Machine learning in trading: theory, models, practice and algo-trading - page 171

New comment
 

we need to give up the price in the representation of BP, it is the most ridiculous representation for MO, in the case of the market. imho...

But how to represent it, I don't know...

 
In classifying for truth or falsehood, what matters is not how the network has divided these concepts, but how it stably divides them in the future. And the division itself is not important, it is important that it is stable. This is from the area that stably draining TS is as difficult to make as pouring. For example, I do it this way: I train the network every day, according to the volume of the previous day and the open interest. Then for 3-4 signals I build the model (sometimes I invert signals to the opposite), so to say orient it towards the market, and voila, the network divides good from bad signals stably. This is what is most important....
 
BlackTomcat:

1) Do you have a time gap between these periods?

2) The pattern is exhausted, recognized, and exploited by so many. Because of this, it turns into an inverse pattern.

3) I am currently working on a TS that uses graphical methods. In my opinion, if there are any working patterns, it's here.

4) I would like to make some more details to my previous post. There I seemed to have gone over the analysis of individual bars. But actually this is not true. The analysis of individual bars has the right to exist, but these key bars usually do not lie in the area of the tops.

1) As far as I remember it is not, the picture is not new, I don't remember it anymore...

2) it's nice that I am not the only one who thinks so

If i'm not the only one who thinks so ... i even may enter the market with a stop in three ticks and take 1k2 , 1k5 in 50% of cases, but it's impossible to mathematically analyze it, so it's garbage

4) everything should be able to search

 

п

So if there is someone who is good at programming divergences, then you can try to implement such a tricky pattern and test it

 
mytarmailS:

3) Me too ..., I can even enter with a stop at three ticks and take 1k2 , 1k5 in 50% of cases, but it's impossible to mathematically analyze, so it's garbage.

I don't agree with you there. :) I am pretty sure that graphical methods can and must be formalized. Perhaps due to a certain complexity, there is a division between 95% and 5% of those who have succeeded. But if there is a way to success in the exchange, it lies exactly in this area. At any rate, I see a lot on the screen, although it does not protect me from mistakes. However, there is always an alternative scenario. And the good news is that if you recognize everything correctly and in time, it is not very difficult to switch to an alternative scenario, even with some (small) losses.
I also want to add that at some point I became very skeptical of all mathematical (indicator) methods. They bribe me with their simple realization, but in this simplicity lies their uselessness. The history of exchange trading is very old (one can even say "ancient"), and hardly in those days anyone was sitting and calculating stochastics and RSI. :) But drawing lines on a chart - that's easy. And if generations of traders were raised and trained on it, why should it suddenly stop working? With the advent of computer technology, everything has become more complex and faster, and now trend channels can be seen even in tick movements. But the fact that they still exist suggests that they should be used.
Graphical methods have another major plus: they show you the PURPOSE! They show you where (or should I say, where) the price is going. When you know where the target of a price movement is, the question of its direction disappears by itself.

 
BlackTomcat:
I don't agree with you there. :) I'm pretty sure that graphical methods can and should be formalized. Perhaps due to a certain complexity there is a division between 95% and 5% of those who have succeeded. But if there is a way to success in the exchange, it lies exactly in this area. At any rate, I see a lot on the screen, although it does not protect me from mistakes. However, there is always an alternative scenario. And the good news is that if everything is correct and TIME to recognize, it is not very difficult to switch to an alternative scenario, even with some (small) losses.
To formalize it is formalized, it is not possible to explain to the machine, I do not know exactly how
 

I won't make references, because several of the posts leave out one important detail.

The value of the target variable cannot match in time the value of the predictors, namely, the value of the target variable must be shifted backward. If it's 1 then it's one step forward, if it's 10 then it's ten steps forward.

The target variable, the teacher, must look ahead.

As an illustration of this point, an idea has been expressed here on the thread that more clearly highlights the nuance of the target variable looking ahead in relation to the predictors.

