/sci/ - Science & Math » Thread #12289429

943KiB, 1200x675, file.png

View Same Google iqdb SauceNAO

Mathematical finance thread

Anonymous Fri 30 Oct 13:32:53 2020 No.12289429 View Reply Original Report

Quoted By: >>12289449 >>12289526 >>12289544 >>12289742 >>12290201 >>12290415 >>12290419 >>12292531 >>12292948 >>12292951 >>12295517 >>12296002 >>12299338

How do I use non-linear transforms to beat the market?

Anonymous

Anonymous Fri 30 Oct 2020 13:44:13 No.12289449 Report

Quoted By:

>>12289429
Thread your fiber-optic cable through a wormhole so it gets to the exchange quicker than a straight line does.

Anonymous

Anonymous Fri 30 Oct 2020 13:51:15 No.12289465 Report

Quoted By:

something something autocorrelation

Anonymous

View Same Google iqdb SauceNAO 1604058971033.png, 289KiB, 425x445

Anonymous Fri 30 Oct 2020 14:16:49 No.12289526 Report

Quoted By:

>>12289429
gradient descent

Anonymous

Anonymous Fri 30 Oct 2020 14:23:24 No.12289544 Report

Quoted By:

>>12289429
This is a pure speed and bruteforce problem. Attach non-synchronous computers to a state of the art internet connection, coupled with a decent buy/sell algorithm and you are done. BTW you need to spend 1M dollars a month too.

Anonymous

View Same Google iqdb SauceNAO 1590276788415.png, 86KiB, 1745x612

Anonymous Fri 30 Oct 2020 15:45:26 No.12289742 Report

Quoted By:

>>12289429
only market timing is true

Anonymous

Anonymous Fri 30 Oct 2020 18:12:09 No.12290201 Report

Quoted By:

>>12289429
this is the dumb thinking and honestly not worth the effort. Better setup an exchange platform and front run your traders on minor spreads.
so if stock is going for $2 execute the trade for $2.02.
but thats only 2cents you say...multiply this by millions of trades per day every day.

Anonymous

Anonymous Fri 30 Oct 2020 19:19:17 No.12290415 Report

Quoted By: >>12290882

>>12289429
Things that work(from my experience at a fairly known algo hedge fund that is well past its glory days)
- advanced cointegration
-fractal approaches for volatility products
-phase transition approaches
- Long Vol/ vol convexity(but only if you can ensure that you have a very good OTC derivatives trader who keeps the costs very low and can extend long vol strategies to more exotic derivatives)
Things that don’t work(or rather that will end up costing you your job):
-Any sort of arbitrage that requires leverage to work
- Short vol strategies with unlimited downside
- Trend following

Anonymous

Anonymous Fri 30 Oct 2020 19:22:30 No.12290419 Report

Quoted By:

>>12289429
Optimizing Sharpe Ratios.

Anonymous

Anonymous Fri 30 Oct 2020 22:50:36 No.12290882 Report

Quoted By: >>12291506

>>12290415
In your experience what do you think Renaissance Technologies does? Can't just be advanced algorithms and math right? I know they cap the fund's size which helps but 80% every year? Cmon...

Anonymous

Anonymous Sat 31 Oct 2020 03:39:54 No.12291506 Report

Quoted By: >>12291519 >>12291527 >>12292403

>>12290882
From what I was told by ex-Rentech i used to work with they mastered the art of proprietary data pipelines
When everybody else was running around dicking with fads Rentech amassed an enormous amount of proprietary market data nobody else wanted and then capitalized on that information. It takes like 10 years to put together the kind of data analysis Rentech has been doing imagine 100+ proprietary data feeds, each with terabytes of noisy data being dumped every ms and being able to make sense of that in real time in order to 'beat the market' that's what they do and anybody else can do too if they invested in the data pipeline areas and not blowing their wad on silly strategies and other voodoo.

Ito calculus and the like is all on MIT open courseware if anybody is interested

Anonymous

Anonymous Sat 31 Oct 2020 03:48:02 No.12291519 Report

Quoted By:

>>12291506
by 'ex rentech' i mean interns that only worked there 2 months.

Anonymous

Anonymous Sat 31 Oct 2020 03:52:30 No.12291527 Report

Quoted By:

>>12291506
What sort of data do these firms work with and what is meant by data pipeline exactly?

Anonymous

Anonymous Sat 31 Oct 2020 13:26:47 No.12292403 Report

Quoted By:

>>12291506
More or less accurate. Not a well-known fact was that Simons hired a bunch of ex-NSA statisticians, people who are very good at sorting huge datasets. Their strength absolutely lies in the ability to collect and make sense of data better than everyone else.

Anonymous

Anonymous Sat 31 Oct 2020 14:52:16 No.12292531 Report

Quoted By: >>12292577

>>12289429
Use Fourier transforms to analyze your data outside the time domain. Maybe you can spot patterns; or even better, use Fourier transforms on volatility data to check how it behaves outside the time domain.

