We've learned how to sell trading robots. But have we learned how to make them reliable?

We've learned how to sell trading robots. But have we learned how to make them reliable?

19 November 2026, 19:23
Roman Kondratev
0
5

These days, it's easier than ever to make a technical product look impressive.

With a few hours of work and some AI tools, almost anyone can create a beautiful logo, a professional presentation, an attractive trading dashboard, and a product description that sounds like it came from a major investment company.

Charts, indicators, animations, dozens of settings, impressive numbers.

Everything looks sophisticated, expensive, and reliable.

But I think there's a question we should be asking more often: are we buying actual technology, or just the feeling that we're looking at something advanced?

When marketing becomes more important than the product

I have nothing against marketing. A good presentation helps people understand a complicated product.

The problem starts when the presentation becomes more convincing than the evidence behind it.

A beautiful interface gives us a feeling of control. Lots of settings suggest flexibility. A perfect equity curve makes future profits seem almost inevitable.

But feeling in control isn't the same as being in control.

And looking reliable isn't the same as being reliable.

This is especially dangerous when we're talking about financial products, where a false sense of security can cost real money.

Today, marketing content can be produced almost endlessly. The internet is filling up with the same promises, the same beautiful pictures, and the same impressive trading charts.

There is so much of this material that it becomes increasingly difficult to recognize serious engineering work.

And that's a real problem for a new product with strong technology behind it.

You're no longer competing only with other algorithms. You're also competing with people's ability to create a convincing illusion of quality.

Past trades have already happened. We're investing in the future.

Imagine two trading robots.

The first has an impressive backtest, a beautiful equity curve, and an excellent presentation.

The second looks less exciting, but its developer can explain exactly what happens when trading costs increase, market conditions change, or the strategy encounters periods that are very different from those used during optimization.

Which one is actually better?

We don't know. Not without further testing.

Historical trades are events that have already happened. Every profitable trade belongs to the past, whether it was executed on a live account or simulated in the strategy tester.

Of course, a real trading history matters. It shows that the system actually operated under real market conditions, rather than only producing results in a simulation.

But even a verified live trading history cannot prove that the strategy will remain profitable.

As investors, we're not buying yesterday's profits. We're accepting the risks of tomorrow.

So perhaps the most important question isn't "How much did this robot make?"

It's "What happens to this strategy when things get worse?"

Don't just try to improve a backtest. Try to break it.

One way to understand the strength of a trading algorithm is to examine how sensitive it is to additional trading costs.

For example, we can test the same strategy under several conditions:

0 bps - no additional cost penalty.
1 bps - an additional cost of 0.01% of the traded notional value.
2 bps - an additional cost of 0.02%.

And then continue increasing the costs.

Naturally, we must define whether these costs apply per side or to a complete trade, and avoid counting existing commissions or spreads twice.

What interests me here isn't just the original profit.

I want to see how quickly the strategy deteriorates.

How fast does the Profit Factor decline? What happens to the expected return per trade? At what point does the strategy lose its statistical edge? How do drawdowns change?

A robot might produce fantastic results under ideal conditions and lose almost all its advantage after a relatively small increase in execution costs.

Another robot might have a less impressive initial result but remain profitable under much more difficult conditions.

Does this guarantee that the second robot will make money in the future? Of course not.

But it tells us something important about the robustness of its historical performance.

Personally, I'm not interested only in how high an equity curve can climb.

I'm interested in how difficult it is to destroy the advantage behind it.

Optimization should find robustness, not a perfect past

One of the biggest traps in algorithmic trading is finding parameters that perfectly explain historical data.

The more combinations we test, the more likely we are to discover an excellent result purely by chance.

That's overfitting.

It becomes especially dangerous when a developer repeatedly improves an algorithm while looking at the same historical price movements.

Eventually, the system may start memorizing the past rather than identifying repeatable market behavior.

This is why I've been exploring a different approach in my own work.

For RIZA, I use an optimization approach based on mixed historical quarters and Monte Carlo methods.

Instead of relying entirely on one continuous historical period, the idea is to construct different combinations of quarterly market data and examine how stable the strategy parameters remain across those combinations.

The goal is to reduce dependence on one particularly favorable period or one specific sequence of market conditions.

There is an important limitation, however.

Monte Carlo cannot tell us what the future will look like. Rearranging historical periods does not create genuinely new market information, and optimizing repeatedly on the same data is not equivalent to testing on unseen data.

That's why independent validation periods and tests with less favorable execution conditions are still necessary.

I don't see this approach as a way to make a trading system invulnerable.

I see it as a way to understand its weaknesses before the market exposes them.

Why is real engineering so difficult to sell?

Because a beautiful picture takes one second to understand.

Understanding the quality of an optimization process takes considerably more effort.

It's much easier to show someone a spectacular profit chart than to explain Monte Carlo distributions, parameter stability, out-of-sample validation, or sensitivity to trading costs.

Marketing naturally focuses on whatever is easiest to demonstrate.

Engineering focuses on problems that are much harder to explain, measure, and solve.

This creates an uncomfortable situation for developers of genuinely technical products.

Should we spend more time making our products look impressive, or more time investigating whether the algorithms actually deserve that impression?

Obviously, good technology also deserves a good presentation.

But I think the order matters.

First, build the technology. Then test it properly. Then present it.

Not the other way around.

What I would suggest to investors

Don't trust beautiful pictures simply because they look professional.

Don't ignore historical results either. They are important, but they are not forecasts.

Don't confuse a complicated dashboard with a sophisticated trading model.

Ask developers how their strategies were optimized. Ask what happened when execution conditions became worse. Ask which historical periods were excluded from optimization and used for independent validation.

And remember that even a carefully tested strategy can eventually lose its edge.

My own project, RIZA, combines multiple trading strategies with risk management mechanisms. I've described its structure and operating principles in the RIZA EA user guide.

But I'm not asking anyone to accept these ideas simply because I'm developing a trading robot.

Quite the opposite.

I'd genuinely like other traders and developers to challenge this approach.

Does optimization across mixed historical quarters offer meaningful advantages? How rigorous is this form of Monte Carlo testing? What additional validation methods would make the results more trustworthy?

And how useful is measuring strategy degradation at 0, 1, 2, or more basis points of additional trading costs?

I'm interested in technical arguments, criticism, and alternative approaches.

Because I believe a well-engineered trading system should be able to withstand more than difficult market conditions.

It should also withstand difficult questions.

And perhaps those questions are exactly where the evaluation of any trading robot should begin.