The point is this. Let's take reversals, such as the Machka. From these reversals on the history we move forward and mark the reversal in question in the past, after which the price has changed by a certain number of pips, for example by 100. We have found it. We take the next reversal and look for a change of 100 pips and thus form the teacher. This idea demonstrates very clearly the approach of formation of a target variable: a target variable must realize "looking ahead", which is quite realizable on historical data. It is the target variable that provides predictions from the model, not the application of the predict operator .

There is another important nuance to this idea. It is absolutely clear WHAT we predict: we predict the future growth/decline of the price by 100 points. This, for example, advantageously differs from ZZ, in which the teacher marks "1" for an upward knee and "0" for a downward knee. If you think about it, WHAT are we predicting?

So the requirements for the target variable are:

1. The target variable must look forward

2. There must be a clear understanding of WHAT we are predicting.

Seemingly obvious thoughts, but in practice it is not possible to implement them: either the boots are too tight, or something else interferes...

PS.

At my request, the idea was tested, but I could not find predictors for its implementation.

 
Alexey Burnakov:
I answer you both.

A model is worthless if it is evaluated on the data on which the model was chosen. EVEN if it is a period of data on which the model has not been trained.

Think about it.

There is 1) overtraining. This is when you catch up with the model on the training data to a state of near-perfection. On other data there is no generalizing ability.

And there is 2) selection bias (optimistic model selection). This is when the best model or committee is chosen on data on which the behavior of the model is ALREADY known. And again - even if it's a test segment.

The resulting reality is this. The untrained model selected by the crossvalidation test blocks (the kind that goes to the plus side on the test) is potentially fitted to the TEST. To reduce this effect, nested crossvalidation was invented. Already selected model (or committee) must still be tested on other data.

That is, it is a validation of the model selection method.

Once again, I also have dozens of models, and I also tumble through predictors and parameters. And these models go to a solid plus over a period of 8 years each! And that's the test period. But when the "best" models selected by the test are tested by deferred sampling, there are surprises. And it's called - Model fit crossvalidation.

When it's clear, pure experimentation continues. If it's not clear, you'll see a multiple drop in quality on the real. Which is observed in 99% of cases.

Alexey!

To my mind, you overestimate the importance of formal tools like "cross validation" or "committees of models".

When developing models, there should be an evaluation criterion, which has NOTHING, ANYTHING to do with the learning process.

Let me list these criteria:

1. Validation of the model at a time interval FOR the learning interval.

2. Running of the Expert Advisor that uses the model in the strategy tester. Moreover, the Expert Advisor has no MM, it is the most primitive one. No stops, Take Profits, etc.

If at any point the results differ greatly from those obtained during training, the model is REPROVED, i.e. predictors have no predictive ability for the target variable. These criteria do not say: what to do, how to change - these criteria say one thing: THE MODEL IS RETURNED.

PS.

For ardent supporters of MCL I note that without all those actions and tools, which are discussed in this thread, the tester gives no basis at all for speculating on the future behavior of a trading system. The tester says: "These are the results for this time period. That's all. The tester gives exactly one figure, such as the profit factor, which refers to a certain period of history. And the statistics can only be obtained in R. And the tester is the final part of model design, but not a substitute for the whole development process.

 
mytarmailS:

So if there is someone who is good at programming divergences, you can try to implement such a tricky pattern and test it

I have seen indicators in kodobase that mark divergences with lines.
 
SanSanych Fomenko:

Alexey!

1) As I see it, you overestimate the importance of formal tools like "cross-validation" or "model committees.

2) The tester gives exactly one figure, such as the profit factor, which refers to a specific historical period. And you can only get statistics in R. And the tester is the final part of model design, but not a substitute for the entire development process.

Mr. SanSanych,

You don't have to talk about committees, it's a special case in the model selection process. About validations, no, I don't overestimate it.

2) MT does not give distribution of statistics.

1...164165166167168169170171172173174175176177178...3673
New comment