Machine learning in trading: theory, models, practice and algo-trading - page 2118

 
mytarmailS:

What?

Nothing, you take two mas and a martingale... profit for a while

 
Maxim Dmitrievsky:

It's okay, you take two mas and a martingale... ...for a while...

what does that have to do with it?

 
Alexander_K:

Gorgeous pictures. Perhaps this is what the Grail looks like in the projection from N-dimensional space...

Or a noose around the neck, in multidimensional space ))

 
Maxim Dmitrievsky:

Or a noose around the neck, in multidimensional space ))

:))))

Interesting comments here:

https://smart-lab.ru/blog/658290.php

Машинное обучение в трейдинге - насколько эффективен подход?
Машинное обучение в трейдинге - насколько эффективен подход?
  • smart-lab.ru
Не в коем случае не хочу оскорбить людей, кто занимается ML. Наличие кандидатской или пр. международных сертификатов могут говорить что-либо о человеке? Это нужно для трудоустройства и чтобы пускать пыль в глаза тем, кто платит деньги. Тоже самое касается и мира сомелье, где сертификаты WSET ничего не стоят и любой гик-сноб уделает по матчасти...
 
Alexander_K:

:))))

Interesting comments here:

https://smart-lab.ru/blog/658290.php

well there are no links to any research, only to this forum... which is frustrating

 
Maxim Dmitrievsky:

Well, there are no references to any research, only to this forum... which is frustrating

It argues, quite reasonably, that a neural network by itself is unable to find patterns in the market time series. Alas - I agree with that. The non-stationarity of both variance and expectation is to blame. The key link is the constant significant shifts of the expectation relative to 0 for any sample of increments.

Therefore, it is important to single out stationary sections and trade only on them.

1. I will try before the New Year to collect data for each hour within a day, glue it together, and see if the hourly plots are stationary.

2. Someone named Demko, at one time, argued that the OPEN price series for bars consisting of 100 ticks (equal tick bars) is stationary. I watched his research - yes, he seems to be right.

3. Warlock also did some preprocessing of the data.

Although I don't use MO, I sincerely wish the sufferers in this thread good luck and profits. I worry, so to speak.

 
Alexander_K:

It argues, quite reasonably, that a neural network by itself is unable to find patterns in the market time series. Alas, I agree with that. The reason is the non-stationarity of both the variance and the expectation. The key link is the constant significant shifts of the expectation relative to 0 for any sample of increments.

Therefore, it is important to single out stationary sections and trade only on them.

1. I will try before the New Year to collect data for each hour within a day, glue it together, and see if the hourly plots are stationary.

2. Someone named Demko, at one time, argued that the OPEN price series for bars consisting of 100 ticks (equal tick bars) is stationary. I watched his research - yes, he seems to be right.

3. Warlock also did some preprocessing of the data.

Although I don't use MO, I sincerely wish the sufferers in this thread good luck and profits. I worry, so to speak.

The analytical model of the draw implies the difference between the real and the unsteady probability of the forecasting. Predicting the price is not the purpose, in order to trade we need to predict the future returnee, the returnee series is quasi-stationary if it is aligned with the seasonal volatility. But the future returnee is predicted very badly, totally crap and it has nothing to do with non-stationarity, what does it have to do with anything, if it were at the cumulative price, it would be an obvious pattern like seasonal volatility, but it is not and so why keep talking about it?

 
kapelmann:

Stationarity/non-stationarity is not the cause of all traders' troubles, you can improvise a series with the same statistical characteristics as the financial series, the same non-stationary one, but from which it is easy to make the Grail. Predicting the price is not the purpose, in order to trade we need to predict the future returnee, the returnee series is quasi-stationary if it is aligned with the seasonal volatility. But the future returnee is predicted very badly, very badly and it has nothing to do with non-stationarity, what does it have to do with anything, if it were at a cumulative price, it would be an obvious pattern like seasonal volatility, but it's not and that's it, why keep talking about it?

Um... Then why don't the MOs have positive stats? After all the ACF of a market series of increments is not equal to 0 and therefore should be predictable. Obviously, because the 2nd condition for predictability according to Kolmolgorov is not met - there is no constancy of expectation for any sample of data. What's wrong with that?

 
Alexander_K:

It argues, quite reasonably, that a neural network by itself is unable to find patterns in the market time series. Alas, I agree with that. The reason is the non-stationarity of both the variance and the expectation. The key link is the constant significant shifts of the expectation relative to 0 for any sample of increments.

Therefore, it is important to single out stationary sections and trade only on them.

1. I will try before the New Year to collect data for each hour within a day, glue it together, and see if the hourly plots are stationary.

2. Someone named Demko, at one time, argued that the OPEN price series for bars consisting of 100 ticks (equal tick bars) is stationary. I watched his research - yes, he seems to be right.

3. Warlock also did some preprocessing of the data.

Although I don't use MO, I sincerely wish the sufferers in this thread good luck and profits. I'm worried, so to speak...

I read somewhere that ticks are thinning and that they are in different amounts for all brokerage companies. Some have 100 ticks per minute, other have 300. The values are seldom on demo accounts. For example one broker has 1 liquidity and quotes provider, the other one has 3, and the other one combines them.
On something unstable from one brokerage company to another, it is impossible to make something stable.

 
Alexander_K:

It argues, quite reasonably, that a neural network by itself is unable to find patterns in the market time series. Alas, I agree with that. The reason is the non-stationarity of both variance and expectation. The key link is the constant significant shifts of the expectation relative to 0 for any sample of increments.

Therefore, it is important to single out stationary sections and trade only on them.

1. I will try before the New Year to collect data for each hour within a day, glue it together, and see if the hourly plots are stationary.

2. Someone named Demko, at one time, argued that the OPEN price series for bars consisting of 100 ticks (equal tick bars) is stationary. I watched his research - yes, he seems to be right.

3. Warlock also did some preprocessing of the data.

Although I don't use MO, I sincerely wish the sufferers in this thread good luck and profits. I'm worried, so to speak...

I assume that the issue is not stationarity, but regularity. There are no patterns. If you add patterns to a random series, then MO starts working dramatically

but, unfortunately, on smradlab do not know what regularity is, so they mention stationarity %)