neural network and inputs - page 39

New comment
 

The quality of modelling in out of sample:
*
* TruePositives: 83
* TrueNegatives: 111
* FalsePositives: 96
* FalseNegatives: 47
* Total patterns in out of samples with statistics: 337
* The remainder patterns in out of samples without the statistics: 78
* Total errors in out of sample: 143
* Sensitivity of generalization abiliy: 46.36871508379888%
* Specificity of generalization ability: 70.25316455696202%
* Generalization ability: 16.621879640760895%
* Indicator by Reshetov: 0.004123184591376475
*/
double x0 = 2.0 * (v0 - 19.0) / 35.0 - 1.0;
double x1 = 2.0 * (v1 - 12.26) / 57.8300000000005 - 1.0
double x2 = 2.0 * (v2 - 21.0) / 13.0 - 1.0
double x3 = 2.0 * (v3 - 12.11) / 24.3800000000003 - 1.0
double x4 = 2.0 * (v4 - 18.0) / 40.0 - 1.0
double x5 = 2.0 * (v5 - 11.61) / 58.5 - 1.0;
double decision = -0.03426154451479057 + 0.09136531101334192 * x0 -0.16115463032514218 * x1 + 0.3774761240476446 * x0 * x1 -0.149536367886396967 * x2 -0.2182655506670959 * x0 * x2 -0.686972851164288 * x1 * x2 -0.7274492971348857 * x0 * x1 * x2 -0.06979110777265085 * x3 + 0.27356476016739995 * x0 * x3 -0.026662374360625248 * x1 * x3 + 0.12474514432879064 * x0 * x1 * x3 -0.291989483838501985 * x2 * x3 -0.2863737167793397 * x0 * x2 * x3 + 0.04656257824516221 * x1 * x2 * x3 + 0.11427907143112637 * x0 * x1 * x2 * x3 + 0.01709410880995815 * x4 + 0.21856958901169654 * x0 * x4 -9.925957720785493E-4 * x1 * x4 + 0.9723342991021926 * x0 * x1 * x4 + 0.04599384769467396 * x2 * x4 -0.05459813284687198 * x0 * x2 * x4 + 0.37290192411918303 * x1 * x2 * x4 + 0.010296169116858033 * x0 * x1 * x2 * x4 + 0.058584612082841506 * x3 * x4 + 0.531371391780234 * x0 * x3 * x4 -0.025018778838931215 * x1 * x3 * x4 + 0.1861984476159817 * x0 * x1 * x3 * x4 + 0.07319097184962621 * x2 * x3 * x4 + 0.0968827127374181818 * x0 * x2 * x3 * x4 + 01411041957291555 * x1 * x2 * x3 * x4 + 0.16417712916264263 * x0 * x1 * x2 * x3 * x4 -0.1726597989770004 * x5 + 0.36239224523375185 * x0 * x5 -0.008892292227349143328 * x1 * x5 -0.04417677147047251 * x0 * x1 * x5 -0.7319687377043317 * x2 * x5 -0.7806416997531924 * x0 * x2 * x5 + 0.01225632222209106843 * x1 * x2 * x5 + 0.04393711771649319 * x0 * x1 * x2 * x5 -0.006563651321672569 * x3 * x5 + 0.0627642424509067496 * x0 * x3 * x5 -0.015999570769395857 * x1 * x3 * x5 -0.05302786422005222 * x0 * x1 * x3 * x5 + 0.03534892871195049 * x2 * x3 * x5 + 0.1463193475694817 * x0 * x2 * x3 * x5 -0.027476124047644598 * x1 * x2 * x3 * x5 + 0.052884787352004865 * x0 * x1 * x2 * x3 * x5 -0.018202954537325178 * x4 * x5 + 1.0 * x0 * x4 * x5 -0.07118968415781378 * x1 * x4 * x5 -0.003138748792788926 * x0 * x1 * x4 * x5 + 0.2624137067639589 * x2 * x4 * x5 -0.02015595378617162 * x0 * x2 * x4 * x5 + 0.08019279607969382 * x1 * x2 * x4 * x5 + 0.06399649461673285 * x0 * x1 * x2 * x4 * x5 -0.02596308616804378 * x3 * x4 * x5 + 0.18361769860857746 * x0 * x3 * x4 * x5 -0.08407017017920377723 * x1 * x3 * x4 * x5 + 0.03014271917587724 * x0 * x1 * x3 * x4 * x5 + 0.07432306756805093 * x2 * x3 * x4 * x5 + 0.20722895875809277 * x0 * x2 * x3 * x4 * x5 + 0.0075079586507851345 * x1 * x2 * x3 * x4 * x5 + 0.20670493972886933 * x0 * x1 * x2 * x3 * x4 * x5

