Machine learning in trading: theory, models, practice and algo-trading - page 422
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
Dear professors and associate professors of programming, have you finished the code?
Can I try it? At least a trial.
Alyosha:
I repeat once again double-check your results with a random source, On a random walk arithmetic or geometric. With ZZ and other fake targeting you will get a prediction well over 50%, you can easily get 90%. You will have to prove that you can predict random, which is not reasonable.
Probably, you yourself understand what you are talking about, but alas, you explain yourself vaguely. I join the request that you should demonstrate on a specific dataset, on specific algorithms, what you mean, what is wrong and how it should be according to you.
Probably, you understand what you are talking about, but alas, your explanation is vague. I join my request to show you on a concrete dataset, with concrete algorithms, what you mean, what's wrong and how you want it to be.
Alyosha explained everything clearly. It's up to you to agree or disagree with his opinion.
Ah, well, if we are talking about accepting/disapproving a point of view, then forget it, I thought we were talking about something objective, such as the fact that the square of the hypotenuse is the sum of the squares of the cathetuses, then "prove", and you have "convince".
Mixing inputs with outputs is cheating yourself. The results of the training will be exorbitantly good, the work on the real will be exorbitantly unsuccessful. I've already mentioned more than once that using a zigzag is possible to identify the market trend, but it has no final value, hence all the problems when working with real time. There are classical output functions and if you get the result below 50% with them then your input is bad and in this case we should try to correct the input, but not the output.
The first classic target is forecasting of changes of Close one bar forward (believe me, it is enough). It's done simply: Lead((Close[0]-Close[1]),1) is used in approximating or predicting networks, where network response is value of predicted value. In such networks, the target is used ONLY in the learning process. It is not involved in network operation, because the network itself produces a value.
The second classic targeting is a classification of the color of the next candle for the recognition or classification networks. But it should also look into the future, otherwise it is of no use. To do this we label the up candle with 1 and the down candle with -1 and then shift it one bar backward when preparing the training set. Therefore we do not know the last value until the zero bar closes. We train the network and wait for its response on the future candle.
And now here is the essence. If your inputs can't go higher than 50% when training to any of the above mentioned bars, it means the input is shit, not the output is too complex. So you need to look for an input set that can do that. Raise it above 50%. The answer lies in a simple understanding of causality in relation to price. What is the cause and what is the consequence of the price. And there's a growth formula I've talked about more than once and it looks like this.
market expectations (option trading is the cause of volume) - trade volume (traded either in correspondence with expectations or not, but it is the cause of price movement) - price (changes according to the traded volume) - indicator (changes according to price movement)
So? Are you still using indicators???? Well, then you're at the end of the train, because indicators are a consequence of price changes, but not the cause. The indicator has become so because the price has become so. Now take a look at the following phrase........
The price has become so because the volume was traded so. Feel the difference????? Think about it........
And there's no need to coddle your grandmother :-)
With the right causal model approach, that is, when the inputs are the cause of the price, you can build ANY strategy. As complex or as simple as you want, it will work because it is the reason for Close. Here is an example.
I thought about the following question, which sounds as follows. "Will there be or will there not be a pullback when a signal appears?" It seems to be a problem not related to direct price prediction, HOWEVER!!!!!!
I have a data set for a general classification of the underlying strategy for true signal or false signal. And what I did, I changed the rollback targeting for the same input data and got a pretty tolerable model above 50%. The main thing here is to keep the targeting clean, without errors. As a result I can apply any target with a semantic value to one and the same input data set and get quite an adequate model, because input data is the cause for price and it is enough to answer such questions as "is the signal true or false", "will it rollback", "what will be the next candle after the signal", "will the price reach take profit" and so on.
And on the subject of tinkering, I'm surprised this issue even came up here.... It's like picking at the bough you're sitting on. Deceiving yourself......
Mixing inputs with outputs is cheating yourself.
You mean output data to input?
The price became that way because the volume was traded that way. Feel the difference????? Think about it........
This is like arguing about the egg and the chicken.) Imho, there is a mistake in understanding the cause and effect relation. We knowthe closing price at the end of the day, right? We know the traded volume too. So how can one be decisive for the other? We get both of these parameters at the same time - post factum.
Okay, I'll give you an introduction. Or as they say in such cases, learn the basics. Volumes are received in real time. For each bar, you know the traded volume, delta, maximum volume in the bar at a certain price, etc. You are confused with the volume at the end of the day, which is used to determine the market context.
You mean output to input?
Well, yes. You kind of brought up the point that if the output peeks into the input, you get unrealistically nice numbers in training, but not in real trading.