Machine learning in trading: theory, models, practice and algo-trading - page 480
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Again, this depends on how the classification is constructed. In the example above, the classification was based on the distance from the center line (boundary) without taking into account the absolute value of the increment. If you introduce the absolute value of increment, the classification will be different in principle. Its scale will also be different.
there are 2 outputs, to buy and to sell, their sum is always equal to 1.
We teach the principle:
10 points on the buy - we put 0.6; 0.4.
20 points - 0.7; 0.3
30 points - 0.8; 0.2.
The reverse is true for sales. Will it be correct to train in such a way and will these values then indicate that the more probable a class is the stronger the increment will be? :)
Or is it necessary to do not 2 but N-classes, each will be responsible for % increment, let's say 10 classes, each subsequent class is 10 points more
So the problem is: always use 2 classes or make more if we want to predict not only buy/sell belonging but also the degree of price change
there are 2 exits, to buy and to sell, their sum is always equal to 1.
We teach about the principle:
there was an increment to buy of 10 points - we put 0.6; 0.4
20 points - 0.7; 0.3
30 points - 0.8; 0.2.
The reverse is true for sales. Will it be correct to train in such a way and will these values then indicate that the more probable a class is the stronger the increment will be? :)
Or is it necessary to do not 2 but N-classes, each will be responsible for % increment, let's say 10 classes, each subsequent class is 10 points more
So the problem is: always use 2 classes or make more if we want to predict not only buy/sell membership but also the degree of price change
This is the wrong approach.
First, there must be at least three states: buy, sell, stop.
Secondly, in the future you will come to the necessity and usefulness of introducing distinctions for different movement states (buy-increase, buy-brake and sell-increase, sell-brake).
And most importantly, you have to distinguish between "state" and "action. For this it is not enough to take into account only the number of points traveled.
This is the wrong approach.
First, there must be at least three states: buy, sell, and stop.
Second, you will come to the necessity and usefulness of introducing distinctions for different states of motion (buy-increase, buy-brake and sell-increase, sell-brake).
And most importantly, you have to distinguish between "state" and "action. For this it is absolutely not enough to take into account only the number of points passed.
It is clear that you can add other states. But still, wouldn't they be derived from the probabilities of belonging to this or that class (of 2)?
Probability is the degree (a relative measure, a quantitative assessment) of the possibility of someevent occurring. When the reasons for a possible event to happen in reality outweigh the opposite reasons, the event is calledprobable, otherwise it is calledunlikely orimprobable. The preponderance of positive grounds over negative ones, and vice versa, may be in varying degrees, as a consequence of whichprobability(andimprobability) isgreater orlesser[1]. Therefore, probability is often assessed at a qualitative level, especially in cases where a more or less precise quantitative assessment is impossible or extremely difficult. Different gradations of probability "levels" are possible[2].
In our case, the classifier will give probabilities of assigning the target to 1 of 2 classes, which means that the higher the probability to buy the stronger the signal, if it decreases or increases, it is inhibition, amplification and whatever you want, ie again, maybe only 2 classes may be enough, and then interpret their results, or not? :)
http://radikal.ru/video/u5ibCyXdqGF
Sweet man, can you help me with my question? I'm confused :)
https://www.mql5.com/ru/forum/86386/page479#comment_5807576
It is clear that other states can be added. But still, wouldn't they be derived from the probabilities of belonging to one or another class (of 2)?
Probability is the degree (a relative measure, a quantitative assessment) of the possibility of someevent occurring. When the reasons for a possible event to happen in reality outweigh the opposite reasons, the event is calledprobable, otherwise it is calledunlikely orimprobable. The preponderance of positive grounds over negative ones, and vice versa, can be in varying degrees, as a consequence of whichprobability(andimprobability) isgreater orless[1]. Therefore, probability is often assessed at a qualitative level, especially in cases where a more or less precise quantitative assessment is impossible or extremely difficult. Different gradations of probability "levels" are possible[2].
In our case, the classifier will give probabilities of assigning the target to 1 of 2 classes, which means that the higher the probability to buy the stronger the signal, if it decreases or increases then it is inhibition, amplification and whatever you want, ie again, maybe only 2 classes may be enough, and then interpret their results, or not? :)
Are you in the mood for an argument? or are you still in the mood for reflection? You don't want to get out of the 2 classes of classification. Make sense of it.
Are you in the mood for an argument, or are you in the mood for reflection? You don't want to leave the two classes of classification. Make sense of it.
On reflection of course, yes, 2 classes is a comfort zone :) I just haven't got into it yet... if the price increments can be taken as probabilities (the greater the increment the higher the probability of the event), then the probabilities of 2 outputs can be interpreted in a similar way... why this cannot be done for me is not yet obvious
on comprehension of course, yes, 2 classes is a comfort zone :) I just haven't got into it yet... If price increments can be taken as probabilities (the higher the increment the higher the probability of the event), then the probabilities of 2 outputs can be interpreted in a similar way... why this cannot be done is not yet obvious to me
Gross error. So how can you take increments as probabilities??? You gave a definition of probability, but do you understand it yourself?
To make sense of it: What is the range of possible increments? Which of the increments within the range is the most frequent?
When you understand this part, you can move on.
Gross error. How can you take increments as probabilities? You gave a definition of probability, but do you understand it?
To make sense of it: What is the range of possible increments? Which of the increments within the range is the most frequent?
When you understand this part, then you can move on.
I look at it this way:
The increments that are close to zero indicate a 50\50 probability, i.e. we would give the 2 classes a probability of 0.5 each (uncertainty, neither buying nor selling)
respectively, we can normalize the increment from 0 to 1, where 0.5 would be uncertainty, >0.5 buy, < sell. The closer the value is to the extremes, the higher the probability of the event (higher price change).
After the model is trained, it gives the same values from 0 to 1, which can be interpreted as probabilities of occurrence of this or that event (the higher the probability, the stronger will be the change in absolute values)
The question is whether it is correct to interpret the probability of the target belonging as a probability of price change in absolute values (the higher the output probability the greater the change is expected)
The way I look at it is this:
Increments close to zero suggest a 50\50 probability, i.e., we would give the 2 probability classes 0.5 each (uncertainty, neither buying nor selling)
respectively, we could normalize the increment from 0 to 1, where 0.5 would be uncertainty, close to 1 buy, close to 0 sell. The closer the value is to the extremes, the higher the probability of the event.
After the model is trained, it gives the same values from 0 to 1, which can be interpreted as the probabilities of occurrence of this or that event (the higher the probability, the stronger the change in absolute values will be)
increments and probabilities are NOT the same thing.
In general, you should start with a textbook on probability theory.
increments and probabilities are NOT the same thing.
In general, you should start with a textbook on probability theory.
The more positive the increments, the more likely it is to be assigned to a class of by{sell}. The larger the positive increment, the more likely it is to belong to the bye class, is that clearer? And since everything above 0.5 is already a buy class, then the absolute increment increases with increasing probability