Machine learning in trading: theory, models, practice and algo-trading - page 2834

New comment
 
Andrey Dik #:

1. My last name is not declined
2. the question is not whether the global will change or not (it will change necessarily), but whether it can find the global extremum at all. if you don't care, you can just initialise the network weights with random numbers and that's it, because what difference does it make if it's global or not? local.))

1. About the surname, I'll take it into account. You are the only person I know with such a surname.

2. I would like to object that we should not bring the idea to absurdity, but in reality the case is the same: we do not need any optimums - we need plateaus, the more extensive, the better, the more profitable, the better. And extremum is always an unstable state, but we need a stable state, and in the future. Therefore, it turns out that the search for optimums is a harmful activity.

Here I am pushing the idea that the stability of profit in the future depends on the properties of predictors and depends very little on the model, and even less on the optimisation algorithms. That is, a completely different approach.

 
СанСаныч Фоменко #:

What I'm pushing here is the idea that the sustainability of profits in the future depends on the properties of the predictors and depends very little on the model, much less on the optimisation algorithms. That is a completely different approach.

Can you send me your sample? We have the same vision of the problem of poor model learning, I would like to compare how much your selection method is better than mine, and whether it fits your sample at all.

 
What is proposed to be substituted in place of logloss?
 
Aleksey Nikolayev #:

Yeah, that's the tricky part. You need to calculate the gradient and the matrix of second derivatives - the Hessian. The Hessian must also be positive definite.

For profit, for example, the Hessian is degenerate like (equal to zero). It is necessary to somehow finesse the loss function to a reasonable compromise between need and necessity.

Here, as I understand it, the Hessian is not taken into account at all.

#  Custom objective function (squared error)
myobjective <- function(preds, dtrain) {
  labels <- getinfo(dtrain, "label")
  grad <- (preds-labels)    
  hess <- rep(1, length(labels))                
  return(list(grad = grad, hess = hess))
}

#  Custom Metric
evalerror <- function(preds, dtrain) {
  labels <- getinfo(dtrain, "label")
  err <- (preds-labels)^2        
  return(list(metric = "MyError", value = mean(err)))   
}

just works to reduce the gradient.

grad <- (preds-labels)  


This is different.

logregobj <- function(preds, dtrain) {
  labels <- getinfo(dtrain, "label")
  preds <- 1 / (1 + exp(-preds))
  grad <- preds - labels
  hess <- preds * (1 - preds)
  return(list(grad = grad, hess = hess))
}


I remember I had a lot of trouble with it, so I gave it up.

But it would be cool to train my physique.)
 
СанСаныч Фоменко #:

But in reality, this is the case: we don' t need any optimums - we need plateaus, themore extensive, the better, the more profitable, the better. And an extremum is always an unstable state, and the

And the function that evaluates "how much plateau/not plateau" is not the search for maximum ?

And a function that evaluates "the more profitable the better" is not a maximum search ?

And the function that evaluates "the more extensive the better" is not a maximum search?

And the function that evaluates "steady/unsteady state" is not a maximum search ?

Or do you need a plateau there too? )))) Well, then see point one ))))

Oh these profane people ...



SanSanych Fomenko #:

I'm pushing the idea here that the sustainability of profits in the future depends on the properties of the predictors and very little on the model, much less on the optimisation algorithms. That is, a completely different approach.

You should not push the idea, but prove it or at least justify it....

There was one here, pushing the idea that Mashka is the best thing that can be, and what's the use of this pushing??????

 
Even cooler would be to include the missing body and think about what kind of relationship could describe the profit factor between chips and tags. Roughly none 🤤
 
СанСаныч Фоменко #:

1. I'll keep the last name in mind. You're the only person I know with that last name.

2. I would like to object that we should not bring the idea to absurdity, but in reality the case is the same: we do not need any optimums - we need plateaus, the more extensive, the better, the more profitable, the better. And extremum is always an unstable state, but we need a stable state, and in the future. Therefore, it turns out that the search for optimums is a harmful occupation.

Here I am pushing the idea that the stability of profit in the future depends on the properties of predictors and depends very little on the model, and even less on the optimisation algorithms. That is a completely different approach.

1. v

2. Here is a figure, like a curve of some hypothetical learning function.

Are you satisfied that the grid will stop at local extremum 1? Or maybe 2? Or what, 3, like in the middle? So it is not known in advance how many local extrema there are, there may be 100500 or even more. That is why it is important to try to find the highest of all local extrema that the algorithm can reach.

 
Maxim Dmitrievsky #:
What is proposed to be substituted for logloss?

If it's a question for me, profit or some reasonable analogue of it. Profit - let it for now just be the sum of all (close[i] - open[i])*prognos[i], where prognos[i] is the prediction of the candle colour and is 1 or -1. It may have to be modified somehow for good gradient and hessian behaviour.

 
Aleksey Nikolayev #:

If it's a question for me, profit or some reasonable analogue of it. Profit - let it for now just be the sum of all (close[i] - open[i])*prognos[i], where prognos[i] is the prediction of the candle colour and is 1 or -1. It may have to be modified somehow for good gradient and hessian behaviour.

In genetics, we take variables and maximise them by a criterion. You can't do that here, because classification. There is no relationship between profit and class labels. At best, you'll get nothing. That's why such criteria are placed in eval_metrics
 
Maxim Dmitrievsky #:
In genetics, we take variables and maximise on a criterion. You can't do that here, because it's a classification. There's no relationship between profit and class labels. At best, you'll get nonsense. That's why such criteria are placed in eval_metrics

There is no happiness in life.)

1...282728282829283028312832283328342835283628372838283928402841...3590
New comment