Machine learning in trading: theory, models, practice and algo-trading - page 1933

 
Aleksey Vyazmikin:

Got this picture from the past results - 4 clear clusters

Earlier you talked about some numerical values of the results evaluation - what numbers do you suggest to look at?

Yeah)) and now look at the test data

scatter3d(x = predict.test[,1], 
          y = predict.test[,2], 
          z = predict.test[,3],
          groups = Y[test.idx],
          grid = F, 
          surface = F,
          ellipsoid = F,
          bg.col = "black",surface.col = c(2,3))

I said it's all nonsense, you need to look at the error of the model on the test data, not the beautiful pictures, the error of any model from any package


By the way, change the indices because you're training on 10k data

train.idx <- 100:8000
test.idx <- 8001:10000

and you have over 500k of data

 
mytarmailS:

Who knows at least something about polynomial or harmonic approximationPLEASE!!!

 
mytarmailS:

Yeah)) and now look at the test data

I told you it's all nonsense, you have to look at the model error on the test data, not the pretty pictures, the error of any model from any package


Now it's the test data that's causing the error again.

> library(car)

> scatter3d(x = predict.test[,1], 
+           y = predict.test[,2], 
+           z = predict.test[,3],
+           groups = Y[train.idx],
+           .... [TRUNCATED] 
Error in complete.cases(x, y, z, groups) : 
  не все аргументы имеют одинаковую длину
mytarmailS:

You're only training on 10k data.

and you have over 500k data.

I tune it first on small data and then on large data.

By the way, I noticed that R is very irrational in its memory usage.

 
Aleksey Vyazmikin:

It's fighting again on the test data now

of course it does

library(car)

> scatter3d(x = predict.test[,1], 
+           y = predict.test[,2], 
+           z = predict.test[,3],
+           groups = Y[train.idx],
+           .... [TRUNCATED] 
 
mytarmailS:

Of course it does.

Oh yes, thank you!

Here's the picture - a little worse, but the trend is there.

I just don't understand what's going on - where did another cluster come from, if there wasn't one before?

 
Aleksey Vyazmikin:

Oh, yes, thank you!

Here's the picture - a little worse, but the trend is there.

Only I do not understand what is happening - how come there is another cluster, if previously it was not?

Wow, just a little)))

new data - new points

 
mytarmailS:

(a little bit)))

New data, new points.

I don't understand what's going on.

We trained something there and then applied it to the test or what? :)

It seems like it's all nonsense.

 
Aleksey Vyazmikin:

It seems like it's all nonsense.

How many times have I written this to you? 3? 5?

 
mytarmailS:

Who knows at least something about polynomial or harmonic approximationPLEASE!!!

And what do you want to end up with?, with harmonics the way to Fourier. Or to orthogonal polynomials, there are many of them. All can be decomposed)

 
I read the answers to my question. Thank you all very much!