Neural networks, how to master them, where to start? - page 11
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
If you were doing 3D graphics and animation, would you be writing your own 3DStudioMAX?
These are different things...
I gave you a concrete example... try changing the error function in your favourite neural network program...
Softly speaking you want to say that 2DMAX is equal to Neuroshell in terms of functions and code size ????
well, well...
And don't you allow for the following idea - in order to get a profit with this program, you don't need to modify this error function in it?
Thanks to this gracious statement of yours, we now know that you have no or no concept of optimisation in your kinspecks from NS's best experts.
And you in turn do not allow the following idea - that you live in a country somewhat different from usay,
and therefore you are not supposed to have a software to fully optimize the creation and use of NN?
Thanks to this gracious statement of yours, we now know that you have no and no concept of optimisation in your kinspecks from NS's finest.
:-)
Did you understand what you wrote? )))
No matter how you twist the NS, whatever you give it to the inputs, there are no miracles, of course!
So, what turns out: On the one hand, the more the layering of NS, the higher its predictive power, but it is senseless to build more than three layers - three-layer grid is already a universal approximator. On the other hand, all the black box called NS does is find and exploit quasi-stationary processes in the market. There is no other way around it. Exactly, quasi-stationary, not stationary at all (there are simply no such processes on the market) and not non-stationary (such processes cannot in principle be exploited). I gave above a link to derive the optimal relation between the number of NS synapses - w, dimension of its input - d and the optimal length of training sample P (4 or 5 pages of the topic): w^2=P*d
Hence, the more is the complexity of the NS, the larger in length a training sample should be used for its training. Not only does the complexity of training increase as P^3, but the data may also be insufficient! But the biggest trap lies where you don't expect it - quasi-stationary processes (those that our NS identifies in the cotier and then exploits) have a characteristic lifetime (at all, different from zero and smaller than some). It is clear that on a large training sample, there is a higher probability of change of the selected process... you see? The shorter the training sample, the better - less chance of getting screwed by a change in market sentiment! Here, it seems the answer to the question "Which is better - a 2 layer NS with a short training sample, or a mighty 3 with three universities behind it (while learning, everything became unnecessary)?", will be given by a simple experiment.
For this, I threw three grids - 1,2 and 3 in Mathcad, and compared the results of predicting a sign of kotier increments one count ahead (I gathered statistics from 100 independent experiments). The results are as follows:
1 - p=10% correctly guessed signs (probability=1/2+p).
2 - 15-16%
3 - 12%
There are some free parameters here: dimension of input and number of neurons in the layer(s). First parameter was the same for all architectures, second was chosen personally. We see that 3-layer NS-grid is not a panacea, and perhaps for us as traders, the best option for the MTS analytical unit is a two-layer grid - from the viewpoint of maximum forecasting accuracy and minimum requirements for training complexity (power of RS, large history and its non-growth).
No matter how you twist the NS, whatever you give it to the inputs, there are no miracles, of course!
So, what turns out: On the one hand, the more the layering of NS, the higher its predictive power, but it is senseless to build more than three layers - three-layer grid is already a universal approximator. On the other hand, all the black box called NS does is find and exploit quasi-stationary processes in the market. There is no other way around it. Exactly, quasi-stationary, not stationary at all (there are simply no such processes on the market) and not non-stationary (such processes cannot in principle be exploited). I gave above a link to derive the optimal relation between the number of NS synapses - w, dimension of its input - d and the optimal length of training sample P (4 or 5 pages of the topic): w^2=P*d
Hence, the more is the complexity of the NS, the larger in length a training sample should be used for its training. Not only does the complexity of training increase as P^3, but the data may also be insufficient! But the biggest trap lies where you don't expect it - quasi-stationary processes (those that our NS identifies in the cotier and then exploits) have a characteristic lifetime (at all, different from zero and smaller than some). It is clear that on a larger training sample, there is a higher probability of the selected process changing... you see? The shorter the training sample, the better - less chance of getting screwed by a change in market sentiment! Here, it seems the answer to the question "Which is better - a 2 layer NS with a short training sample, or a mighty 3 with three universities behind it (while learning, everything became unnecessary)?", will be given by a simple experiment.
For this, I threw three grids - 1,2 and 3 in Mathcad, and compared the results of predicting a sign of kotier increments one count ahead (I gathered statistics from 100 independent experiments). The results are as follows:
1 - p=10% correctly guessed signs (probability=1/2+p).
2 - 15-16%
3 - 12%
There are some free parameters here: dimension of input and number of neurons in the layer(s). First parameter was the same for all architectures, second was chosen personally. It's clear that 3-layer NS-grid is not a panacea, and perhaps for us as traders the best option for MTS analytical unit is a two-layer grid - from the viewpoint of maximum forecasting accuracy and minimum requirements for training complexity (power of RS, large history and its non-growth).
I wonder if anyone tried to use NS for lottery number prediction?
gpwr, you're making fun of everyone! - it's about predicting the lottery number. And please remove the quotation of my post - it will add even more localism to your post :-)
I already gave above a link to the conclusion of the optimal relation between the number of NS synapses - w, its input dimension - d and the optimal length of the training sample P (4 or 5 pages of the topic): w^2=P*d