Anonymous

Anonymous Sat 31 Oct 2020 14:54:30 No.12292537 Report

Quoted By:

You don’t and go poor instead.

Anonymous

Anonymous Sat 31 Oct 2020 15:12:27 No.12292577 Report

Quoted By: >>12292797 >>12292825 >>12293419 >>12293438 >>12293567

>>12292531
Don't. If you use a discrete FT on a finite sample you just end up with a frequency decomposition of THAT sample, with no predictive power at all. Most you can do is find power spectra but they only work for stationary processes and nothing in finance is stationary in any sense unless you assume a lot. This has been done but anyone trading on that is a fool.

If you want to make money you have to get huge undisclosed data and feed machine learning algorithms that. This is why ML is THE industry standard right now in all hedge funds. If you come up to HF with time series, Ito calculus or the likes they will laugh at you and send you back to where you came from.

If you want to work in funds, become an expert in cutting edge machine learning. I talk about CS PhD level.

Anonymous

Anonymous Sat 31 Oct 2020 16:56:28 No.12292797 Report

Quoted By: >>12292802 >>12292989 >>12293004

>>12292577
which hedge funds have cutting edge ml? they're all shit except rentech or tgs right

Anonymous

Anonymous Sat 31 Oct 2020 16:57:57 No.12292802 Report

Quoted By: >>12292811

>>12292797
>I’m smarter than Ivy League mathematicians

Anonymous

Anonymous Sat 31 Oct 2020 17:02:23 No.12292811 Report

Quoted By: >>12292989

>>12292802
No its a genuine question, I'm just curious. What firm other than those two are using ML to produce outsized returns?

Anonymous

Anonymous Sat 31 Oct 2020 17:08:22 No.12292825 Report

Quoted By: >>12292878

>>12292577
>This is why ML is THE industry standard right now in all hedge funds
LMAO you are literally just making that shit up. Satellite / fiber optics, FPGAs and pretty much anything low latency related are the main concerns of any decent HFT firm. ML is a buzzword that never actually caught up in the world of automated trading because everyone knows this shit is 100% unpredictable, it's all about being 1 ns faster than the others.

Anonymous

Anonymous Sat 31 Oct 2020 17:29:45 No.12292878 Report

Quoted By: >>12292929

>>12292825
Pretty sure he's talking about mid frequency quant firms. There are firms that don't rely on speed but rather high level statistical arbitrage, amplified by extremely high leverage.

Anonymous

Anonymous Sat 31 Oct 2020 17:44:48 No.12292929 Report

Quoted By: >>12293004

>>12292878
Are those firms even relevant or competitive at this point? Unless they lack a really good "public relations" division you don't really need @ bunch of traders or expensive computing power to run a company like that in 2020.

Anonymous

Anonymous Sat 31 Oct 2020 17:51:04 No.12292948 Report

Quoted By:

>>12289429
You can get good analytical system, depends on how much do you pay.

Usually for decent information people pay.

Anonymous

Anonymous Sat 31 Oct 2020 17:52:06 No.12292951 Report

Quoted By:

>>12289429
Feed them into sum of sum.

Anonymous

Anonymous Sat 31 Oct 2020 18:02:56 No.12292989 Report

Quoted By: >>12292992 >>12293004 >>12293046

>>12292811
>>12292797
All 'no-name' funds (the vast majority) now base all their decisions on neural networks. I have had contacts with quants that told me that all HF industry bought HARD into the meme and now all quants basically just do that and investors want the ML badge on their money.

In any case data is key to any successful trade or investment, at any term. If you base your information on data that other people have, you're busted. Data mining is the next step.

Anonymous

Anonymous Sat 31 Oct 2020 18:04:09 No.12292992 Report

Quoted By:

>>12292989
But neural network is not correct key for that. There's a bit magic in it to do precise.

Anonymous

Anonymous Sat 31 Oct 2020 18:05:55 No.12293004 Report

Quoted By:

>>12292929
Only a handful and they're all private, so no PR. Read "13 billion dollar mystery angels" on Bloomberg for example. >>12292797
>>12292989
Yeah, precisely what I was getting at. Rentech et. al are a league above because they're truly quantitative. Full end to end automation of data extraction, processing and executing trades. Very different to what firms like Two Sigma or Citadel is doing.

Anonymous

Anonymous Sat 31 Oct 2020 18:15:36 No.12293046 Report

Quoted By: >>12293349

>>12292989
hell, I've seen even PE funds using rudimentary machine learning to mine their data. These people care about two things:

1) Fads. ML has been a big fad for over a decade and still hasn't been dethroned. Investors love the latest fad, so fund managers and quants buy into the memes to get money.

