Forums  > General  > Is anything interesting happening in quant finance right now?  
     
Page 2 of 2Goto to page: 1, [2] Prev
Display using:  

mtsm


Total Posts: 212
Joined: Dec 2010
 
Posted: 2018-02-06 17:47
Also, what is your take on DLT jslade? Thanks.

Osiris2


Total Posts: 15
Joined: Sep 2017
 
Posted: 2018-02-07 01:28
Here's a thing ...
Lots of what we are calling "fintech" companies, which I would more say are more accurately specialty finance non-bank lenders, are selling their capability to improve on legacy credit underwriting and analysis. As of now, this remains 60% hype, 40% bullshit, and approximately 0% reality.
But that won't remain the case forever. The industry and it's biggest players have reached the scale where they will have to deliver on this promise soon, or else acknowledge they are just certain pieces of a bank with a shiny front-end and give back all the "value" they created by claiming otherwise.
That's a quant problem, and a hard one. Anyone that can improve on the predictive power of current credit scoring, do it at scale, and do it for lower marginal cost (whether in consumer unsec., real estate, or SME) should be able to turn that into real money.

NeroTulip


Total Posts: 1010
Joined: May 2004
 
Posted: 2018-02-07 12:41
It recently occurred to me that not much interesting is happening, or I am out of the loop... A friend is doing a PhD in machine learning after a dozen years in trading, and is looking for a research topic applying deep learning to finance. I was like "meh". Am I missing something?

"Earth: some bacteria and basic life forms, no sign of intelligent life" (Message from a type III civilization probe sent to the solar system circa 2016)

jslade


Total Posts: 1123
Joined: Feb 2007
 
Posted: 2018-02-07 22:32
@deeds I don't have any pointers beyond "get good at Bayesian modeling." A dumb starting point would be something like Steve Skienna's book on modeling horse races.
@mtsm I don't know what DLT means. If it's some new shitcoin, I haven't thought about it at all.

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

mtsm


Total Posts: 212
Joined: Dec 2010
 
Posted: 2018-02-07 23:32
Distributed Ledger Technology. I am being told it's no longer civilized to talk about blockchain, for whatever reason.

I meant to ask if you had a take on start ups pushing this sort of tech in a proprietary way for various applications, first and foremost smart contracts for finance I suppose?

jslade


Total Posts: 1123
Joined: Feb 2007
 
Posted: 2018-02-08 18:30
Whoever told you this is an idiot. Most of the DLTs which exist in the corporeal world are blockchain.

Start ups are pushing blockchain for the same reason they're pushing "AI." Investors invest in things with those terms in the bizplan.

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

katastrofa


Total Posts: 420
Joined: Jul 2008
 
Posted: 2018-02-23 21:59
Deep Learning is not successful in quant finance for several reasons:

- it's not data-efficient
- it's not suitable for online learning (you need multiple passes over a dataset), which is a problem for non-stationary data
- it's not stable: people don't talk too much about it, but you deep neural network training is very unstable, requires a lot of hand-tuning and depends on the random seed
- theoretical understanding is poor, so you're never sure about convergence, generalisation or confidence intervals

Deep Learning so far has worked best when applied to problems which humans are too lazy to do at scale (e.g. image recognition). But in trading, the incentive to analyse data is large ($$$), so anything humans *could* do in order to make a profit, they *will*. Hence, AI has to beat human performance, not just match it or replace accuracy with throughput (e.g. no image recognition network meets human performance, but we don't care because it's way cheaper to use a PC to tag 1,000,000 images than to pay a human to do it).

mtsm


Total Posts: 212
Joined: Dec 2010
 
Posted: 2018-02-24 04:21
Not wanting to play devil's advocate, but to mitigate your statements a bit:

1. not necessarily an issue, DL isn't the only technique that isn't data efficient. Data inefficiency is generally an issue on the buy side, which is mostly backward looking along a single path...

2. Alright, not much to say. It's true it's hard to make it online. It's not really the only reason non-stationarity is an issue.

3. Yes, but nobody in his right mind uses a single net to make predictions. Even computer vision applications tend to compute averaged predictions. If you spend too much time on tensorflow tutorials and such, you walk away with the impression that every learning curve is exponentially decaying and you end up with a latest and greatest trained net. That's kind of naive though...

4. That's a problem, although I would say that convergence, generalisation, confidence is generally a problem when applying statistical methods to empirical data in the absence of knowledge of the data generating process.

katastrofa


Total Posts: 420
Joined: Jul 2008
 
Posted: 2018-02-24 12:34
1. But DL is more data inefficient than others. Data efficiency is an issue in all applications.
2. Of course it's not the only one, I never said it is.
3. The problem is that if you need to train the network 40 times, and each time takes you 1h, then you either have to throw a lot of compute at the problem (which is expensive) or forget about daily recalibration.
4. You never *know* the data generating process, but you can *model* it. But DL setups are often not interpretable statistically, e.g. because they use L2 loss.

nikol


Total Posts: 439
Joined: Jun 2005
 
Posted: 2018-02-26 23:15
a lot.
finance is coming back to senses now.

nikol


Total Posts: 439
Joined: Jun 2005
 
Posted: 2018-02-26 23:16
a lot.
finance is coming back to senses now.

nikol


Total Posts: 439
Joined: Jun 2005
 
Posted: 2018-02-26 23:16
[delete duplicate]

deeds


Total Posts: 378
Joined: Dec 2008
 
Posted: 2018-02-27 00:22
Hi Nikol,

possible to provide any color or examples?


katastrofa


Total Posts: 420
Joined: Jul 2008
 
Posted: 2018-03-02 00:19
This time it's different!
Previous Thread :: Next Thread 
Page 2 of 2Goto to page: 1, [2] Prev