Machine learning in trading: theory, models, practice and algo-trading - page 1856

New comment
 
Maxim Dmitrievsky:

You'll be surprised if you do it right

Confusion Matrix and Statistics

          Reference
Prediction   1   2   3
         1 266  72  15
         2  54 130  50
         3  17  55 293

Overall Statistics
                                          
               Accuracy : 0.7237          
                 95% CI : (0.6941, 0.7519)
    No Information Rate : 0.3761          
    P-Value [Acc > NIR] : <2 e-16          
                                      
 
mytarmailS:

is this on oos? ok, but it could be higher

 
Maxim Dmitrievsky:

is it on oos? ok, but it could be higher

yes oos

with two classes / clusters is better

Confusion Matrix and Statistics

          Reference
Prediction   1   2
         1 381  79
         2  67 425
                                         
               Accuracy : 0.8466         
                 95% CI : (0.8222, 0.869)
    No Information Rate : 0.5294         
    P-Value [Acc > NIR] : <2 e-16         
                                         
                  Kappa : 0.6927         
                                    



Data - two hours 5 mins from opening, diffused and then into cumulative sum

clustered with kohonen - Data - rows for 2 hours and 5 minutes

clustering by forest - Data - rowsfor first hour and 5 min.

 
mytarmailS:

yes oos

with two classes / clusters is better



Data - two hours 5 mins from opening , diffused and then into cumulative sum

clustered with kohonen - Data - rows for 2 hours and 5 minutes

clustered by forest - Data - rowsfor the first hour and 5 minutes

there you go, and you were yelling.

 
Maxim Dmitrievsky:

There you go, and you were yelling.

how did you get 100%? what are your signs?

 
mytarmailS:

how did you get 100%? what are the signs you have?

on the trail 100, I wrote, on the oos 0.84, the same approximately (3 clusters)

signs themselves prices relative to the first. The video says it all from minute 25

From here on, you can use it without forest, just predict the cluster by clustering algorithm. The forest was used to make it easier to transfer it to the metaque. Then in the tester to race different strata on clusters.
 
Maxim Dmitrievsky:

on the trail 100, I wrote, on the oos 0.84, the same approximately (3 clusters)

signs themselves prices relative to the first. The video says everything from minute 25

I can use it further without forest, just predict the cluster by the clustering algorithm. The forest was used to make it easier to transfer it to the metaque

Yes , I agree , you can do everything by clustering .


Proclustered Kmeans , the result is better than kohonen by +- 10%

Kmeans 

 Reference
Prediction   1   2   3
         1 106  40   0
         2  53 590  42
         3   1  22  98

Overall Statistics
                                          
               Accuracy : 0.834    

kohonen

Confusion Matrix and Statistics

          Reference
Prediction   1   2   3
         1 266  72  15
         2  54 130  50
         3  17  55 293

Overall Statistics
                                          
               Accuracy : 0.7237    


But it is not clear how clusters are distributed, I see from the matrix that the cluster algorithms see quite differently

 
mytarmailS:

Yes, I agree, you can do everything by clustering.


I clustered Kmeans, the result is better than Kohonen by +- 10%.

Kmeans

kohonen


But it is not clear how clusters are distributed, I can see from the matrix that the cluster algorithms see quite differently

I have a kameans as well. It is usually better and faster than others. Apparently the problem is simple.

I've tried Gaussian mixtures - they show shit, worse. Although I haven't tried them for this task yet.

Visualize the clusters and look at the statistics for different methods, you'll see which is better. I like kamins for boxplots.

 
Maxim Dmitrievsky:

I have a Kamins too. It is usually better and faster than others. Apparently the challenge is simple.

Tried the Gaussian mixes - it's not as good, it's worse. Haven't tried it for this task yet though.

Visualize the clusters and look at the statistics for different methods, you'll see which is better. I like kamins for boxplots, in principle.

I do not like kamins, some very bad balance does

1    2    3 
 529 3093  378

"2" is a flute

kohonen like funnier

1    2    3 
1243 1497 1260
 
mytarmailS:

I don't like kamins, it's a very bad balance of classes

"2" is a flute.

kohonen more fun

show me pictures of clusters on kamins and kohoos, so it is not clear

1...184918501851185218531854185518561857185818591860186118621863...3475
New comment
Reason: