Machine learning in trading: theory, models, practice and algo-trading - page 1741

New comment
 
May I cut in, is the cluster, if I understand you correctly, the same price behavior found? If so, why is this behavior searched for at the formed search target - the trend? We should search before and after. I think that is where the dog is buried. I respect the division by time and the search, but it is not the answer to the problem, although the profit is there for sure.
 
Valeriy Yastremskiy:
If I understand you correctly, is the cluster found the same price behavior? If so, why is this behavior searched for on a search target - a trend? You should search before and after. I think that is where the dog is buried. I respect the division by time and the search, but it is not the answer to the problem, although the profit is there for sure.

This is a division into several conditional states, according to the specified characteristics

Let's assume the states - growth, decline, flat

For each state a strategy is written. When you switch clusters the strategies change, on new data

 

I'm trying to sb predict, does the type of distribution matter?

Is this a way to prognosticate?

int size 10000;
double A[]; ArrayResize(A,size);
for(int i=0;i<size;i++)
  {for(int j=0;j<15;j++)
     {rand();
     }
   A[i]=rand();
  }

As far as I understand, a simple regression of 2 coefficients is enough to predict?

Imho, it's peeking somewhere, you can't predict such outliers, or was that a joke? It's a grail then.

 
mytarmailS:
Mihail Marchukajtes:

Now the whole irony of a trading strategy relying on frequency transformations that takes place when we get a perfect circle at the end of the futures, there and any fool can build it. You just get the last week of the last well-functioning TS, and at the beginning you almost have to do it in the morning and then in the afternoon as well. It's work, well, if I had the opportunity at least once to run and look for these figures. I would be glad.

To me it is due to the correctness of the wave, and from wave theory, the correct wave fades correctly, so we have a short-term forecast. The wrong curve has a lot of waves inside, so a forecast without wave separation cannot be made.

 
Maxim Dmitrievsky:

This is a division into several conditional states, according to the specified characteristics

Let's assume the states - growth, decline, flat

For each state a strategy is written. When you switch clusters, the strategies change, on new data.

In principle, I understand it correctly. Could you elaborate on the signs. Increments (so far I have only seen them here), velocities, on how many bars, may we watch the price behavior on lower TFs inside bars of higher ones?

The reverse task is closer to me. The general strategy: Option parameters of advisors with not enough correct clustering rules will always have zones of loss, the beginning of which the not enough correct rules can not accurately determine.

It is more correct to opt for clustering rules, or rather parameters by which we divide states. Determine the states on the history and look at all parameters that can be invented at the beginning and at the end of a stable state and define what parameters have the same changes, or watch the parameters not one by one, but in pairs or threes.

 
Rorschach:

I'm trying to sb predict, does the type of distribution matter?

Is this a way to prognosticate?

As far as I understand, a simple regression of 2 coefficients is enough to predict?

imho, it's peeking somewhere, you can't predict such outliers , or was that a steal? It's a grail then.

here they are predicted from the previous hour, also thinned...i.e. they are there too

In the last article search for "Parameters for all cluster clocks are pretty close"

 
Valeriy Yastremskiy:

to me it is due to the correctness of the wave, and from wave theory, the correct wave fades correctly, so we have a short-term forecast. Irregular curve has a lot of waves inside, so a forecast without the separation of waves can not be made.

Does wave theory have a concept of right and wrong wave?



I tried to make predictors using spectrum analysis.

I tried to do it on eurodollar clocks (classification). I wanted to compare my results with my results of Vladimir Pereorvenok's articles, which I deeply respect.

As far as I saw his articles his maximal result was if I'm not mistaken 76% of correct predictions, mine was 80-82%, but I want to note that it was for tray crossvalidation, I cannot check it on the test because the target is built in a special way and the new data simply do not have marks.

But judging by the chart, the error is in the same ranges.

I also want to note that I have a suspicion that I have managed to kill the affect of model dying after training.


This is what the prediction graph looks like at 800 bars OOS after training

eur watchmaker

I dont know if i can make any profit on it, but it is what it is.

 
mytarmailS:

In wave theory, is there a concept of right and wrong wave?

This is what the prediction chart looks like on 800 bars of OOS after the training

I don't know if you can make any profit on it, but it is what it is.

The sine wave is usually considered))). And the circle comes out then))

It's a nice chart, you should try it.)

 
Valeriy Yastremskiy:

Beautiful chart, we should try it)

What is it, a lot of false signals, we need to strive for 95%

And I have ideas how else to improve the quality, but I do not know how to do it all

 
Rorschach:

What we see, the more stable the period, the more round. That is, I can take wavelets, for example, and get the same picture.

In general, it does not work very well. The picture below has good periods (2 and 3), but they are not very smooth and make the circle fall apart.

It sayshere that cssa is a ssa built on a neural network prediction. That's what I wrote before, lag can only be eliminated by prediction. In usual ssa most likely instead of prediction they duplicate last known price, but in cssa they build prediction with neural network.
Thank you very much for the information, well, bottom line. Did you make these pictures yourself or did you take them out from somewhere else, and if yes, what was the result of working with OOS?
1...173417351736173717381739174017411742174317441745174617471748...3496
New comment
Reason: