Machine learning in trading: theory, models, practice and algo-trading - page 2486

[JeeyCi]

JeeyCi #:

Mihail Marchukajtes

- Can you tell me what multipliers you have in Classifiers in the variable double decision are weights?

ah, these are probably initially set, which in the learning process are then self-tuned by the NSET ... i.e. initially, probably, random... (but is there any logic at least in their signs?)...

but the question of the logic of operations in functions and constants (why are they exactly like that and functions why exactly like that) remains?

[iwelimorn]

"Ehhh...If only someone would help..." ©

Could you tell me what algorithm can be used to separate these three classes, especially interested in the class designated in blue. This dog is divided into two separate parts, and unfortunately I do not know how to separate the marking of the target, so as to separate the right and left parts. Maybe something advise?

[Aleksey Nikolayev]

iwelimorn #:

"Ehhh...If only someone would help..." ©

Could you tell me what algorithm can be used to separate these three classes, especially interested in the class designated in blue. This dog is divided into two separate parts, and unfortunately I do not know how to separate the marking of the target, so as to separate the right and left parts. Maybe something to advise?

By eye, separated by two straight lines.

[Andrei Trukhanovich]

Judging by the picture, any classification algorithm can do it

[Mihail Marchukajtes]

JeeyCi #:

Mihail Marchukajtes

I got up the strength/courage to look through your code (often there is more truth in the code than in all textbooks) - can you tell me what are those multipliers in your Classifiers in the variable double decision - are they weights?... and how did you initially find them? i.e. why exactly those?

or better yet, comment please - what variables does it take, and the function code

thanks in advance!

p.s.

1. I see that you use sigmoid (S-shaped) as an activation function... it is "often used as a squeeze function"...

2.

maybe a squared one would be better?

That's right, you write the solution of the network, if you write it differently it would look like this.

double decision=1.5260326743246075*x0-0.13861638107404117 * sigmoid(x0)-0.06391652777916389 * sigmoid(x2)-0.44591870340615364 * sigmoid(x0 + x2)+0.14661031327032664*sigmoid(x3)-0.024191375335575492*sigmoid(x0+x3);

This is called a polynomial in mathematics, the coefficient multiplied by the value of the input plus another coefficient multiplied by the activation function with the values of another input or with the sum of the values of the inputs as shown below, minus the coefficient and so on...... As a result we get a number either above zero or below, which corresponds to this or that class, but for AI systems the method of trend classification is applied, when in addition to "Yes", "No" the answer "Don't know" appears as well. This is achieved by using two NS in one AI system, the so-called committee. What is remarkable the committee itself does not greatly improve the quality of the overall model. That is to make a committee of 5 or more models makes no sense, but two models would be fine, the effect of improving training is still present.

This code

double x0=2.0 *(v0+327.0)/650.0-1.0;

It normalizes input value, this is internal, technical normalization for polynomial before it goes directly into equation. The normalization itself does a range reduction. That is, it does not change the ratio of the series and the series itself looks like a perfect original, but after this normalization it starts to lie in the range of the maximum of the minimum of the current series. In general, the reduction to the range.

Regarding the asset function, yes there is a code of it, it is intended that the solution would not be linear in each neuron! Basically this is one neuron of the network

-0.13861638107404117 * sigmoid(x0)

It's not hard to calculate that this polynomial has 6 neurons and uses 4 inputs

[Mihail Marchukajtes]

JeeyCi #:

Mihail Marchukajtes

2.

Maybe it's better to squared?

Squared, we get the speed of change, how fast the variable has changed, and the standard difference will give us the degree of change, that is, the actual value of how much the smile has changed. That's why I'm talking about the smile and I still can't do it. On Ubuntu office under the wine crashes the system that then I can not boot, I think it's related to updating the DDE and recording, in general, I screwed up a lot when the boot problems began, but I was lucky and managed to boot and kind of fix it. What, but linux systems are stronger in recovery than vinux. If the probability of recovery wind 5-10% and linux about 30-40% of recovery. I respect linux a few years ago and I still do :-)

In general try to do on used data as less mathematical changes as possible, maximum plus for integration, minutes for finding out not only a sign of change but also a degree, how strong was this change, that's probably all, and then normalize center, scale etc.

[Mihail Marchukajtes]

iwelimorn #:

"Ehhh...If only someone would help..." ©

Could you tell me what algorithm can be used to separate these three classes, especially interested in the class designated in blue. This dog is divided into two separate parts, and unfortunately I do not know how to separate the marking of the target, so as to separate the right and left parts. Maybe you can advise something?

Look, if you can't create a target using code, when there are no clear rules for collecting it and we basically need to find it out. To be more exact, to find out whether or not the vector to be presented relates to the blue point, in that case you should use NS where no target is needed, like Kohonen's self-orgonizing maps or something of that series. There are types of networks that do not need a target, but after training they give out how many classes are in the sample, i.e. how many groups can be divided into the training sample. Or you can set this parameter forcibly. If you know for sure that you have 4 groups, then you force the sample into 4 classes, find the blue one and check it....

[mytarmailS]

iwelimorn #:

"Ehhh...If only someone would help..." ©

Could you tell me what algorithm can be used to separate these three classes, especially interested in the class designated in blue. This dog is divided into two separate parts, and unfortunately I do not know how to separate the marking of the target, so as to separate the right and left parts. Maybe you can advise something?

Send me the data and I will try it.

[Mihail Marchukajtes]

mytarmailS #:
Send the data, I'll try it

I was going to offer to try that too!!!

[JeeyCi]

Mihail Marchukajtes #:

Why am I actually already talking about the smile and still can not do.

By the way, yes, its change in dynamics would be more interesting (with a statement about which options are getting more expensive/decreasing due to demand, I guess) - as an alternative one can use a slope line (elasticity) on +/-Delta the same from central strike (better specifically from seattle Fut by linear regression)... imho (to simplify calculations)... But in the elasticity estimation variant the contribution of rt should be somehow neutralized anyway... And/or study the series in dynamics by dt - so that the skewness of rt variable (%*days till exp.) would not distract from it ... it is exponential after all

*****************

I keep thinking about the web model (c59) (in the context of striving for balance/disbalance)... the mathematics of the model scares me

Mihail Marchukajtes #:
In general try to do as little mathematical changes as possible on the data used, maximum plus for integration, minus to find out not only the sign of change but also the degree to which this change was strong, that's probably all, and then normalize, center, scale, etc.

Thank you... I'll try, because I used to automatically divide everything to get the ratio (e.g. Call-Put Ratio by price and/or volume)... really, apparently, there are other operations in math - just to model horizontally (aka by dt) to trace the dynamics