Machine learning in trading: theory, models, practice and algo-trading - page 3460

New comment
 
Maxim Dmitrievsky #:
🙀

I encounter misunderstanding everywhere. I'm used to it, not interested.

 
If we input price increments with some transformation (0...1) or (-1...1), the system will work infinitely.

Infinitely bad.

If you give pure prices as input and predict the next price as output, a simple MLP produces the following picture:

The best quality averaging that can be had.

That said, it only works on the forward if the price chart does not change dramatically. At some point the NS simply stops working.

And even additional training does not help, it just fades on new data and does not open. Then, when the market either calms down or something else happens to it, the NS after training with direct prices continues to work.

And sometimes it works very well. Especially on EURGBP and other flat markets.

UPD

Stationarity seems to kill some... important information
 

I am curious to know the result in the case of offset TF construction.

Using H1 as an example. Now the TF is built on this data [HH]::[00]-[HH+1][00]. If I build with an offset of 12 minutes: [HH ]::[12]-[HH+1][12], will it affect the results of the TF based on MO?

If the MO was trained on M1, the offset is not 12 minutes, but 12 seconds.


Obviously H1 and H1+12 are different data. Will other patterns be found?

 
fxsaber #:

Curious to know the result in the case of offset TF construction.

Using H1 as an example. Now the TF is built on this data [HH]::[00]-[HH+1][00]. If we build with an offset of 12 minutes: [HH ]::[12]-[HH+1][12], will it affect the results of the MO-based TF?

If the MO was trained on M1, the offset is not 12 minutes, but 12 seconds.


Obviously H1 and H1+12 are different data. Will other patterns be found?

Real-time translate now?

Forum on trading, automated trading systems and testing trading strategies

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

fxsaber, 2024.04.06 12:48 AM

I am curious to know the result in the case of offset TF construction.

Using H1 as an example. Now the TF is built on this data [HH]::[00]-[HH+1][00]. If I build with an offset of 12 minutes: [HH ]::[12]-[HH+1][12], will it affect the results of the TF based on MO?

If the MO was trained on M1, the offset is not 12 minutes, but 12 seconds.


Obviously H1 and H1+12 are different data. Will other patterns be found?

 
fxsaber #:

Whatever.

Ivan Butko #:
If you input price increments with some transformation (0...1) or (-1..1), the system will work infinitely.

Infinitely bad.

If we give pure prices as input and predict the next price as output, then in a simple MLP we get the following picture:

the best quality averaging that can be.

But, at the same time, it works on a forward only if the price chart does not change dramatically. At some point, NS simply stops working.

And even additional training does not help, it just fades on new data and does not open. Then, when the market either calms down or something else happens to it, the NS continues to work after training with direct prices.

And sometimes it works very well. Especially on EURGBP and other flat markets.

UPD

The stationarity seems to be killing some..... important information

my opinion is that nonstationarity contains the lion's share of useful information, which disappears if the series are stationary.

 
fxsaber #:

Curious to know the result in the case of offset TF construction.

Using H1 as an example. Now the TF is built on this data [HH]::[00]-[HH+1][00]. If we build with an offset of 12 minutes: [HH ]::[12]-[HH+1][12], will it affect the results of the MO-based TF?

If the MO was trained on M1, the offset is not 12 minutes, but 12 seconds.


Obviously H1 and H1+12 are different data. Will other patterns be found?

Other patterns will be found (besides those that can be found on standard TFs). Imho, the patterns of non-standard TFs are less "chewed out" by market participants (there are strategies that still work on non-standard TFs, but haven't worked on standard ones for a long time).
 

What does this have to do with the IO?

There is a huge set of tools that are grouped together in a science called "Statistics". IOs are some of the tools.

The whole science of "Statistics" is not just a list of tools, but also the conditions in which these very tools are applicable. It's like a 12 spanner only with 12 nuts. Nobody argues at the level of the spanner, but the MO is somehow not so clear.

MO gives the better result the closer the initial data are to stationarity, and the initial series is fundamentally NOT stationary. Therefore, MO is fundamentally not applicable to the original series as a 12 spanner is to any other nut.

 
fxsaber #:

Real-time translate now?

yeah but can't follow , i'm sure the translation is equally bad to Russian

 

I believe that whether there is stationarity or not depends on the point of view of the object,

from which point of view of the feature space one looks at it.


Saber's writing is correct, albeit a bit sloppy.

fxsaber #:

Perhaps somewhere stationarity manifests itself on profitable TCs not in an explicit form - stationarity is not created on purpose.


For example, someone trades only in the evening session because it is profitable. He has no idea about the reasons of this fact. It is just a fact.

If you start digging into the reasons, it may turn out that the market is more likely to be stationary in the evenings than at other times.


But the person who discovered evening trading didn't even know anything about the concept of stationarity. He never looked for it. He just tried it and it worked. This man is much dumber than the MoD. But for some reason he had a primitive idea to put a trading time limit on his stupid TS.

 
Maxim Dmitrievsky #:
A time filter is almost the same as a volatility filter with a period of 20.
Does the price (series) contain time data?
1...345334543455345634573458345934603461346234633464346534663467...3526
New comment
Reason: