Machine learning in trading: theory, models, practice and algo-trading - page 483
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 see :) so, all the same, probability of assignment to a class on the output has nothing to do with probable value of increments... you see here some people get confused too, for newcomers this is an ambiguous point... On the other hand, if a neural network outputs probabilities (e.g. softmax layer), what for do we need them if belonging to class will be defined by probability more than 0.5 all the same? And then you can really try to use regression model and get rid of all normalization of output values... by the way, I use a random forest where inputs normalization is not needed
Fuck, and he's drawing pictures here))))
x = increments (the first difference)
target = x > 0
target = f(x)
target = x > 0
lloss=0
probability=1
Yes, thank you, I really do not understand whether it confirms or refutes my hypothesis, most likely refutes :)
Probabilities, look in the code of your forests, it's just the percentage of trees that voted for one class or another.
That makes perfect sense :)
Yes, thank you, I really do not know what the hell it confirms or refutes my hypothesis, most likely refutes :)
I used these .
did not even understand the code
48 parrallel saw what I found. I don't remember, I killed Arima with the data I used, and then I did it again adding time and days of the week or something.
I'm now experimenting to go into pure MO, without the old forex habits. No plotting profit charts to evaluate the pattern, no wands or other indicators. Instead, regression (gains per bar ahead), complex crossvalidations and special patterns. I managed on eurusd m5 to train the model on 10,000 bars of history, get r^2 about 0.001, or almost 52% accuracy, and stay with the non-deteriorating result in the future. At zero spread it looks good, even a spread of 2 five digit points would pull in profits. But that's not enough.
There is also a strong fear that dealing centers have seriously started to destroy working Advisors and this entire search for the universal grail makes no sense. There is one dealing company that has been destroying working EAs for a couple of years already and even if the EA has been bringing profit for months, all of them have lost the balance within a week. This time there are more and more strange dealers, and I have only one dealer left, whom I trust. These are troubled times to come.
And there is also a strong fear that dealing centers have seriously started to destroy working EAs and this whole search for the universal grail makes no sense. I have one dealing company that has been destroying working EAs for a couple of years already and even if an EA has been bringing profit for months, all of them have lost their balance in a week. This time there are more and more strange dealers, and I have only one dealer left, whom I trust. These are troubled times to come.
If the system cannot be beaten by prosklazovyy, they draw a spike, how many times has this happened?
But there is an exchange.
I'm now experimenting to go into pure MO, without the old forex habits. No plotting profit charts to evaluate the pattern, no wands or other indicators. Instead, regression (gains per bar ahead), complex crossvalidations and special patterns. I managed on eurusd m5 to train the model on 10,000 bars of history, get r^2 about 0.001, or almost 52% accuracy, and stay with the non-deteriorating result in the future. At zero spread it looks good, even a spread of 2 five digit points would pull in profits. But that's not enough.
There is also a strong fear that dealing centers have seriously started to destroy working Advisors and this entire search for the universal grail makes no sense. There is one dealing company that has been destroying working EAs for a couple of years already and even if the EA has been bringing profit for months, all of them have lost the balance within a week. This time there are more and more strange dealers, and I have only one dealer left, whom I trust. These are the dark times to come.
I have not got to the experiments with TS, but the results of the trained NS on random samples are impressive. Only a 6% classification failure rate of -(0.1). It seems in this thread earlier gave exact figures.
But learn long and painful - pure time - 23 hours, not counting the intermediate checks, adjustments. For 3 months of work training enough, beyond that was not checked.
Imho, MLP is quite enough.
I did not get to the experiments with TC, but the results of trained NS on random samples are impressive. Only 6% of unsuccessful classification -(0.1). It seems in this thread earlier gave exact figures.
But learn long and painful - pure time - 23 hours, not counting the intermediate checks, adjustments. For 3 months of work training enough, beyond that was not checked.
Imho, MLP is enough.
How many layers/neurons are there in backprop? That's why I switched to scaffolding, because classical NS is slow
You're learning by backprop? How many layers/neurons? That's why I switched to scaffolding, because the classical NS is slow.
51 neurons. Configuration -15,20,15,10, 5,1. Yes, BP training with simulated annealing.
You definitely saw the results earlier. I can repeat, if anyone has a desire to see.
In general, if even once every 3 months to train, then 23 hours - yuck, compared to scratching a turnip with the usual TC on the logic.
I still have a slow computer, on a modern computer at least 2 times faster will be.
I see the work is booming... Well done!!!
Vizard huge respect for the video dedicated to me!!!! I really do not understand what is in them and what they are for, but the man is trying, I can`t thank him enough :-)
I`m going back to MT5 again. I have two questions on the agenda.
1. If i´m ready to use the indicator for MT5, it may be so because the copying has not helped me :-() I´m afraid that the indicator will be redesigned for МТ5.
2. How to make the calculation of the indicator after the update of all related indicators? Or at least a delay in the calculation for 30 seconds....
I do not understand this OnTimer function.
Thank you!