Machine learning in trading: theory, models, practice and algo-trading - page 854
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
For me? I've already solved the problem. Now I'm thinking about what else to do. Python or R. I don't have any new ideas yet.
So what is the way to evaluate the model? Or you always get the same model when you build the same set over and over again. Is this how it works?
So the task of the AI is precisely in the non-stationary series in which the pattern is floating. The task of the AI is to maintain performance when this dependence runs away, at least for an insignificant time, but enough to earn money. After all, the regularity does not change by leaps and bounds. In place of the main, first entry, there is another one, but the main one is still in the set, and here it is the AI that takes upon itself the load of holding the line, as they say. That's why in the first month of a futures contract, you have to train very often, especially when the market does not know where to go. Looking at the Vtrite, I can see this pattern dancing. But in the middle and at the end of the futures, as a rule, the market becomes more orderly and one entry dominates for a long time.
Man, everybody works from the moment or a couple of days before the previous futures close. What the hell is the first month?
the pattern varies chaotically and the deviations in the patterns increase exponentially with time
Any approximator (except, in part, RNN or LSTM) cannot solve such problems
all articles on statistics, with attempts to apply them to the market in their current form - can be thrown out and do not pay any attention to them
The main efforts should be focused on methods of working in a non-stationary environment, one of which was proposed by Alexander (provided you don't have signs that have a stationary effect on a quotient, that can't be extracted from the quotient itself, a-priori)So what is the way to evaluate the model? Or do you always get the same model when you build the same set repeatedly? Is that how it works?
Probably different, who knows. It learns from a random sequence.
Probably different, who knows. It's trained on a random sequence.
Well, how do you choose the right one??? Or do they all end up giving the same result on the feedback loop???
I have all models will work differently on the feedback loop....
Well, how do you choose the right one??? Or do they all end up giving the same result on the feedback loop???
I have all the models that work differently on the OOS....
I only have one model, the NS -60 neurons. I don't need to choose anything. We train, we work.
Yes, feedback - what is it?
I have only one model - NS-60 neurons. I don't need to choose anything. We train, we work.
Yeah, what's that?
Yeah... Geez...... You have one NS, but when you train it, you will always get different neuron weights. ALWAYS DIFFERENT. It will work the same in the training area. But every time you train it, you will always get a different NS, and the difference is in the neuron coefficients. So you can statistically determine that this NS with this particular set of rates will work better in the future than this one. Isn't it? Or I don't understand something. In P it's all written in...... as I understand it...
Yeah... Geez...... You have one NS, BUT when you train it you will always get different weight coefficients of neurons. ALWAYS DIFFERENT. It will work the same in the training area. But every time you train it, you will always get a different NS, and the difference is in the neuron coefficients. So you can statistically determine that this NS with this particular set of rates will work better in the future than this one. Isn't it? Or I don't understand something. It's just that in R it's all sewn up inside...... as I understand it...
I don't work in R.
Yes, it's always a different NS with every training. I check it, the only one, on an independent BP, and off to the real. By the way, for futures.
I don't work in R.
Yes, with each training is always a different NS. I check it, the only one, on an independent BP, and go to the real. By the way, on futures.
I've been testing it on an independent BP as well. I have a basic strategy that allows me to create such a BP without losing time. But as it turned out, it's better to apply the method calculated in my example. So, statistically more reliable to understand how much your model carries information about output....