Forums  > Trading  > Symbolic string representation of time series (pattern recognition)  
     
Page 1 of 1
Display using:  

Igor


Total Posts: 5
Joined: Nov 2014
 
Posted: 2016-11-11 05:18
I am exploring a few ideas relating to symbolic representation of time series for pattern recognition.

Does anyone have any references (textbook or papers), perhaps from the bioinformatics community, on applying techniques like string based symbolic representation to time series for classification and indexing?

Cheers

"..tempus casumque in omnibus"

jslade


Total Posts: 1064
Joined: Feb 2007
 
Posted: 2016-11-12 06:48
SAX is pretty good:
http://www.cs.ucr.edu/~eamonn/iSAX/iSAX.html

I've used other, hand rolled discretization schemes as well. Throwing data away is key.

"Learning, n. The kind of ignorance distinguishing the studious."

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-11-12 11:52
Isn't this the Bruno Dupire type stuff? ie, digitizing prices/returns into streams of letters?

Chiral is Tyler Durden

jslade


Total Posts: 1064
Joined: Feb 2007
 
Posted: 2016-11-12 21:58
Yes Nonius; that is an example of TS discretization (an ad hoc one). The sequence prediction and TS classification fields all use discretization though. Think of it as a sort of information theory inspired filter. If you can convert your TS to a symbol string, there's all kinds of information theory inspired tricks you can do with them. Simply storing the data in this form has serious benefits when your data gets big.

"Learning, n. The kind of ignorance distinguishing the studious."

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-11-13 08:56
interesting jslade. Taking a deeper look.

Chiral is Tyler Durden

JTDerp


Total Posts: 38
Joined: Nov 2013
 
Posted: 2016-11-13 17:46
jslade, your mention of 'throwing away data' as well as the SAX paper's mention of random sampling - is this somewhat similar to 'bootstrapping' returns of a time series? Is there a 'smoothing' effect as result of discarding and/or random sampling?

The clouded mind seeks; the emptied mind finds.

jslade


Total Posts: 1064
Joined: Feb 2007
 
Posted: 2016-11-14 01:37
More like, "you don't need that many bits to capture the signal in a noisy time series." Prediction and pattern recognition is compression. That sort of thing.

"Learning, n. The kind of ignorance distinguishing the studious."
Previous Thread :: Next Thread 
Page 1 of 1