Forums  > Pricing & Modelling  > Machine learning for pre announcement index rebal  
     
Page 1 of 1
Display using:  

Nonius
Founding Member
Nonius Unbound
Total Posts: 12671
Joined: Mar 2004
 
Posted: 2016-12-04 15:38
Thoughts, memes, links and bags of themes?

Chiral is Tyler Durden

nikol


Total Posts: 422
Joined: Jun 2005
 
Posted: 2017-04-08 15:50
https://cloud.google.com/blog/big-data/2017/01/learn-tensorflow-and-deep-learning-without-a-phd

no need for thanks

jslade


Total Posts: 1089
Joined: Feb 2007
 
Posted: 2017-04-09 12:29
I bet you could get something alpha-like using simple PGMs.

Some nincompoop at two sigma tried to convince me dweeb learning was the future. I happen to know one of the god-fathers of this subject (accident; I'm not interested, mostly because I have good inside information about it); he figured two sigma guy was trolling me.

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

EspressoLover


Total Posts: 237
Joined: Jan 2015
 
Posted: 2017-04-09 23:13
Seems like engineering a good feature set is much more important than model selection here. I'd just start with a random forest. RF's are pretty much the closest thing to a free lunch in machine learning. It might not be the optimal model, but if your features work at all, they'll almost definitely work in an RF right out of the box. If not, back to the drawing board.

ryankennedyio


Total Posts: 12
Joined: Nov 2016
 
Posted: 2017-04-10 08:15
On the face of it, PGM's seems like they were designed to deal with a lot of things that pop up in finance.

This seems like a good MOOC to learn more, which I wouldn't mind doing. https://www.coursera.org/learn/probabilistic-graphical-models

Pretty dubious of 'deep' learning in this area. Maybe something worth looking at for spacial structure in the order book, but probably overkill.

katastrofa


Total Posts: 360
Joined: Jul 2008
 
Posted: 2017-05-01 22:56
Deep learning (no need to mangle the name) makes sense if your model can benefit from being non-linear (in feature space). If yes, try it. If no, stick to linear models.

katastrofa


Total Posts: 360
Joined: Jul 2008
 
Posted: 2017-05-02 22:36
The problem with deep neural networks is that they're not robus to noise (http://www.cs.toronto.edu/~tang/papers/deep_robust_rec.pdf).
Previous Thread :: Next Thread 
Page 1 of 1