Serving the Quantitative Finance Community

 
User avatar
bbtac
Topic Author
Posts: 3
Joined: April 11th, 2010, 9:20 pm

Machine Learning topics in Model Validation?

May 21st, 2018, 11:42 am

Hello,

I am working in a Model Validation team and an intern will join shortly in the team for 6 months. 
His background is mainly Machine Learning, he did not do much of mathematical finance.
I am looking for Machine Learning topics on which he could worked on. The topics need to be related to Model Validation (at least vaguely).
From what I have seen so far, there are at least two types of topics he could work on:
  • Missing data topics: for example  predicting the ratings of bonds (the application is for bonds for which we don't have ratings)
  • Fraud detection topics: for example detecting crazy counterparty PV or crazy Markit prices etc ...
I would like to find some more sexy topics.
Any Ideas?

Regards
 
User avatar
staassis
Posts: 27
Joined: April 12th, 2014, 5:10 pm

Re: Machine Learning topics in Model Validation?

August 4th, 2018, 2:41 pm

This is fairly general but applies to model validation in particular: apply gradient boosting to model development. That is, use the current model as the starting point. Augment it with various other modeling ideas for those data points (instruments, assets) for which the current modeling iteration does not perform well. Do everything under the guidance of gradient boosting principles, as introduced by Friedman.

The philosophy of this isn't new. In my experience at an investment bank, some of the best performing models were not those based on one underlying distribution but those composed of a series of corrections and premiums.
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning topics in Model Validation?

August 5th, 2018, 10:45 am

His background is mainly Machine Learning, he did not do much of mathematical finance.

How long does this intern project last? Is is a MSc/MFE student?

My take would be to write up an unambiguous specification of an existing problem that you have solved before and then solve/extend it using ML.
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning topics in Model Validation?

August 5th, 2018, 11:40 am

This is fairly general but applies to model validation in particular: apply gradient boosting to model development. That is, use the current model as the starting point. Augment it with various other modeling ideas for those data points (instruments, assets) for which the current modeling iteration does not perform well. Do everything under the guidance of gradient boosting principles, as introduced by Friedman.

The philosophy of this isn't new. In my experience at an investment bank, some of the best performing models were not those based on one underlying distribution but those composed of a series of corrections and premiums.
If you had 60 seconds to explain what gradient boosting is, how would you pitch it? (I'm not lazy, it's just there are too many (unreferred?) articles to read.)

Stupid question: why use gradients?

Thank you.
 
User avatar
staassis
Posts: 27
Joined: April 12th, 2014, 5:10 pm

Re: Machine Learning topics in Model Validation?

August 5th, 2018, 5:57 pm

@Cuchulainn

I understand your constraints very well. I am in the same boat, in a way... "Gradient" is just a word in the name of the method (related to original inspirations). The whole method is much broader... More importantly, in the next 10 years the method will become even broader, through automated frameworks like the one I suggested. To the best of my knowledge, my suggestion has not been addressed in industry or academia... One of the best references on boosting is chapter 10 (Boosting and Additive Trees) in the following book:   

https://web.stanford.edu/~hastie/Papers/ESLII.pdf