Machine learning in trading: theory, models, practice and algo-trading - page 1116

 
Mihail Marchukajtes:

What do you need time for, if the data are mixed before training, unless you want to isolate a section of OOS... Don't sweat it, just show me the results...

I need it.

Results. See my thread. There they are on the case.

 
Vizard_:

(I ran for Validol)))

dataset of 42 examples! 42 examples!


don't even mention datetime in the time series!

 

See.

Divide the predictor into two parts: one part belongs to one class, and the other part belongs to another class. Draw a histogram of each half and combine.

So.









Different quality predictors, but all of them have a much better predictive power than before (from memory)

We need to introduce a measure of the distance between the histograms that will more realistically show the difference between them, which will be more accurate than as a picture.

 
itslek:

dataset for 42 examples


Look, I don't understand you.... If your AI is so cool that it can learn 1000 examples, then this sample will be like a nut. What's the problem?

 
SanSanych Fomenko:

See.

Divide the predictor into two parts: one part belongs to one class, and the other part belongs to another class. Draw a histogram of each half and combine.

So.









Different quality predictors, but all of them have a much better predictive power than before (from memory)

We need to introduce a measure of distance between the histograms that will more realistically show the difference between them, which will be more accurate than as a picture.

Great... continues. We need the result of a trained model. Data analysis is good, but the most important thing is profit, if I'm not mistaken of course. This is why I'm asking you to trade them, if you got it ...

 
Mihail Marchukajtes:

Look, I don't understand you.... If your AI is so cool that it can learn from 1000 examples, it will be like a walnut. What's the problem?

Actually, it's the other way around...

better a worse algorithm but more examples than a cool algorithm but less data.

Even 1000 is not enough, especially for the market...

 
itslek:

Actually, it's the other way around...

agree.... It depends on which AI tool to use. Some need a large sample size, and some, like vector of reference vectors, don't need a large sample, because the method is resource-intensive and with a large sample works extremely long...

 
Mihail Marchukajtes:

agree.... It depends on which AI tool to use. Some require a large sample size, while some, like a vector of reference vectors, do not need a large sample, because the method is resource-intensive and with a large sample works for an extremely long time...

What is AI?

 

Not bad results in predictive ability will NOT lead to stable models, because simply ridiculous number of observations = 51. We need at least 10 times as many, or better, 100 times as many.

If you build models on that number of observations, the results are abysmal.


Predicted

Actual [0,0] (0,1] Error

[0,0] 42.9 28.6 40

(0,1] 28.6 0.0 100


Overall error: 57.1%, Averaged class error: 70%


Rattle timestamp: 2018-10-18 21:29:39 user

======================================================================

Error matrix for the Linear model on Mic1.txt [validate] (counts):


Predicted

Actual [0,0] (0,1) Error

[0,0] 1 4 80

(0,1] 2 0 100


Error matrix for the Linear model on Mic1.txt [validate] (proportions):


Predicted

Actual [0,0] (0,1] Error

[0,0] 14.3 57.1 80

(0,1] 28.6 0.0 100


Overall error: 85.7%, Averaged class error: 90%


Rattle timestamp: 2018-10-18 21:29:39 user

======================================================================

Error matrix for the Neural Net model on Mic1.txt [validate] (counts):


Predicted

Actual [0,0] (0,1) Error

[0,0] 2 3 60

(0,1] 1 1 50


Error matrix for the Neural Net model on Mic1.txt [validate] (proportions):


Predicted

Actual [0,0] (0,1] Error

[0,0] 28.6 42.9 60

(0,1] 14.3 14.3 50


Overall error: 57.1%, Averaged class error: 55%


Rattle timestamp: 2018-10-18 21:29:39 user