Forums  > Pricing & Modelling  > Forecasting Methodologies  
     
Page 2 of 3Goto to page: 1, [2], 3 Prev Next
Display using:  

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-11 13:23
I like #2. Making zee fuck. Sometimes I'll be mid trade and I run to the restroom to get a few pumps in because I know the trade will go even better if I have the right chemicals coursing through me. As he recommends I keep "sexy pictures" with me at all times. Women, cats, horse genitals, picnic tables. All the things that get me going.

Nonius is Satoshi Nakamoto. 物の哀れ

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-11 14:08
They do publish a bunch. Not enirely clear to me how some of the pubs connect exactly to, say, Stratus performance. Good for them that they seem to be making money with little or no beta.

Nonius is Satoshi Nakamoto. 物の哀れ

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-06-11 14:45
They wrote this paper about how much to bet when you have a nominal limit, fixed transactions costs, and a predictor that is correlated with returns. The upshot in most interesting cases was, uh, bet the full amount if predictor is above cost in absolute value....Head against Wall

Chiral is Tyler Durden

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-11 14:55
So if I go all in when E[] of trade net of tx is > 0 then I make money? Conversely I should spend nothing. Got it.

Nonius is Satoshi Nakamoto. 物の哀れ

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-11 15:41
I am going through radikal's list of references. It's long. Trying to avoid the obvious speech / pattern recognition papers. Visual renderings of fisher matrices can be pleasing in an artistic aesthetic way.

Nonius is Satoshi Nakamoto. 物の哀れ

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-06-11 19:32
"Always compare the complex methods with something really simple, like exponential smoothing or moving average. You may be surprised how good the simple stuff can be, especially if you have a lot of noise."
-Katastropha.

Time and time and time again lots of people come back to this. I think even Ronin or Radikal had said this before. The phrogs are big on this concept as well, believe it or no.

Chiral is Tyler Durden

jslade


Total Posts: 1070
Joined: Feb 2007
 
Posted: 2016-06-12 02:53
There are specific branches of statistics and such which probably merit deeper study for phynance beyond MA and linear regression in furtherance of making money trading. I'm not going to lemonade stand it, but you're basically looking for weak signal in a sea of shitty noise, so .... who else gets paid to do this well?

@Espresso I'd be curious what kinds of real world problems you're referring to. If it's just some math thing (aka "muh NP hard"), my curiosity is lower.

I actually like Sornette's essay on optimizing life, though screw not eating meat. Doing the Sornette diet during periods of high stress and little sleep seemed to help me. Amusingly a lot of it is completely consistent with "physical culture" guys (Bernarr MacFadden & etc) advice from the 1900s

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

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-12 03:10
If I didn't eat basically high fat / protein vegan and zone train my heart five days a week I'd keel over and die with my current work and family demands.

Oh, and sexy pictures.

Nonius is Satoshi Nakamoto. 物の哀れ

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-06-12 03:50
yeah, fuck the vegetarian thing. I keep on coming back to the basic principle that vegatarians are inhumane- they eat things that can't run away.

Chiral is Tyler Durden

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-12 05:36
Meat mainly on weekends. Maybe I am just feeling older too. Digestion is an expensive calculation. Drags down he whole grid. If I had a 35 hour work week it would be less of an issue. Quel dommage.

Nonius is Satoshi Nakamoto. 物の哀れ

Maggette


Total Posts: 943
Joined: Jun 2007
 
Posted: 2016-06-12 11:30
I have to defend the "more fancy" forecasting methodologies here, at least to a certain extend.

At present state I live outside the world of financial time series and earn my money on forecasting and optimizing "physical stuff". There are several scenarios where a kalman filter won't do much.

On Bouchaud: back in the days for my master thesis I wrote him an e-mail since I was writing on a topic he publishes good stuuf (IMHO): Hedging and Pricing in Incomplete Markets.

He was very kind and we had a nice conversation for a couple of mails. Very friendly and humble person. Really appreciated it back then.


