Machine learning in trading: theory, models, practice and algo-trading - page 427
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
In principle, it doesn't matter. The predictive ability there is zero....
What does the forecast have to do with the result? It's like looking at equity with your eyes and dividing profit by drawdown, it will be approximately SR
The worse the forecast, the higher SR
So what is the forecast, it is an assessment of the result, it's like looking at the equity and profit divided by the drawdown, it will be about SR
Well, the steeper the forecast, the higher the SR
And in this case. Well, then yes. I completely agree....
Can you tell me, here you wrote https://www.mql5.com/ru/blogs/post/160258 that RF is the best classifier among the proposed ones, I agree with that. And what to do with 3rd generation diplerning (with autoencoders), did you have any comparative tests?
About GARCH as far as I know it is a linear model to predict volatility, it does not predict the market direction, or am I wrong?
GARCH is a volatility model, it predicts the direction and size of the next candle. The timeframe may be different: M5, M30, D1 or any. But usually it is the small TF up to the highfrequency trade.
I find GARCH attractive because in order to apply it I need to analyze the initial quote to obtain the proof that the trained GARCH will behave the same on history and in future.
The idea is to fight with non-stationarity (variable mean and variable deviation from the mean), which is considered the main evil for the random performance of Expert Advisor in the future.
The model itself has three components:
Initially, price increments are calculated. Then formulas are written for this increment:
1. for the average behavior
2. for the deviation from the average (volatility). There are a great number of variants, including exponential and threshold variants (I mean the linearity of the model).
3. the law of distribution of the average.
If we managed to "correctly" choose parameters of all these parts, then the residue should be normally distributed, which is a guarantee of the model behavior in the future similar to the historical data.
Something like this, very roughly.
Tell me, here you wrote https://www.mql5.com/ru/blogs/post/160258 that RF is the best classifier from the proposed, with this I agree. And what about 3rd generation diplerning (with auto encoders), did you have any comparative tests?
I heard that diplinking is a little different, it's all about automatic search of SIMPLE features, in hierarchical structures, such as pictures or text, where from the edges of the corners and lines of which simple shapes are more complex, and so on. As for time series this is not quite the case. Thus if deep convolution neural network abruptly finds cats and faces on pictures, it does not mean that it will abruptly predict the market, besides all "deep" is a huge hassle for users, it is a huge hassle even for those who recreated the whole network from scratch on C++ 10 times by themselves, a lot of parameters are very unstable, the results are random. This is something you have to do for many years. But this is not my personal opinion, I haven't experienced and probably will not.
This is not a bad result, just a fantastic result, I'm sure even on Renaissance it's not even close, with their terabytes of data per day. Look at live-score on numer.ai and think why they have at least 45% error(logloss~0.69) and you have 30%.
But what you're saying is true, you've created your synthetic target function, which is functionally bound to features in a weird way (obviously not very obvious for you) and you have such a nice scan on Lern and test, everything looks right... But why aren't you a billionaire yet, although you could easily become one in about a year, if you had 30% error predicting color of the next candle, and it's because you predict not future but past mixed with future, by indicator. Try predicting a pure future returnee and everything will fall into place.
You have said a lot, mister, but never said anything in fact, and you were immediately offered to show, instead of tiresome shouting to everyone what is wrong with ZZ and what is needed.
https://ru.stackoverflow.com/a/683632/1084
I heard that diplerning is a little different, it's all about automatic search for simple features, in hierarchical structures, such as pictures or text, where the edges are corners and lines of which are simple shapes of which are more complex, and so on. As for time series this is not quite the case. Thus if deep convolution neural network abruptly finds cats and faces on pictures, it does not mean that it will abruptly predict the market, besides all "deep" is a huge hassle for users, it is a huge hassle even for those who recreated the whole network from scratch on C++ for 10 times, a lot of parameters are very unstable, the results are random. This is something you have to do for many years. But this is not my IMHO, I have not tried it on my own skin and probably will not try.
The devil is not as frightening as it is made to look.
It's not very difficult, start with these few articles(1, 2, 3, 4). Not everything will work at once and will be clear, but it will be useful.
Good luck
I heard that diplerning is a little different, it's all about automatic search for simple features, in hierarchical structures, such as pictures or text, where the edges have corners and lines with simple shapes of which are more complex, and so on. As for time series this is not quite the case. Thus if deep convolution neural network abruptly finds cats and faces on pictures, it does not mean that it will abruptly predict the market, besides all "deep" is a huge hassle for users, it is a huge hassle even for those who recreated the whole network from scratch on C++ 10 times by themselves, a lot of parameters are very unstable, the results are random. This is something you have to do for many years. But this is not my IMHO, I have not tried it on my own skin and probably will not try.
I see, I think in practice, no one here and did not compare :) I will seek information, that in the end, not to be fooled if it turns out that the diplerning does not give advantages over woods. And since the component part there is an MLP, it may well be that it does not...
By the way, diplerning is anything that has more than 2 layers, MLP with 2 hidden layers is also diplerning. I was referring to deep nets, which Vladimir described in the article at the link above.
Although they write that predictors are most important, as models work approximately the same... but this is theory, in practice it turns out that the choice of model is also very important, for example, a compromise between speed and quality, because NS is usually long...
For example, I've tried to buy a copy of MT5 for the first time, but in practice it is not so good as to make it faster and more convenient. I want to rewrite the neural network I need on mql with C++ and that's all.
ah, forgot to add IMHO
Ahem...ahem...I've been pondering and pondering...
There are a lot of models, from classics like ARMA/GARCH to complete breakthroughs like CNN/LSTM, somewhere in the middle of everyone's favorite MLP\RF\XGB and so on. But how to unify it all? So that, for example, we could exchange / sell trained models in a universal format, readable anywhere without special libraries and training, who tried to pass a non-trivial model to someone will probably understand what I mean)))
And then I realized that numer.ai - solved this problem! Actually, it should be so, otherwise how could you get fiches from future to give an answer, fiches are in real time, you can not know them beforehand. But if you fill N-dimensional space grid of points, then the prediction of a new point, just the closest(s) of the grid, which are sought instantly as an element of the array index. And any simple and tricky model can be folded into such a model, easily transferred and used without revealing its secrets.
Respect numer.ai, the idea is super!
I propose to think over how it can be used in our miserable community)))