Machine learning in trading: theory, models, practice and algo-trading - page 2518

 
Ivan Butko #:
Gentlemen, regulars of this thread, tell me if there are any successes in machine learning, ready-made products that work?

The branch has grown fiercely, the most active, probably.

There are no ready-made ones. Neither are there any long-playing ones. And learning as you go along, it's a heavy algorithm. This is not optimization.

 
Dr. #:

Yes, the right answer is known to everyone and no one is interested in it. What is of interest is the answer of such colorful characters as the automaton. And Alexander's answer is interesting. With his energy and passion).

Hello, Doctor! Things of bygone days... Well... Does he need the right answer?

The very formulation of the question provokes an active discussion, whole volumes are already written on the smartlab. People are still researching Wizard's method - giving test results, both confirming and denying it, trading with it. People are interested and that's the main thing.

 

Alexander_K #:

People are still researching Koldun's method - giving test results, both confirming and refuting it, trading with it.


Is that where the channel with the sum of increments is?

Obviously, this system will not work directly because of the trends (spikes). At first glance, volatility prediction seems to be asking for itself, but who knows...

 
Evgeniy Chumakov #:


Is that where the channel with the sum of the increments is?

What can I say, obviously this system will not work directly because of trends (spikes). At first glance, volatility forecasting seems to be in the cards, but who knows...

Yes. People are just fine-tuning it to their needs and desires. I personally - use some kind of average and volatility prediction, which, as you know, depends on the number of ticks received per unit time.

 
Alexander_K #:

Yes. People are just fine-tuning it to fit their needs and aspirations. I personally - use a certain average and volatility forecast, which, as you know, depends on the number of ticks per unit time.

The speed of ticks in what period do you measure relative to TF. And how. as number of ticks per period or average time between ticks per period?

 
Valeriy Yastremskiy #:

The speed of ticks for what period are measured relative to the TF. And how. as the number per period or the average time between ticks per period?

When we use the formula S*sqrt(T) to calculate the standard deviation of the process, we must understand that in the market T is a function of the number of ticks coming (or events on a larger scale) from time.

It is this T that we need to be able to predict. In the simplest case it is the maximum possible number of events during a definite time interval, for example during a day. This eliminates the non-stationarity of the process of occurrence of events.

For periods < a day, it is necessary to count the number of events for each hour from 0 to 23, and predict them for the chosen time period, for example =8 hours for the current time of the day.

 
Alexander_K #:

When we use the formula S*sqrt(T) to calculate the standard deviation of the process, we must understand that in the market T is a function of the number of incoming ticks (or events on a larger scale) from time.

It is this T that we need to be able to predict. In the simplest case it is the maximum possible number of events during a definite time interval, for example during a day. This eliminates the non-stationarity of the process of occurrence of events.

For periods < a day, it is necessary to count the number of events for each hour from 0 to 23, and to predict their number on a chosen time period, for example =8 hours for the current time of the day.

Is this like a reasonable multiple of the TF? At 15 minutes all 15 minutes count sliding window every minute or take an hour, or 5 minutes to determine the rate.

 
Valeriy Yastremskiy #:

Is this like a reasonable multiple of the TF? At 15 minutes all 15 minutes count as a sliding window every minute, or take an hour, or 5 minutes to determine the speed.

I do not work on such scales, alas. I have a window of 8 hours and higher on astronomical time (and, accordingly, variable on market time).

 
Alexander_K #:

When we use the formula S*sqrt(T) to calculate the standard deviation of the process, we must understand that in the market T is a function of the number of incoming ticks (or events on a larger scale) from time.

It is this T that we need to be able to predict. In the simplest case it is the maximum possible number of events during a definite time interval, for example during a day. This eliminates the non-stationarity of the process of occurrence of events.

For periods < a day, we have to count the number for each particular hour from 0 to 23, and predict their number for a selected time period, for example =8 hours for the current time of day.


S*sqrt(T)

Then it is possible to suppose, that if we work in a fixed window according to astronomical time, then instead of forecasting T (using hourly statistics of ticks amount arrival, by type of histograms of hourly volatility etc.) i.e. the amount of events (ticks in this case) it is necessary to forecast the amount of price change points in the future some P.

Though it's more of a delusion... ))

 
Evgeniy Chumakov #:


S*sqrt(T)

Then it is possible to suppose, that if we work in a fixed window on astronomical time, then instead of forecasting T (using, as I understand, hourly statistics of ticks amount arrival, by type of histograms of hourly volatility, etc.) i.e. the amount of events (ticks in this case) it is necessary to forecast the amount of price change points in future some P.

Though it's more of a delusion... ))

But actually no, it's not bullshit, it works approximately the same way. It's almost that way... price change points are predicted as long as sqrt(t) is within certain limits.
Half of the forum is devoted to it and everybody says it's not always "in the frame" and getting out of the frame is always a surprise.
As long as sqrt is close to a typical one, even the timing of reversals is very well known. But as soon as it exceeds them the reversals "slip" to the right and to the left.