Forums  > Books & Papers  > how to use "Machine Learning" to build trading system?  
     
Page 2 of 2Goto to page: 1, [2] Prev
Display using:  

EspressoLover


Total Posts: 320
Joined: Jan 2015
 
Posted: 2016-03-31 04:44
Well I have very little background in options, so take what I say with a grain of salt. But here's how I would approach that problem. In trading, your actual objective is to maximize PnL, but that's a noisy, discontinuous, largely intractable function to optimize. So instead we use MSE because it's much better behaved and in most cases closely proxies the thing we care about.

But you want to think about how you might modify the objective function to more closely align with what you care about. I think the easiest solution for that type of problem is to use weighted MSE. If there's some method or algorithm that uses MSE, it's usually trivial to adapt it for weighted MSE. You want to put a higher weight on the points representing options that you care about. There's a number of metrics that you might use to scale weight: volume, log-volume, OI, near-moneyness, inverse of spread as a percent of price, etc.

You'd probably want to play around until you find something that looks intuitively correct. But the jist is that far-OTM illiquid, .04delta options are going to have much less of an influence on parameter training and model scoring than the important contracts.

Good questions outrank easy answers. -Paul Samuelson

levkly


Total Posts: 28
Joined: Nov 2014
 
Posted: 2016-04-06 09:59
Goldorak,

I use ensemble of many fix length slidings windows.
Do you have other dynamic approach?

Lebowski


Total Posts: 66
Joined: Jun 2015
 
Posted: 2016-04-06 16:01
@EL, thanks for sharing your ideas above. I've never really thought of MSE as a "proxy," but that explanation makes a lot of sense.

Azx


Total Posts: 35
Joined: Sep 2009
 
Posted: 2018-03-28 11:01
Marcos Lopez de Prado has written a new book on the subject: Advances in Financial Machine Learning.

jslade


Total Posts: 1132
Joined: Feb 2007
 
Posted: 2018-04-19 21:32
His book is pretty terrible; it has a chapter on quantum computing for crying out loud.

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

Maggette


Total Posts: 1047
Joined: Jun 2007
 
Posted: 2018-04-19 22:07
My Business Partner ist fiddeling around with Quantum Computing as a hobby....he even went to conferences on QC in Maschine learning (at present there is no real application). He is tanken seriously by the IBM guys and researchers in the field. If I get his opinion right:

Pretending that in current state of QC you could do anything useful with it is simply fraud

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

Nexis


Total Posts: 2
Joined: Feb 2018
 
Posted: 2018-04-20 11:28
And as far as I am aware Lopez de Prado has never made the claim that anything useful can be done with it at the moment.

@jslade Could you elaborate a bit? Would be very interested to hear what you think about the book.

Maggette


Total Posts: 1047
Joined: Jun 2007
 
Posted: 2018-04-20 11:35
+1 for a book review by jslade :).

My copy will probably arrive today. Hopefully I won't regret it too much.

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

Martingale
NP House Mouse

Total Posts: 2635
Joined: Jun 2004
 
Posted: 2018-04-20 14:41
Any books/papers to recommend on this subject?

jslade


Total Posts: 1132
Joined: Feb 2007
 
Posted: 2018-04-24 23:37
Part 4 of Lopez de Prado's book is reasonably useful if you don't know these things already. The rest of it? Mostly so bad it's not even wrong.

I have a friend who publishes in this space; he's my friend, and probably trying to get tenure or something so I won't tell you who he is, but he's fucking wrong too. His ML is weak-ass (srsly bad), and his "building of trading strats" statistics is pretty questionable also. Nobody publishes useful information in this field. Why would they?

Real talk: you want to trade using machine learning, you need to study basic stats, signal processing and linear regression to the point you can make money without machine learning. Then you can go read Cesa-Bianchi and Lugosi or Vovk, Gammerman and Shafer if you want to be all fancy. Or hire a ML domain expert to get side information you fold into your trading strat using linear regression and t-stats.

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

goldorak


Total Posts: 1042
Joined: Nov 2004
 
Posted: 2018-04-25 16:39
Big viva for the reference to Shafer !

> Nobody publishes useful information in this field. Why would they?

I was actually thinking about this the other day and came to an evidence: it is not even publishable (as a book and even more as a paper as it would not satisfy certain academics egos). Being able to put together years of experience, heuristics, technicalities, tricks, assumptions mixed personal goals and risk tolerances, etc... in a consistent and readable book seems to me like the ultimate job, a lot harder than just making (or losing) money with all that. Even the bibliography would look horrible. I am even certain I would not be able to explain things done years ago and looking at them I would do it completely differently now. I guess writing a book would become an unending task.

Finally, I would say that what is important are the mistakes done. Calling them mistakes is even a misnomer.




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

deeds


Total Posts: 393
Joined: Dec 2008
 
Posted: 2018-05-25 16:01
Saw LDP give his talk on 7 reasons Quant ML funds fail at BBQ (Bloomberg Quant Seminar) in NY last night. I've heard presentation and paper are also on SSRN.

Topics range more widely than you may expect. None are breaking news.

His scope and delivery were excellent for the audience, in my opinion, recommend trying to catch him in person just for that aspect (regardless of views on content).

Would like to hear thoughts on an interesting provocation from him -

One of the 7 reasons is 'integer differentiation'...the use of returns to get to stationarity. Interesting subsequent comments about the tradoff between memory (in a Hurst sense, sum of deviations) and stationarity.

- recommends fractional differentiation to preserve memory while producing stationary series
- his empirical studies show all financial series his team looked at (dozens) can be expressed this way in a range of exponents of 0.1 to 0.5 or so, without exception
- suggested that perhaps efficiency of markets and unpredictability was a consequence of looking at returns (int diff) rather than frac diff'd price series (*)

Fractional differentiation is an old chestnut from the heady days of Santa Cruz complexity team, 80's maybe, mandelbrot before that...but i'd never heard (*)

Thoughts on whether suggestion is tenable?

[My suspicion is estimation in fractional world is pretty ill-conditioned, not very robust, but i know nothing]


jslade


Total Posts: 1132
Joined: Feb 2007
 
Posted: 2018-05-25 21:46
I fooled around with fracdiff/ARFIMA for a few days when I was fooling around with other wildcard stuff like Sornette's model and Hurst exponents. I think it's a stupid idea, and having skimmed LDP's book, I'm not surprised at his suggesting it. It's trying to capture "long term memory" in a particular way that doesn't really reflect any realistic generating process, but which would probably look good in backtests on some kind of semi-Markov process. FWIIW I think Lo & McKinlay's book has some chapter on this also.

Machine learning guys fail at trading because they have never been signal processing guys. The ML guys who have been signal processing guys seem to do pretty well, which is why RenTech hired Bob Mercer.

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

Azx


Total Posts: 35
Joined: Sep 2009
 
Posted: 2018-05-25 22:54
Well, a fractional differentiation of the price is just a weighted moving average of past returns, with the decay of the weights being controlled by the order of differentiation. Moving averages of returns isn't exactly new, that's what people have been doing with MA crossovers for ages. No need to introduce fractional calculus into the subject.

The correlation between the price and it's fractionally differentiated price should be obvious, the price is also a MA of returns so naturally they will correlate. But perhaps I am missing the point.

I think the main reason "Quant ML funds" fail is that they take simple common methods and concepts and turn them into overly complicated models that makes inference impossible.


deeds


Total Posts: 393
Joined: Dec 2008
 
Posted: 2018-05-29 13:04
Azx, jslade, thanks for the kind responses
Previous Thread :: Next Thread 
Page 2 of 2Goto to page: 1, [2] Prev