Ich kam hierher und sah dich und deine Leute lächeln, und sagte mir: Maggette, scheiss auf den small talk, lass lieber deine Fäuste sprechen...

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-06-12 16:57
here's been my problem on the more complicated stuff. I'd like to be able to use it and maybe it's just I'm being stupid. but it's hard to track all the moving parts in a fast-pace environment and it's also kinda slow to update complicated models, er, in a fast-paced environment. On the NN and other machine learning stuff, I guess that once they are calibrated maybe they can run fast, but aren't NN more for discrete classification problems? I read in another NP post some guy saying that NN can make you do what you do smarter, but I'd like more clarity on what that means. On Kalman seems you can do Kalman super lite and in fact you can just do an even simpler version of Recursive Least Squares, which is Kalman Lite already. MA works pretty nicely not only on averages but also on covariances.

Chiral is Tyler Durden

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-12 17:06
I could write a big list but here's a simple example: for me, at least as it relates to P&L, more complicated schemes become harder to attribute. Further, P&L attribution elements become more abstract and arcane. In my experience there definitely is a relationship between complexity and surprises - at least in equity vol and rates space. Being long gamma helps in many of these instances because they will often involve over-fitting, high turn over, low-latency, higher reactivity. Longer range methods work well over the characteristic time scale you are targeting but can generate path dependent P&L at shorter time scale you are attempting to avoid realizing. Sometimes there's just no benefit. In fact, just this past week we ditched basically a NN for OLS in an application that I could describe roughly but accurately as longstaff-schwartz.

Nonius is Satoshi Nakamoto. 物の哀れ

EspressoLover


Total Posts: 224
Joined: Jan 2015
 
Posted: 2016-06-13 19:32
@jslade - "I'd be curious what kinds of real world problems you're referring to"

Definitely not an esoteric math thing. Deep learning's comparative advantage is when the problem has layers of abstraction. This is why it does so well in image classification. E.g. detecting a certain type of animal in an images requires building up recursive layers of representation. Extract the edges from raw pixels. Then convert edges to curves and counters. From there try to map to simple body parts: eyes, mouths, paws, ears, antenna, etc. Finally if those body parts are arranged the right way, then pick out which animal. (Check out the first link below for an example of this with image classification)

Directly mapping raw pixels to a kitty is not impossible. But there's so many different ways that a cat can be oriented in a picture that a single hidden layer or shallow learner ensemble grows exponentially with the number of pixels. First mapping to a simplified intermediate layer, like edge and contour detection, compacts the parameter space by discarding mostly irrelevant data like noisy variations between neighboring pixels. (Second link is the best non-hype-y paper I could find related to layers of abstraction).

All that being said, I'm still generally skeptical that deep learning has that much to offer quant trading. Just don't think that there's actually that many layers of abstraction to markets. Sometimes you see something like it shoehorned in by some egghead who watched Pi too many times. "At Fund X we use a hyper-advanced system of different trading models depending on a classification of 1 of 67 different market regimes." Almost every single time, those systems tend to be overfitted vaporware. The second major problem is that deep learning is really really susceptible to adversity in the distribution. (Third link). A DNN trading system without some sort of clever mitigating kludge would probably get chopped up by adverse selection.

http://engineering.flipboard.com/assets/convnets/yann_filters.png
http://arxiv.org/pdf/1502.04042.pdf
https://arxiv.org/pdf/1511.07528.pdf

EspressoLover


Total Posts: 224
Joined: Jan 2015
 
Posted: 2016-06-13 20:03
Re: Meat Consumption [sorry il_vitorio for completely chopping up your thread]

The problem with vegetarianism is that it's a lifestyle choice that mostly suits the young. The older we get the more important meat consumption becomes. Basically leucine consumption is really critical to prevent age-related muscle loss. And the number of plant-based proteins high in leucine is basically zip. Maintaining high lean muscle mass is really *really* important to keeping a lid on advanced age mortality.

