I have a question about setting the "start" value in a multi-parameter optimize run:

The manual says:

and also:


Can you explain briefly, for example, how Zorro finds the optimize value for parameter 1? Does it do this by assuming only the "start" value on all the remaining parameters at first? The manual says to use the "estimated best value" and in cases where I don't know this I've used only the minimal value, which might be a mistake on my part.

Also, on subsequent parameters, does Zorro then assume the found-optimal value on the earlier parameters already optimized, or some other value? (In other words, after determining the optimal value for parameter 1, will this value be assumed during the parameter 2 optimize, or another value like start?)

Sorry if I've not explained this very well. I am just trying to understand the logic behind multi-parameter optimization, so that I can make sure I am feeding-Zorro the best inputs.

Yes to both questions. For subsequent, not yet optimized parameters, Zorro uses the start value; for earlier, already optimized parameters it uses the optimized value.

Thanks for explaining!

Please correct if this is wrong understanding:

INITIAL OPTIMIZE RUN with 3 optimize calls:
Parameter 1 optimize--> assume "start" value on P2 and P3 to find optimal P1
Parameter 2 optimize--> use found-optimal P1, P3 "start" to find optimal P2
Parameter 3 optimize--> use found-optimal P1 and P2 to find optimal P3

Yes, that's correct.

As an interesting aside... I've been learning a ton in trying to put together a methodical process for designing a strategy (I'll share my findings for comments once I get them a little more stable) ...

But based on this new understanding of how the optimizer works, I've actually modified my process like this:

If I'm creating a new range for optimizing a parameter and do not know how it will perform, I therefore do not have an "estimated best value".

So what I do is pick a guesstimate value and run the WFO optimize. Then I look at all the optimized parameters from each cycle. I take an average of those optimal parameters and then use that figure as a "start" value and then apply a new WFO optimize. I do not think this is overfitting, because I'm simply giving Zorro a cleaner starting point (that it will use for assumptions when doing a multi-parameter optimize).

Is it possible to set a flag so that all parameter combinations are run.

At times I've found myself running many combinations manually, as there can be profitable zones e.g. with a large stop and close target and a closer stop with a further target.

I don't think there is. But I support it as a request for future Zorro releases.
I think the reason why optimize does not work that way is because the time the program runs grows exponentially with each optimized parameter if all combinations are tested.
And there is a huge danger of overfitting.

Another possible danger that I've run into several times is the memory limit of ~1.5gig for 32-bit applications. I usually hit this when trying to do a WFO while looping across multiple assets.

I found what I think is the same thing you are talking about swingtraderkk -- the only solution I've come up with is a multi-step process to figure out which ranges work best, then work on optimizing only that range.

For example, on equity-curve trading, I wanted to know if the LowPass-filtered crossover curves should be narrower or wider, so I actually came up with 4 different scenarios and I've found that usually 1 of them outperforms the others: narrower, wider, mega-wide and uber-wide. Once I identify (generally) which one works best with the given strategy, then I look to optimize only that one.

Optimizers normally do not run all parameter combinations. This would take too much time and limit you to maybe two parameters. A genetic algorithm is used instead.

Problem is that systems optimized this way are normally extremely overfitted and useless in real trading. That's why we did not implement such an optimizing method. Of course, we can implement it anyway, so you can use it at your own risk.

Dusktrader that is what I have been doing, but it is tedious when you have multiple strategies to test, and variants of each also.

Clearly this would be just for an initial screen, without wfo's etc. when the screening is done, I would optimise in detail one I know the ball park.