It's complicated..... How well the data describe the dependent variable????

 

I have submitted one training file. No test interval yet :-( But that's no problem.....

And as far as I understood it took 337 entries, although I submitted 600.

Here, I will try to reproduce the result in my environment. I wonder what will be the result, even on different machines?

Files:
reshetovtest.txt  23 kb
 
And about the data it is not quite clear how to take them to train the network on clean data..... How to select TruePositives: TrueNegatives: FalsePositives: FalseNegatives from the training sample and try to train the network. See what happens. Normalization of incoming data, that's a good thing.... I don't know how to use.... Just to make it look good...
 

The result matches....... Suppose we found 83 real positive examples. How do we separate them from the total sample...... And feed purely these 83 records naturally normalized. And if the network learns with minimal error to these 83 records. Then it will (theoretically) be able to classify such records in the input noise...... Like this....

 
nikelodeon:

I have submitted one training file. No test interval yet :-( But that's no problem.....

And as far as I understood it took 337 entries, although I submitted 600.

Here, I will try to reproduce the result in my environment. I wonder what will be the result, even on different machines?

VMR divides the total sample into two parts: training and control. I.e. if the total sample contains 600 examples, it means that 600 - 337 = 263 examples were included in the training sample, on which the model was created (trained), and 337 examples were included in the control sample, on which the model was then tested (but not trained).
nikelodeon:
And about the data, it is not quite clear how to take it to train the network on clean data..... How to extract from a training set TruePositives: TrueNegatives: FalsePositives: FalseNegatives and try to train the net.
There is no point in extracting anything from the training set. The training sample is only for creating the model, and the model is for other data that will not be in this sample, so VMR does all the calculations only on the control sample.
 

In fact, JPrediction was not created to predict financial instruments, but to predict the profitability of signals for the next month.

In other words, I compile a sample that includes current signal characteristics: number of trades, term, monthly profit %, percentage of profitable trades, percentage of losing trades, profit factor, Sharp's ratio, etc. Then I wait a month and mark with 1 a month profitable and 0 unprofitable trades.

Then I train the model on this sample and use it to forecast signals for the next month.

The idea is that signals are easier to forecast because they have a lot of additional useful information in addition to historical data. Financial instruments do not have any additional data other than historical data.

 
Wizard, what program do you use to interpret the data? E-excel???
 

So I'm thinking... how to interpret the result... to make it faster and certainly not by hand........

 
Reshetov:

In fact, JPrediction was not created to predict financial instruments, but to predict the profitability of signals for the next month.

In other words, I compile a sample that includes current signal characteristics: number of trades, term, monthly profit %, percentage of profitable trades, percentage of losing trades, profit factor, Sharp's ratio, etc. Then I wait a month and mark 1 mark for signals that have had a profit within a month, while mark 0 mark for unprofitable trades.

Then I train the model on this sample and use it to forecast signals for the next month.

The idea is that signals are easier to forecast because they have a lot of additional useful information in addition to historical data. Financial instruments do not have any additional data other than historical data.

I fully support that point of view, I have an indicator that gives signals. The same trades. I think it may be possible to run it in JPrediction too, but it is not clear how to choose the training interval? And it would be convenient to save the file with calculated indicator for each record.....As it is done by Vizard...... And the data itself so that it could be obtained.... On them you can try to train another network later..... That's it. Say Yuri, is this planned????
 
The most annoying thing is that Excel doesn't support such long formulas either :-(
1...32333435363738394041
New comment
Reason: