Machine learning in trading: theory, models, practice and algo-trading - page 608
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
sounds just like the ensemble from NS
if I'm not mistaken
Then we choose model parameters (activation function, number of layers and their sizes, etc.) each time we do all these steps (train 5 models, predict 5 pieces unique for each model, combine them, R2), achieving better and better estimates.
That's hundreds of networks to build and train! But so far I don't see any other option(
I think if you write to him and offer a thousand bucks an hour, you can take individual lessons from Perchik, Perepelkin will be a little more expensive, but it's worth it
seriously about perepchik?)
Paying a grand to a hired driver.
I don't know anything about the other one... and I don't want to know.)
i think it's interesting, but i'm gonna go to bed and finish reading my books this week
sounds like just an ensemble from the NS
if I'm not mistaken.
At the end you will get a regular ensemble, yes. But the result will be much better than compared to "just train 5 neurons on the whole table".
Well yes, standard option, although I prefer without kv, I wrote already...
Doc, try fixing the steepness with different parameters and test.
I have LibreOffice, it did not work that neuronics.
That's hundreds of networks to build and train! But so far no other option is visible(
That's why I like gbm package in R, for example, its training speed is orders of magnitude faster. This is not neuronics, this is scaffolding and boosting.
It's also interesting, that k-fold crossvalidation worked fine for me even with small number of epochs of neural network training. The number of epochs was one of the training parameters that I selected. Small number of epochs = fast learning, that's a plus. But the possible accuracy of the model is lower, that's a minus.
I suggest you learn how to k-fold crossvalidation. I have seen several different ways, this one works well -
...
There is also a nuance that the initial weights of neuronka are set randomly, and the final result of training may very much depend on it, including for forest and other models.
Each time before training the model, I set the value of the random number generator to the same state:
set.seed(12345)This way I get reproducible results and stability. The gpsc grain value can also be picked up instead of 12345, which sounds pretty funny, but is sometimes necessary.
There's also a nuance that the initial weights of neuronka are set randomly, and it may very much depend on the final result of training, including for the forest and other models.
Each time before training a model, I set the value of random number generator to the same state:
This way I get repeatability and stability. The gpsc grain value can also be selected instead of 12345, which sounds rather funny, but sometimes it is necessary.
Throw your network in the trash, if it reacts so to the values of gpsc. Normal network works and learns at any initial values, even at zero.
If you give that neuron a lot of neurons and layers and infinite epochs of learning, it will easily train to the desired accuracy with any initial grain.
I, for example, am learning to predict the increase in price per bar. The problem is that there is a lot of noise in prices (real price +- some random shifts), and it is impossible to predict noise. But I can use crossvalidation to pick parameters where the model will not yet remember noise, but will generalize that data somehow and make correct predictions at least in a small percentage of cases. And with some initial values of weights model immediately starts to remember noise instead of trying to generalize these data, it's bad, and then you should look for another initial grain for initialization of weights.
Throw your network in the trash if it reacts so much to gpsh values. A normal network works and learns at any initial values, even zero.
This is just another explication that you can't use static methods in dynamic systems.
this is just another explication of why you shouldn't use static methods for dynamic systems. i also use setseed, any MOH fluctuates a lot from time to time, at least i can see it in the residuals.
Well let's go now look at architecture for learning on time lines rather than pictures of SEALs, option one:
Variant 2:
That is, the combination of NS and automata seems a good solution, with positive and negative circuits, and who and how will implement it - another question. Personally for me this approach is the most obvious