I've set up a model exactly as described by Thomas Ott of 'neuralmarkettrends' in videos 8-10 - and it's working well so far.
But what I would still need is the output of the probability for the predicted label (horizon = 1). The model only gives the average values in form of
prediction_trend_accuracy: 0.807 +/- 0.067 (mikro: 0.807).
As far as I understand the process is as follows:
- Order your data by date
- Split your data into two parts
- Use data before date X for training, use data after date X for testing.
- Features for training use created using windowing
- SVM is used as learner
* This process does not deal with horizons very well, neuralmarkettrends1 is aware of this fact, but does not want to complicate his video
Now to answer your question:
My suggestion would be to rescale absolute error to fall into range 0 to 1, and use this as a measure of probability.
This is the best answer I can give right now.
You need to provide better information to get a better answer.
You are right the question was a bit too unprecise, however you got it right that's the way I'm doing it.
Unfortunately I don't know what to do exactly regarding your answer "Now to answer your question:
My suggestion would be to rescale absolute error to fall into range 0 to 1, and use this as a measure of probabilit".
You should get a result looking like this:
( I have problems uploading images, will edit this image later, just go into results dataset and plot "predicted" and "label" and maybe "abs_pred_minus_label" ).
Try figure out why absolute error is different from average(abs_pred_minus_label)
Also note that I'm not using a fixed split, instead I'm using a sliding window validation, because this is the proper way to validate time series models).
This XML shows how you can use the Regression Performance Operator.
Thank you so much for your answer. Due to the fact that I'm a beginner I don't know how to import your data as a new operator into my process of video 8 to 10 & I'm not sure at which position of the chain to position this operator then.
Thank you wessel for your tips but I'm afraid it looks too complicated for me, I think I cannot handle (understand) it completely. Therefore I've created a PDF - file that you could view using this link:
Bild 1 shows my original process, Bild 2 is the content of the validation operator.
Bild 3 shows the general performance output.
Bild 4 is my latest progress :-) I've inserted the "Log - Operator" and defined here the values for performance and prediction accuracy.
Bild 5 shows the result of the latter.
My question is: Did I insert the Log - operator at the correct position in the process (Bild4) to be sure it delivers the performance of the predicted n+1 value, that's content of "Read Excel (2)" or do I have to rearrange / add something ???
As usual I'm looking forward to anybodies comments.