Discussion of article "Deep Neural Networks (Part VI). Ensemble of neural network classifiers: bagging" - page 3
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
I watched different elmNN videos, the idea is that the model gradually adds new hidden neurons one by one, calculating their weights with special formulae. A neuron with one hidden layer can describe a function of any complexity, so allowing the algorithm to use a lot of neurons in the hidden layer - it can add them until it reaches error = 0.
And to fight overfit, the maximum number of neurons should be reduced.
Here's another idea about creating an ensemble.
I have noticed that models with the highest results combined in an ensemble do not give the best results.
In one experiment I generated hundreds of models and tried to find the best ensemble by searching their combinations. As a result, such "best" ensembles often contained rather weak models, and attempts to remove these weak models or replace them only worsened the result.
Now I, for example, first find the best model that gives the best result. Then I look for the second model, which in ensemble with the first one will give even better result, and the result of this second model (one without ensemble) is not interesting to me. Then I add a third model, which in ensemble with the first two will give even better result. Etc. That is, I don't care about the result of a model when it is on its own, I care about how much it will improve the ensemble. It's like bousting.
Here's another ensemble idea for you.
I have noticed that models with the highest results combined in an ensemble do not give the best results.
In one experiment I generated hundreds of models and tried to find the best ensemble by searching their combinations. As a result, such "best" ensembles often contained rather weak models, and attempts to remove these weak models or replace them only worsened the result.
Now I, for example, first find the best model that gives the best result. Then I look for the second model, which in ensemble with the first one will give even better result, and the result of this second model (one without ensemble) is not interesting to me. Then I add a third model, which in ensemble with the first two will give even better result. Etc. That is, I don't care about the result of a model when it is on its own, I care about how much it will improve the ensemble. It's like bousting.
I thought you wrote that you do regression - only an average will do, voting is out.
I'm averaging, yes.
I wonder what the prediction result is on the forward?
What's the problem?
There are quotes, scripts. Repeat the calculations by shifting the beginning of the set by 500-1000 bars and get the result you are interested in.
According to my calculations, after training the ensemble with optimal parameters, at least the next 750 bars give good classification quality. For reference, 500 bars of M15 is one week. Re-training/re-optimisation takes about 20-30 minutes (depends on hardware and R release).
An example of such a test is given in the previous article.
Good luck
What's the problem?
There are quotes, scripts. Repeat the calculations by shifting the beginning of the set by 500-1000 bars and get the result you are interested in.
According to my calculations, after training the ensemble with optimal parameters, at least the next 750 bars give good classification quality. For reference, 500 bars of M15 is one week. Re-training/re-optimisation takes about 20-30 minutes(depends on hardware and R release).
An example of such a test is given in the previous article.
Good luck
You have ZigZag increments as your teacher. A function with many monotonic areas. These are not profit increments. That is why 70% of correctly guessed directions do not give an understanding of profitability and you need to test on quotes.
That's right. И ?
Taking into account market realities (spreads, commissions, requotes...), the accuracy of classification with ZigZag as a target should usually be much higher than 80%.
How did you determine/calculate this figure? If it is not a secret, of course.