Machine learning in trading: theory, models, practice and algo-trading - page 3368

 

Optimise for what?

Adapt to what?

There are always some discussions, as if there is nothing around. And most importantly, it is not clear what problem we are solving with the help of optimisation or adaptation.

If the problem is not specified, then all this is just another empty rubbish.

And the problem is exactly the same: a non-stationary market. That is why any talks about optimisation or adaptation of different mashka even with ATRs are just empty rubbish, and about any indicators and even the most sophisticated filters and so on and so forth....

There are two approaches in non-stationary markets, which include financial time series:

1. Modelling of non-stationarity with preliminary destruction of its most blatant manifestations, for example, trends - these are GARCHs

2. Searching for patterns in history using MOE. If some effort is put into preprocessing, such patterns will give a future classification error of less than 20%.

All.

 
mytarmailS Mashka 's period is controlled by ATR, it is also impossible... utopia?

Easy, and another 100500 variants can be invented. Which of them will work on the market?

 
Forester #:

This is about science, where the formulas are precise and they only need to change the input parameters. We don't have science, we don't have formulas describing the market.

I'm not talking about a miracle formula that describes the entire market. I'm saying that in a non-stationary environment a system with dynamic parameters is better than a system with static parameters!

That's it!

Forester #:

Easy, and you can come up with 100500 more variations. Which one will work in the market?

The Mashko and APAC example is an abstract simple example just to get my point across, not a guide to what I'm talking about.

 
Somewhere, Oleg Avtomat is quietly sorrowing.
 
Maxim Dmitrievsky #:
Somewhere Oleg Avtomat is quietly sorrowful

)

By the way, his thread started with a link to a book with another normal variant of modelling non-stationarity. There it was presented as random switching between several stationary series. Another thing is that his self-reference had no relation to the contents of the book) He even denied its basic mathematical apparatus - the theorist).

 

The automat swore a lot, but kept doing what he was doing when I told him that his PI/PID regulators can be easily replaced by mashki, and if the furnace temperature is influenced by an external influence equal to the EURUSD chart, he will not be able to regulate the heat of the spirals with the help of regulators so that to extinguish the external influence and get a stationary rad of the furnace temperature (to get a profit). The furnace will either cool down to shop temperature or burn out the coils and scorch the brick walls.

"That's why it happened the way it did..." Eduard Severe.

))

 
Aleksey Nikolayev #:

)

By the way, his thread started with a link to a book with another normal variant of modelling non-stationarity. There it was presented as a random switching between several stationary series. Another thing is that his selections had not the slightest relation to the contents of the book) He even denied its basic mathematical apparatus - the theorist).

The question is whether it needs to be modelled. There is a good method of estimation in MOE - it is CV. Through it you can even select the parameters of the TC and rearrange the datasets.

A couple of ideas came to mind recently, I'll have to check it out. For example, I shuffle data, and it would be nice to use CV for time-series.
 
Maxim Dmitrievsky #:
Question: and whether it should be modelled.

There is some theorem that states, roughly speaking, that any model predicting an object is always equivalent to some model describing the object. It turns out that we always build some model of the market in addition to the desire.

 
Aleksey Nikolayev #:

There is some theorem that states, roughly speaking, that any model predicting an object is always equivalent to some model describing the object. It turns out that we always build some model of the market in addition to the desire.

modelling can be done in different ways))))

 
Aleksey Nikolayev #:

There is some theorem that states, roughly speaking, that any model predicting an object is always equivalent to some model describing the object. It turns out that we always build some model of the market in addition to the desire.


There are identity theorems. Simply put: "If something looks like a camel, drinks like a camel, smells like a camel, then it is a camel."
You can even calculate the degree of model's correspondence to the market, for this purpose you should calculate the sum of all price increments on a given area, the sum of pips earned by the system and compare them.

In the same way, you can estimate the degree of drift of the system's compliance with the original on oos