Discussing the article: "Developing a robot in Python and MQL5 (Part 1): Data preprocessing" - page 3

New comment
 
Maxim Dmitrievsky #:

Then the price must be pseudo-stationary. This is not observed in trending markets.

Okay, let's forecast the increments. Then let's take the increments of increments. Are they pseudo-stationary?

However, if we have learnt to forecast increments, does it mean that we have learnt to forecast their derivative - price?

 
fxsaber #:

I'm a zero in MO, so I'm relying on the article.

If I understand it correctly, an automaton is a wider field of human choices. If a human can choose a cumulative amount, then an automaton even more so.

The point is that any features fed into the model must be pseudo-stationary, otherwise the model will not work correctly on new data. This is a limitation of the classifier.

The data must be represented in a rank of values that persists on the new data. If this condition is not fulfilled, the model will stick in the boundary values when going beyond this range.

For example, predict always 0 or always 1.
 
fxsaber #:

Okay, let's predict the increments. Then let's take the increments of increments. Are they pseudo-stationary?

However, if we have learnt to predict increments, does it mean that we have learnt to predict their derivative - price?

Yes

 
Maxim Dmitrievsky #:

the data must be represented in a range of values, which is saved on new data. If this condition is not fulfilled, the model will stick in the boundary values when going beyond this range.

I.e. these signs should be obviously thrown out?

    # Add simple features
    raw_data['raw_SMA_10'] = raw_data['close'].rolling(window=10).mean()
    raw_data['raw_SMA_20'] = raw_data['close'].rolling(window=20).mean()

Because they can easily reach values outside the training interval.

 
fxsaber #:

So these signs are supposed to be discarded?

Because they can easily reach values outside the training interval.

Yes, they can't be used, just like the opening prices.

Well, you can't. Technically, you can, but it is fraught with the risk of model stalling on new data.

If you have some vision that they will not go beyond their range in the near future, then you can.

 
Classification learning realises that the features price, returns^1, returns^2, ... are one and the same - there is an unambiguous relationship between them, and therefore only one attribute should be left from this list?
 
Maxim Dmitrievsky #:

If there is some vision that they won't go out of their range in the near future, you can.

Gold and crypto are not suitable for this.

 
fxsaber #:
Classification learning realises that the features price, returns^1, returns^2, ... are one and the same - there is an unambiguous relationship between them, and therefore only one attribute should be left from this list?

It is not the same, there is still a difference. The solution will depend on the total number of attributes. If it is too large, then it can be cleaned up. If not, you can leave it without problems. Moreover, they are slightly different (increments with different lags) and can provide additional useful information to the model.

Price is deleted for the reason mentioned above.
 
fxsaber #:

Gold and crypto don't qualify for this.

not much fits, so all signs are at least detrending.

 

It seems that MO is positioned as a being (mega-human) with super-computing abilities and memory.

What signs does an ordinary person use to create a TC? I have never seen a person look at a graph of increments and try to predict it.

1234567
New comment