Hello,
I think there is a missing video in section 16 of "Feature engineering" course. In the first video, it ends saying "we will see how to keep price levels information while keeping input stationary", but there is no next video on that section. At the end of the section, we do an ADF statistic test and then get over a section called "implementing your own datasets", but there it explains how to install anaconda.
Where I can find the rest of the material?
Hi Mario,
This topic "how to keep price levels information while keeping input stationary" is covered in the Fractional differentiation notebook (Section 16 Unit 5).
We will update the video as soon as possible, which says let's see how we can include the price level information along with keeping the input stationary in the next video.
Hi and thank you for your answer. So, as far as I understood in the notebook, the best way to keep prices levels information in the notebook example would be to use the log(close) prices, right? And I guess that I should apply that log to the full OHLC. After that, I need to recalculate all the features using those new OHLC? I'm asking because I got confused with the tick bars, volume bars and all the other bars.
In the "feature extraction" section (13) it was not very clear how to extract features (usefull ones). At least not for me :(.
Hi Mario,
Machine learning algorithms work best with stationary features.
How can you make OHLC prices stationary?
One of the ways of making OHLC stationary is to take percentage change (diff = logMSFT.pct_change()). But this approach results in loss of information. That is the loss of price levels.
But is there a better method?
Yes, you can avoid complete loss of information of price levels bu using fractional differentiation. The more on it is covered in the notebook.
As seen in the below graph, the red line is of log prices. Log prices are generally not stationary and hence not useful. The blue line is for percentage change. It is stationary but has no information on price levels. So we find a middle ground, a stationary price series with some price levels. This is the green line in the centre or fractional differentiated series.
For the second part on bars, there are many ways in which you can sample data such as on fixed time, fixed volume and fixed dollar volume. Based on the pros and cons, you can finalize how you want to sample the data.
After that, you need to take fractional differentiation of that series and pass it to your ML algos.
I hope this helps.
Thanks