http://www.nature.com/nrcardio/journal/v8/n4/images/nrcardio.2010.209-f4.jpg
http://digitalcommons.wku.edu/cgi/viewcontent.cgi?article=1296&context=ijesab
http://jn.nutrition.org/content/130/11/2630.full

Nonius
Founding Member
Nonius Unbound
Total Posts: 12666
Joined: Mar 2004
 
Posted: 2016-06-13 20:43
thanks EL. counterintuitive but I'll read with great interest. going to go eat some meat now

Chiral is Tyler Durden

pj


Total Posts: 3318
Joined: Jun 2004
 
Posted: 2016-06-13 20:58
> And the number of plant-based proteins high in leucine is basically zip.
Unless wikipedia is overrun by vegans
Soybeans and hemp are way above beef.
Tongue out

I saw a dead fish on the pavement and thought 'what did you expect? There's no water 'round here stupid, shoulda stayed where it was wet.'

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-13 22:35
Eat nuts and nut products.

Nonius is Satoshi Nakamoto. 物の哀れ

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-13 22:36
BTW, nobody could have forecasted where this thread went.

Nonius is Satoshi Nakamoto. 物の哀れ

jslade


Total Posts: 1070
Joined: Feb 2007
 
Posted: 2016-06-14 00:58
I was afraid you were going to say this, Espresso. What makes me make faces the assertions about DL: when I go back and look at the German Traffic Sign contest that caused everyone to shit their pants that the singularity was around the corner, well, look at the results:

http://benchmark.ini.rub.de/?section=gtsrb&subsection=results&subsubsection=ijcnn

Everyone used HOG descriptors. If you scroll down, SVM, Random Forests and KNN did pretty well too (they're all basically KNN), as did plain old 2 layer NNs using HOG descriptors. Did the DL guys really rule the school, or did Schmidthuber and LeCun's grad students merely work a little harder? If one of Gunnar Carlson's students had thrown topological thingamajiggies at a KNN like approach and used linear regression on topological features to separate the hyperplanes, would it have done better? Kind of hard to say, but I'm pretty sure if it had, nobody would be saying linear regression on Vietoris-Rips filtrations are strong AI.

Not that connectionist approaches are bad. They're just not that all fired impressive. If you gave me a neural whosabobber that did something I couldn't do in a more ordinary way, I'd be more impressed. I haven't seen anything yet that showcases any kind of unique capability. The use cases which exist use absurd computing resources which seem better spent elsewhere. Aka, I'm pretty sure I could identify cat videos with linear regression if you gave me a 10,000 node GPU cluster.

I'm now going to go eat a postmodernist phrench steak, choco and look at boobs or something.

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

TonyC
Nuclear Energy Trader

Total Posts: 1245
Joined: May 2004
 
Posted: 2016-06-14 01:09
> Eat nuts and nut products

I believe that Serano ham is an acorn derivative food product.

flaneur/boulevardier/remittance man/energy trader

chiral3
Founding Member

Total Posts: 4986
Joined: Mar 2004
 
Posted: 2016-06-14 03:43
Good point Tony. All delicious cured meats that are acorn derived should be available to vegetarians.

Nonius is Satoshi Nakamoto. 物の哀れ

goldorak


Total Posts: 979
Joined: Nov 2004
 
Posted: 2016-06-14 08:55
Pushing this reasoning even further we are all feeding on nuclear reactions.

#JeSuisGodzilla


If you are not living on the edge you are taking up too much space.

deeds


Total Posts: 346
Joined: Dec 2008
 
Posted: 2016-06-14 19:04
thanks, jslade

big fan of compressed sensing will look at CX, CUR

agree, seems most analyses are best managed with reference to the data first, the method second...

svisstack


Total Posts: 300
Joined: Feb 2014
 
Posted: 2016-06-15 21:46


he meant, we need more hidden layers

Time well wasted.
Previous Thread :: Next Thread 
Page 2 of 3Goto to page: 1, [2], 3 Prev Next