Machine learning in trading: theory, models, practice and algo-trading - page 2214

 
mytarmailS:

I don't know about python, but r-ku is not very good at it, or I don't know how to do it, so I came up with this...

In the same catbust you can write your own loss-factor and estimate by it on the train\test

I just don't see much point, the standard ones are enough
 
Maxim Dmitrievsky:

In the same Katbust you can write your own loss-factor and estimate by it on the train\test

I just don't see the point, the standard ones are enough

Well, katbust has a metric for maximizing profit with minimal drawdown?)

 
mytarmailS:

well, katbusta has a metric to maximize profits at minimum drawdown ? ))

well write an eps and put it as eval_metric

i just use R^2 for optimization

 
Maxim Dmitrievsky:

well, write an eps and put it as eval_metric

I just use R^2 for optimization

I'll have to try that, I didn't know...

what's an eps

 
mytarmailS:

I'll have to try it, I didn't know...

What's an eps?

Well it makes sense to write your own fi's, yes... the question is which is better

eprst

eepst

a guy wrote me a message from here that he's optimizing for Lyapunov's instability

You have to ask https://www.mql5.com/ru/users/alexeynikolaev2 if it's good enough. He certainly knows.)

Aleksey Nikolayev
Aleksey Nikolayev
  • www.mql5.com
Опубликовал пост От игры к чему-нибудь Модель, рассмотренная в предыдущей части , очевидно, нуждается в дальнейшем развитии. Решил сделать для этого новую запись. Более глобальная перспектива... Опубликовал пост От игры к вероятности. Небольшой пример. Ранее я уже немного рассуждал о переходе от модели с игровой неопределённостью к модели с...
 
Maxim Dmitrievsky:

Well, it makes sense to write your own f.i.s., yes... the question is which is better

eeprst

damn it

here's a friend wrote to me in the trolley that optimizes on some Lyapunov's instability

I don't know about Lyapunov, but fitness functions is a whole "world" of new opportunities, much more optimal solutions...

For example we have a "global objective" of maximizing profits with minimal drawdown...

We can say through the function "AMO", and if you think of such signs that will improve the "global target"

And it will look for solutions, it will "think out" by itself what you would never think out, and even more...


 
mytarmailS:

I don't know anything about Lyapunov, but fitness functions are a whole "world" of new possibilities, much more optimal solutions...

For example we have a "global goal" of maximizing profits with minimum drawdown...

We can say through the function "AMO", and if you think of such signs that will improve the "global target"

and it will look for solutions, it will "think of" things you would never think of, and even more...

I still don't understand it. Your goal is set by labels. I don't know what kind of profit and drawdown you'll get, but I'll tell you what you'll get.

write normally )
 
Maxim Dmitrievsky:

I still don't understand it. Your goal is set by the marks. What marks you set, the profit and drawdown will be the same

write normally )

I don't explain well.

Well, there are goals that you can't express with tags, only through finding the minimum.


For example - I want to create an informative attribute

No, it doesn't exist, and I don't even know what it should look like.

But I can describe it in terms of usefulness, an informative sign is such a sign, after training by which the "global objective" profit maximization at minimum drawdown will be significantly improved

You see, I can't create labels on the IP (informative attribute) but I can describe it as a maximization...


Now we can take a net with thousands of weights and play with its weights until it finds the IP, i.e. until it maximizes ...

Or not a network, we can do the same MSUA, or we can create rules...

Playing means to run some optimization algorithm, or search ...

 
mytarmailS:

I don't explain well((

Well, there are goals that you can not express with labels, only through finding the minimum


For example - I want to create an informative feature.

Can I create labels for it? No, it doesn't exist, and I don't even know what it should look like...

But I can describe it in terms of usefulness, an informative sign is such a sign, after training by which the "global objective" profit maximization at minimum drawdown will be significantly improved

You see, I can't create labels on the IP (informative attribute) but I can describe it as a maximization...


Now we can take a net with thousands of weights and play with its weights until it finds the IP, i.e. until it maximizes ...

Or not a network, we can do the same MSUA, or we can create rules...

Playing around means to run some optimization algorithm, or search

So you have a trained network and you just change the weights and see how it works on the new data?
 
Maxim Dmitrievsky:
So you have a trained network and you just change the weights and see how it works with new data?


Basically I have an empty network (I only train it to initialize it, because it's not self-written, but from a package).


I invent any abstraction, any target and write a fitness function.

Then let genetics start to change the weights of the network so that at the train and in the test I (the network) would get something maximum similar to my goal.


And this is "a thousand times" more profound than creating labels and fitting regression or classification