Machine learning in trading: theory, models, practice and algo-trading - page 2841

[mytarmailS]

I decided to trade Eurobucks for a minute, without any indicators.

not a single minus, the entries are accurate, which means that I still have some understanding of the market...

But also analysing myself from the outside, how I make analyses, make decisions, etc... I can't imagine how to put something like this into the machine, given my already obviously not the initial level in the MO....

[Rorschach]

Let me expand on Alexei's point. At the input we have only past data. On the history, an ideal system would be to "memorise" all movements of the available data. This is the task of interpolation; we select a function that perfectly describes historical data. There are infinitely many such functions. As the depth of history increases, they will become less and less, it will be more and more difficult to find them, but beyond the available history, 99% of these remaining functions will be pointing a finger to the sky. We are unfortunately interested in the area outside the history (extrapolation). The first thing we can try to do is to look for internal regularities: check for autocorrelation, construct a Fourier spectrum, look at various statistics, etc. But if we are dealing with a complex system (chaos, PRNG, encrypted signal), then these methods will be ineffective. All we are left with is trying to find a function that approximately describes future and past data (approximation) - through crossvalidation, testing on several pairs and other indirect methods. And here we come to the conclusion that classical TS and NS have much in common in optimisation - to find coefficients of some function by indirect methods. For this purpose, NS use minimisation of ско, variance and so on. But in trading other criteria are more preferable, such as profit maximisation, drawdown minimisation, etc. By and large, regardless of the method of solving the problem (via NS or otherwise), everything comes down to finding the coefficients of the approximating function. Then additional logic is added: stops, take-outs, MM, PM, etc.

[Maxim Dmitrievsky]

Purely technically, you can't find a suitable optimisation surface by going through everything. And the more everything, the lower the chances. Optimisation solves nothing in this matter, and only works in its own area - to improve the good.

Here we can agree with Prado: "stop optimising, look for patterns"

Clean. Technically. You can't. Find a suitable TC through optimisation.

So plateaus and peaks that don't match on new data fall off by themselves. There is simply no such problem, if you don't do nonsense.

[Evgeni Gavrilovi]

What attributes does fxsaber use when designing its trading system?

[Aleksey Nikolayev]

Maxim Dmitrievsky #:

Prado: "stop optimising, look for patterns"

If purely formally, the logic is lame - any optimisation always takes place within some model, as a way of finding its parameters. That is, some model is always already found).

It is clear that he is talking about the fact that no optimisation algorithm can fix a bad model. The question arises - how to distinguish good models from bad ones. If there is no a priori knowledge (for example, you can try to find out what models fxsaber uses 😆 ), then you will have to resort to some a posteriori methods, which will obviously lead to optimisation).

[Maxim Dmitrievsky]

Aleksey Nikolayev #:

If purely formally, the logic is lame - any optimisation always takes place within some model as a way of finding its parameters. That is, some model is always already found)

It is clear that he is talking about the fact that no optimisation algorithm can fix a bad model. The question arises - how to distinguish good models from bad ones. If there is no a priori knowledge (for example, you can try to find out what models fxsaber uses 😆 ), then you will have to resort to some a posteriori methods, which will obviously come down to optimisation).

I can't discuss other people's approaches as I'm not knowledgeable. But based on my own experience, what worked was found either in the internet or through my own inferences and then confirmed in the internet. For the whole history of my trading career, there were probably 2 strategies optimised, both on the return to the average, one of them on Martin, which earned something, but not for a long time :). And there were quite a lot of attempts to optimise, but only 2 strategies in the end, and they were not so good.

One of them earned 1500% for the whole time purely on the falling market and merged when it changed, but the funds were withdrawn with profit. The second one even less.

And this is 10+ years, maybe even 15, of constant/periodic searches through optimisation.

Of course, someone can argue that I am just stupid and he is d'Artagnan, but I don't believe it.

[Valeriy Yastremskiy]

Maxim Dmitrievsky #:
I can't discuss other people's approaches, as I am not knowledgeable. But based on my own experience, what worked was found either in the Internet, or through my own conclusions and then confirmed in the Internet. For the whole history of my trading career, there were probably 2 strategies optimised, both on the return to the average, one of them on Martin, which earned something, but not for a long time :). And there were quite a lot of attempts to optimise, but only 2 strategies in the end, and they were not so good.

One of them earned 1500% for the whole time purely on the falling market and merged when it changed, but the funds were withdrawn with profit. The second one even less.

And this is 10+ years, maybe even 15, of constant/periodic searches through optimisation.

Of course, someone can argue that I am just stupid and he is d'Artagnan, but I have little faith.

Year 14 was calmer and the descent was longer. Now it's similar, but shorter and more unpredictable.

It's a hobby, part of life.)

[Maxim Dmitrievsky]

Valeriy Yastremskiy #:
Year 14 was quieter and the descent was longer. Now it's similar, but shorter and more unpredictable.

It's a hobby, a part of life)

As if there were and partly there are other strategies, but they do not belong to optimisation strategies in any way

At that time those 2 straight gave me a good deal, I did not deny myself anything, but it was not long ).

[Aleksey Nikolayev]

Maxim Dmitrievsky #:
I can't discuss other people's approaches, as I am not knowledgeable. But based on my own experience, what worked was found either in the Internet, or through my own conclusions and then confirmed in the Internet. For the whole history of my trading career, there were probably 2 strategies optimised, both on the return to the average, one of them on Martin, which earned something, but not for a long time :). And there were quite a lot of attempts to optimise, but only 2 strategies in the end, and that's not so good.

One of them earned 1500% for the whole time purely on the falling market and merged when it changed, but the funds were withdrawn with profit. The second one even less.

And this is 10+ years, maybe even 15, of constant/periodic searches through optimisation.

Of course, some might argue that I'm just stupid and he's d'Artagnan, but I don't believe it.

The word "optimisation" has a bad reputation on our forum for obvious reasons. So it is quite understandable to want to somehow get away from it and not even use the word itself. Nevertheless, any training of an MOE model is almost always optimisation, and you can't take words out of a song.

I don't want to hurt anyone's feelings, teach them about life or explain how to do business) I am writing only in the faint hope that metaquotes will take my remarks into account when implementing MO in MT5.

[СанСаныч Фоменко]

Aleksey Nikolayev #:

You are also, in fact, doing optimisation. You have invented some criterion of "stationarity of features" and take the optimal features according to it. It's the same optimisation in history, but in profile.

We should definitely invent a criterion of TS robustness and optimise according to it) We will get the same optimisation on history again,but in a different profile).

Great, in terms of tolerance.

You go to a shop, choose trousers - optimisation according to your figure!

Here we are talking about something completely different - the refinement of the optimisation algorithms available in the models. I object to refinement of already built-in optimisation algorithms. There is an algorithm in the tester - fine. Refining of this algorithm will not allow you to get a profitable TS from a draining one. The same is true for inbuilt algorithms in models.

Also, you should be extremely careful when optimising model parameters, because you can easily retrain the model.

I have reached the most important thought: the undoubted connection between optimisation and overtraining of the model. The model should always be left fairly "coarse" and certainly no global optimums are needed.

When I'm looking for an acceptable list of predictors - optimisation in the trousers sense. But the meaning is quite different: trying to avoid "rubbish in - rubbish out". There's a qualitative difference here from trying to find the "right" algorithm that finds the global optimum. No global optimum will give a profitable TS on rubbish.