Machine learning in trading: theory, models, practice and algo-trading - page 3543

[mytarmailS]

Also, the synthetic price can be hypothetically taken as a "fair" price, and catch any divergences or deviations of the real price from the synthetic price....

You can assume that the synthetic price from the model is the real price, and the deviation of the real price from the synthetic price indicates some kind of intervention in the market and analyse these interventions....

That kind of thing. But you can fantasise.

[Maxim Kuznetsov]

Maxim Kuznetsov #:

With matrices apparently it is not quite clear, here people mostly think in algorithms.

mytarmailS #:
God bless the one who understood what you wrote

Maxim Dmitrievsky #:
It's a bunch of random words.

I flared up, I was wrong...people here don't think at all.

[[Deleted]]

mytarmailS #:
Also, the synthetic price can be hypothetically taken as a "fair" price, and catch any divergences or deviations of the real price from the synthetic price....

You can assume that the synthetic price from the model is the real price, and the deviation of the real price from the synthetic price indicates some kind of intervention in the market and analyse these interventions....

That kind of thing. But you can fantasise.

Well, it's just a regression of signs on prices, and the difference between the initial series and its values are residuals. There could be some inefficiencies in the residuals that could be signals.

The easiest test is to predict the next bar, not the current price :) and there you have everything broken, this will be the true efficiency of such a construction. That is, you will see the average error of exactly the prediction for such a model.

But you can filter these predictions in both cases, look for inefficiencies, rather than using all predictions.

You can, for example, do clustering and see in which cluster the error is smaller. Try to write a ts on these pieces.

[[Deleted]]

Maxim Kuznetsov #:

flared up, was wrong...people here don't think at all.

It's just that people are familiar with regression and how it's calculated :) not some system with memory, but regression coefficients are selected to minimise the error between price and fly.

Superfluous words just, highly complicated description.

And MA is also a filter.

[Aleksey Vyazmikin]

Maxim Kuznetsov #:

flared up, was wrong...people here don't think at all.

If you don't know what he input, it's hard to judge. He could have input only the last value or a vector of, say, even bar values or a convolution.

It is not interesting to discuss such topics without specifics.

[[Deleted]]

I came up with this sequence.

1. predict 1 bar ahead, see the error on the new data. So for all sets of signs.

2. choose the model with the smaller error.

3. Do clustering/filtering of predictions, pick the cluster with the lowest error where the model is least wrong.

4. Do sub-clustering on the clusters already found, if you didn't find anything in step 3.

5. You come up with trading rules for the best cluster. At the same time, you can shift the model back 1 bar, if it is easier for you to get signals.

Since the model is linear and therefore fast, you can loop through fixtures and their combinations or optimisation.

This is roughly the approach I use for classification. It is logical and nechural and understandable.

You end up sifting through a sieve of all the randomness and leaving the signals. There may be variations, this is just a general outline.

[sibirqk]

Maxim Kuznetsov #:

flared up, was wrong...people here don't think at all.

In general you are right of course, but in my opinion - only in general. Here we are talking about a specific bundle with RSI. RSI is calculated as the ratio of upward and downward price velocity changes. The velocity is differentiation, respectively the recovery with AMO in this particular case is integration. It is accurate to a constant, i.e. there is a loss of some information. The chart shows that smoothing occurs, mytarmailS calls it filtering, i.e. some information about volatility is lost. In essence, a kind of filter is created. In any smoothing (filtering) there is a lag of half a period of the filter, and in this case, at least by eye, there is no lag. I still think it's an illusion - I don't believe in miracles. But there is some possibility that there is something in it and the person saw his alpha in the distance.

[Aleksey Nikolayev]

Maxim Kuznetsov #:

flared up, was wrong...people here don't think at all.

There are many ways to think. Imho, the best way is to think in the direction of what and how to calculate. For example, counting the "timetable" for a price reversal and not seeing it. Drawings of borats and other things will not help if they do not lead to calculations.

[mytarmailS]

Aleksey Vyazmikin #:

If we don't know what he input, it is difficult to reason. It could input only the last value or a vector of, say, even bar values or a convolution.

It is not interesting to discuss such topics without specifics.

Everything that he submitted and how he described, you just need to read not from the last page

[mytarmailS]

Maxim Kuznetsov #:

flared up, was wrong...people here don't think at all.

It's up to you.

If you show me where the matrix is in the linear regression, we'll talk further.