Forums  > Trading  > Supervised machine learning targets  
Page 1 of 1
Display using:  


Total Posts: 33
Joined: Aug 2014
Posted: 2016-11-06 11:01
I have noticed in several academic papers that they use the next day return for determining the binary target, 0 or 1. It makes sense but I recently read in a blog that this is not realistic because positions are placed at the open of next bar in any time-frame I guess and using the next day return to classify could lead to some problems. Here is the relevant article and that point is made before the Training and Testing section link

Initially I dismissed these claims but after thinking about it some more there may be some valid points.

What do you think about this problem? Has anyone considered this issue before? I did some tests in Orange and found large variations in results.


Total Posts: 443
Joined: Jul 2008
Posted: 2016-11-06 17:36
I don't understand: if you trade very small sizes, you can assume you will get filled at the opening price if you place a market order before the open.

Founding Member

Total Posts: 8356
Joined: Mar 2004
Posted: 2016-11-25 11:57
For sure using daily returns (i.e. Close-to-close returns) as a forecasting exercise is futile. A very significant amount of volatility comes from overnight movement. So saying you can capture Close(t)-Close(t-1) when you make your prediction and trading decision is going to be made after the close, will simply not be meaningful.

The Figs Protocol.
Previous Thread :: Next Thread 
Page 1 of 1