Serving the Quantitative Finance Community

 
User avatar
katastrofa
Posts: 9921
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: MSc Theses on Machine Learning and Computational Finance

December 17th, 2020, 4:39 pm

From the 2020 thesis.

The results in (Horvath et al., 2019) shows very accurate predictions. How- ever, after examining the source code, we discover that the ANN is validated and tested on the same data set. Thus, the high accuracy results does not give any information about how well the trained ANN generalises to unseen data.
In our experiment, we will use different data sets for validation and testing.

Horvath, Blankaetal.(Jan.2019).“DeepLearningVolatility”.In: URL: https: //ssrn.com/abstract=3322085.
Similar stories here with re-implementing published models. Cheating or gross ignorance.

How much the performance changed after training and testing the model correctly?
 
User avatar
Cuchulainn
Topic Author
Posts: 64120
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

Re: MSc Theses on Machine Learning and Computational Finance

December 18th, 2020, 11:11 am

The original article that grabbed the attention (but I've seen even more ridiculous claims) was this gem.

In this article, the Universal Approximation Theorem of Artificial Neural Networks (ANNs) is applied to the SABR stochastic volatility model in order to construct highly efficient representations. Initially, the SABR approximation of Hagan et al. [2002] is considered, then a more accurate integration scheme of McGhee [2011] as well as a two factor finite difference scheme. The resulting ANN calculates 10,000 times faster than the finite difference scheme whilst maintaining a high degree of accuracy. As a result, the ANN dispenses with the need for the commonly used SABR Approximation. 

Depends on which slide rule you use to measure time.
 
User avatar
Cuchulainn
Topic Author
Posts: 64120
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

Re: MSc Theses on Machine Learning and Computational Finance

December 18th, 2020, 11:20 am

My first exposure to Galerkin and Finite Elements for time-dependent PDE research was in the 70s but don't ask me what Deep Galerkin Methods in 100-dimensional space are.

https://arxiv.org/pdf/1708.07469.pdf

In the references, not a single reference to numerical analysis literature. But lots of ML.
 
User avatar
Cuchulainn
Topic Author
Posts: 64120
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

Re: MSc Theses on Machine Learning and Computational Finance

December 18th, 2020, 11:26 am

How much the performance changed after training and testing the model correctly?

<QUOTE> 
We compare the solutions based on ANNs with more traditional computational solutions; based on our level playing field analysis (that is, we compare “apples with apples”), for this problem the performance of the ANN solution is 7 times slower for option pricing and 17 times slower for implied volatility modelling than traditional methods. Of course, this is only one example but it is hard evidence nonetheless.
<\UNQUOTE>

Chun Kiat ONG 2020
 
User avatar
katastrofa
Posts: 9921
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: MSc Theses on Machine Learning and Computational Finance

December 19th, 2020, 1:34 am

NNs for optimising different steps of FEM seem OK and it's been done a lot. I saw the deep Galerkin paper before - AFAIR, they replace the classic linear expansion with a NN from what I understand. IMHO, not a bad idea. They also provide a proof of convergence. The paper seems solid (not that I am conscious enough to read it carefully now), but who knows what's in their codes ;-) As always, though, NNs can fit everything and more - stressing on more. Testing the model against in-sample and out-of-sample data is a lengthy procedure, and any claims that its gazillions times faster than standard approches are nonsensical. I remember you posted those ML-fin papers before and, out of common statistical sense, that's exactly what I thought about their claims.
Last edited by katastrofa on December 19th, 2020, 1:39 am, edited 1 time in total.
 
User avatar
katastrofa
Posts: 9921
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: MSc Theses on Machine Learning and Computational Finance

December 19th, 2020, 1:38 am

Ar you familiar with Lanczos's 1938 tau method?
 
User avatar
Cuchulainn
Topic Author
Posts: 64120
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

Re: MSc Theses on Machine Learning and Computational Finance

December 20th, 2020, 5:21 pm

Ar you familiar with Lanczos's 1938 tau method?
Yes; it is very elegant, if not somewhat antediluvian. Not taking away from CL in any way who was a brilliant mathematical physicist and a lovely man.He was Schroedinger's successor at DIAS in 1952 after being hounded out of USA by Joseph McCarthy. He sometimes gave us lectures as undergrads.

According to Fermi. he was a Hungarian Martian.

https://en.wikipedia.org/wiki/The_Martians_(scientists)

// Using Galerkin is also a bit altmodiisch.
 
User avatar
Cuchulainn
Topic Author
Posts: 64120
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

Re: MSc Theses on Machine Learning and Computational Finance

February 23rd, 2021, 10:23 pm

Talk on RKHS and Python Library by  John-Marc Mercier et moi.


www.youtube.com/watch?v=S0HnCjKq8-s
ABOUT WILMOTT

Wilmott

Wilmott.com has been "Serving the Quantitative Finance Community" since 2001. Continued...


Twitter LinkedIn Instagram

JOBS BOARD

JOBS BOARD

Looking for a quant job, risk, algo trading,...? Browse jobs here...