2) Results. Traditional mathematical economics and finance are hard wired into looking for consistent theories, not predictions. The industry cares about prediction and machine learning does that. It provides black box models that churn out results. Basically ML destroyed the traditional time series approaches (see cointegration), which now live only in academia and central banks

Anonymous

Anonymous Sat 31 Oct 2020 19:56:31 No.12293349 Report

Quoted By: >>12293422

>>12293046
ML subsumes time series, the standard now is retarded levels of ensembled models.

What is actually cool about hedge funds is that the data scientists are actually real ass dudes with pedigree, unlike in just as prestigious consultancies they are business majors who took a couple intro stat and cs classes in undergrad

Anonymous

Anonymous Sat 31 Oct 2020 20:23:34 No.12293419 Report

Quoted By: >>12293424

>>12292577

> He chose to resort to lies instead

Anonymous

Anonymous Sat 31 Oct 2020 20:24:13 No.12293422 Report

Quoted By:

>>12293349
yeah consultancies are all fake ass shit. They steal money from clients. I've seen what kind of stupid shit goes on in their 'models'. This is because they do not care about theoretical consistency nor results, they just want the fee.

Anonymous

Anonymous Sat 31 Oct 2020 20:25:17 No.12293424 Report

Quoted By: >>12297962

>>12293419
where's the lie anon

Anonymous

Anonymous Sat 31 Oct 2020 20:34:56 No.12293438 Report

Quoted By: >>12293474 >>12293572

>>12292577
Doesn't training neural networks on a finite set of data also gives prediction power only to that sample?

Anonymous

Anonymous Sat 31 Oct 2020 20:45:31 No.12293474 Report

Quoted By:

>>12293438
Hopefully not.

Using fourier series/transforms however gives information only about a specific sample (if you extend the reconstruction in an attempt to 'predict values' you just end up with a periodic function so it's stupid). If you have a stationary process you can use fourier transforms to get the power spectrum and then yes you can predict shit. But no real world data (especially economics) is stationary, so it's useless.

Anonymous

Anonymous Sat 31 Oct 2020 21:18:02 No.12293566 Report

Quoted By:

Consider classical linear regression. HFT firms do not do linear regressions but that is good for the example. Your accuracy metric is the R squared. In HFT, an average R squared is 0.15. A R squared of 0.25 or 0.3 is considered very good.

Anonymous

Anonymous Sat 31 Oct 2020 21:19:13 No.12293567 Report

Quoted By: >>12293614

>>12292577

No ones use advanced ML for maintenance reason and because it is just to slow to be useful for prop shops or funds.

Anonymous

Anonymous Sat 31 Oct 2020 21:22:13 No.12293572 Report

Quoted By:

>>12293438
I mean the idea is that you find the sweet spot between under fitting and over fitting that generalizes well.

Anonymous

Anonymous Sat 31 Oct 2020 21:36:36 No.12293614 Report

Quoted By: >>12293718 >>12293764

>>12293567
Why does everyone assume that hedge funds or prop shops are high frequency?

Anonymous

Anonymous Sat 31 Oct 2020 22:11:46 No.12293718 Report

Quoted By: >>12293731

>>12293614

Mid/low freq = asset management

In low freq it is really hard to "beat the market", 80% of firms that do active management fail.

Anonymous

Anonymous Sat 31 Oct 2020 22:16:58 No.12293731 Report

Quoted By:

>>12293718
Yeah but there are giants that do well like Black Rock or Vanguard.
This is where you cross over from sci and into investment banking

Anonymous

Anonymous Sat 31 Oct 2020 22:33:12 No.12293764 Report

Quoted By:

>>12293614
A lot of traditional hedge funds ARE low frequency whereas algorithmic funds are higher frequency because the signals they generate are very short lived.

Anonymous

Anonymous Sun 01 Nov 2020 13:43:26 No.12295517 Report

Quoted By:

>>12289429
create models that can take non-gaussian distributions

Anonymous

Anonymous Sun 01 Nov 2020 13:48:08 No.12295525 Report

Quoted By:

Why does that chart look upside down?

Anonymous

Anonymous Sun 01 Nov 2020 17:19:17 No.12296002 Report

Quoted By:

>>12289429
>beat the market

Anonymous

Anonymous Mon 02 Nov 2020 05:24:06 No.12297962 Report

Quoted By: >>12299561

>>12293424
Use of ML in finance is incredibly sparse despite the push in recent years turns out that it is not a good tool for work

Anonymous

Anonymous Mon 02 Nov 2020 16:15:13 No.12299338 Report

Quoted By:

>>12289429
you are trying to predict all human economic interaction.

Anonymous

Anonymous Mon 02 Nov 2020 17:20:03 No.12299561 Report

Quoted By:

>>12297962
There are many open lectures from bulge brackets on how ML has been used and where it's failed in recent years in finance. It's being used heavily across the board.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches