellent because its preferred market behaviour dominated the test window. That does not make it useless, but it makes concentration dangerous.

My preferred direction is a multi-strategy architecture in which different modules can respond to different conditions. The important distinction is diversification by mechanism, not diversification by name.

Two systems are not meaningfully diversified merely because:

• they trade different symbols;
• they use different parameter values;
• one runs on M15 and another on M30;
• their full-period correlation appears low.

The more important question is what happens during stress. If several modules respond to the same volatility expansion, liquidity event or risk-off move, their apparently separate positions can become one concentrated exposure.

This is why I treat signal diversification and exposure diversification as different problems.

The portfolio risk layer

When several strategies run inside one MT5 environment, each module can be correct according to its own rules while the combined position is still too large.

The portfolio layer therefore has to operate above the entry logic. Depending on the system, I monitor or constrain:

• total exposure by symbol and currency;
• simultaneous positions created by different modules;
• correlated directional risk;
• portfolio-level floating drawdown;
• maximum active strategies;
• spread and execution conditions;
• recovery or averaging behaviour;
• the conditions required before trading can resume after a block.

This layer is not designed to improve every backtest. Its job is to prevent several reasonable decisions from combining into one unreasonable portfolio.

Robustness is a process, not a checkbox

I do not consider a strategy robust because it passed one real-tick test. Real-tick modelling is an important starting point, not a certificate.

My research process is built around repeated attempts to break the strategy:

1. Deduplicate the evidence. Multiple exported files can represent the same underlying result. Counting them as independent tests creates false confidence.
2. Check neighbouring configurations. A strategy should not collapse immediately when a reasonable parameter is changed.
3. Separate return from sizing. A higher lot size can make an ordinary strategy look exceptional while leaving its underlying quality unchanged.
4. Inspect equity drawdown. Balance-only statistics can hide the path of open risk.
5. Review the trade distribution. A large sample is not automatically good, but a small sample limits what can be inferred.
6. Test the portfolio, not only the modules. Correlation, shared exposure and simultaneous losses matter after strategies are combined.
7. Move to forward observation. Historical simulation is the research phase. Live monitoring is a separate phase with different evidence.

What these results do — and do not — show

The table demonstrates that the current research process can produce several strong historical candidates across EURJPY, AUDCAD and XAUUSD. It also shows that the archive is large enough to require strict filtering and careful disclosure.

It does not show that the same returns will continue. It does not eliminate execution risk, regime change, spread variation or overfitting. It does not turn a short public live record into long-term proof.

My public live track records are still young. For that reason, I prefer to describe them as live tracking rather than long-term validation. The purpose of publishing now is to document the transition from simulation to observation without pretending that the final answer is already known.

What comes next

The next stage is not to generate another thousand attractive curves. It is to strengthen the research funnel:

• preserve every test and configuration in a searchable archive;
• compare strategy families instead of isolated winners;
• run systematic robustness checks;
• monitor live behaviour against test assumptions;
• record execution differences and operational failures;
• publish both progress and invalidations.

The long-term goal is a system that improves its research process continuously — not a machine that only remembers its winners.

I will publish deeper breakdowns of the individual strategy families, their portfolio interaction and the gap between historical simulation and live execution as more evidence becomes available.

Disclosure: Historical simulation—not live performance. Results do not guarantee future performance. This article is for research and educational purposes and is not investment advice.