Machine learning in trading: theory, models, practice and algo-trading - page 3499

New comment
 
Maxim Dmitrievsky #:
So don't rely on crowdsourcing.
Yo gad dem rait
 
Candlesticks are a very versatile thing. They are essentially a lossy compression algorithm.
You can look at harmonics through the prism of candlesticks, you can even look at trades as candlesticks, it's the same as opten hi-close, isn't it? And hi-close is the same mfe if I'm not confused. Or you can aggregate equity into candlesticks, for example, by week or month, and candlestick panner will show all the information on trading and account drawdowns.

 
mytarmailS #:

Yes, that's right, if you change the trend on the ticks, you're fine.

But it is difficult to work with ticks, I need to think how to work with candlesticks directly.

The morning is the morning.

So I solved the problem of how to change the trend on candlesticks without going into ticks.

We have prices

     Open High Low Close
[1,]    1    6  11    16
[2,]    2    7  12    17
[3,]    3    8  13    18
[4,]    4    9  14    19
[5,]    5   10  15    20

slicing prices by rows into one vector

 Open  High   Low Close  Open  High   Low Close  Open  High   Low Close  Open  High   Low Close 
    1     6    11    16     2     7    12    17     3     8    13    18     4     9    14    19 
 Open  High   Low Close 
    5    10    15    20

vector

1  6 11 16  2  7 12 17  3  8 13 18  4  9 14 19  5 10 15 20

change the trend in this vector and reassemble in reverse order back to

     Open High Low Close
[1,]    1    6  11    16
[2,]    2    7  12    17
[3,]    3    8  13    18
[4,]    4    9  14    19
[5,]    5   10  15    20
 
mytarmailS #:

The morning's the morning.

So I solved the problem of how to change the trend on candlesticks without going into ticks.

We have prices

Cut prices by rows into one vector.

vector

change the trend in this vector and reassemble in reverse order back to

You can also change trends piecewise, for example, every n bars. Maybe add some variability.

If we take into account that MO learns better on initial prices, we can add trends that go beyond the known price range. So that the model can work in the future.

Or move the chart up/down by some value.

Then you won't need signs, just prices.

There needs to be some pretty serious research on this topic for MOE tasks.

 
It is possible to "residualise" the original graph using augmentation. This is the most interesting direction, in my understanding.
 
I have tried level shifting and flips. This results in worse signals on the trayne, confusion. But sometimes it improves the OOS.
 
Maxim Dmitrievsky #:
It is possible to "residualise" the original graph using augmentation. This is the most interesting direction, in my understanding.

I don't know what that is.

For me, "rationalise" is to reduce the series to the form of harmonics with unchanging characteristics - frequency and amplitude.

then training and prediction will be trivial.

 
Maxim Dmitrievsky #:
I have tried level shifting and flips. This results in worse signals on the trayne, confusion. But sometimes it improves the OOS.

I don't know, I would only test for OOS.

 
mytarmailS #:

I don't know what it is.

For me, to "stationise" is to reduce the series to the form of harmonics with unchanging characteristics - frequency and amplitude.

then training and prediction will be trivial.

Usually one differentiates a series to achieve stationarity. Here we get the reverse process - augmentation to achieve stationarity of the initial series and training on it afterwards.

has a place, in theory. That is, there is no need to extract features now.
 

I continue my experiments, the last post was here.

Let me remind you briefly, the essence is to take a sample, split it into quantum segments, evaluate these segments for stability of the probability shift relative to the binary probability of all rows of the subsample. We select from the selected set of quantum segments the quantum segment that we consider to be the best candidate for selecting the region with negative class. We remove from the sample the range that fell into the quantum segment and repeat the iteration. In essence, this is very similar to building tree models, but here stump selection takes place. Gradually the sample decreases, the probability of a positive class grows, which should eventually lead to a positive balance on the sample outside of training (exam - I call it).

Earlier I have already published a graph with falling probability of choosing an effective quantum segment at a large number of iterations. Let me remind you how it looks on a concrete example for two classes.



The graph shows that as the iterations increase, the probability of choosing the correct (with positive probability bias on new data) quantum cutoff (split by range) decreases.

I decided to look at how this probability manifests itself on a particular predictor - is there stability from iteration to iteration. At each iteration I counted the number of selected quantum splits and the number with positive probability bias on new data, and then determined the percentage of the latter. It turned out that there are cases when quantum segments from the predictor were not selected at the iteration - then the value of the index is zero, but if they were selected, but there are no correct (necessary/effective) ones among them, then the value gets -100.

The zoo turned out to be quite diverse, here are some samples.







You can see that the probability decreases in different ways, both smoothly and by jumps within individual predictors, why it is so different is not clear yet.

It also seems interesting that there are predictors that supply quantum segments that pass selection on the training sample, but on new data are not effective at all - they will interfere well with training - we need additional methods to reduce the probability of hitting them.

It also seems interesting that predictors become a supplier of quantum segments (selected) not necessarily from the first iterations, which leads to the idea that it is important to follow the sequence of using predictors at different depths of the tree.

If you have any thoughts, please write. I am thinking about some aggregate metric for evaluation of the process, that it would be possible to twist the settings of selection of predictors and see their influence on the change of the result.

1...349234933494349534963497349834993500350135023503350435053506...3590
New comment