[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/biz/ - Business & Finance

View post   
File: 32 KB, 575x367, IMG_0020.jpg [View same] [iqdb] [saucenao] [google]
1806756 No.1806756 [Reply] [Original]

Hey /biz/, was thinking about training a neural net to make long term investments. What are your thoughts? Good idea/bad idea/waste of time? I figured at the very least it might make a good project to teach myself to work with machine learning frameworks. Also anyone know any good neuromorohic computing companies to invest in?

>> No.1806792

Slightly related: any quants on this board? Do you like it? Do you really need a PhD?

>> No.1806793

>>1806756
>Good idea/bad idea/waste of time?
it's not just a bad idea but also a huge waste of time and money.

>> No.1806820

One does not simply train a neural net to predict stock prices...

It's a very difficult problem, but I agree that it would make a good project to study natural language processing, data analysis and statistics.

What experience do you have with ML?

I'd be interested in collaborating on something if you have a bit of experience - I know a bit about state of the art computer vision and reinforcement learning, but I'm only just getting started with NLP.

>> No.1806822

>>1806820
>One does not simply train a neural net to predict stock prices...
it has been done plenty of times tho with some success very short term like predicting that days closing prices.

long term it simply can't work.

>> No.1806836
File: 53 KB, 346x396, 1485185154631.jpg [View same] [iqdb] [saucenao] [google]
1806836

>> No.1806841

>>1806822
Yeah I could imagine that doing sentiment analysis on news and social media feeds to predict short term spikes / falls wouldn't be too hard, but as you say, long term would be near impossible.

Nonetheless, can you point me to some examples of where this has been successfully done before?

>> No.1806848

>>1806841
>do my homework for me

google it yourself dumb nigger

>> No.1807019

>>1806820
Sorry, this would just be something to get my feet wet. No prior experience.

>> No.1807027

>>1807019
Although, I write code for a living if that makes any difference. But pretty much simple CRUD shit. It's not very intellectually fulfilling which is why I'm looking into this.

>> No.1807028

>>1806822
>long term it simply can't work.
What makes you say that?

>> No.1807032

>>1807028
I guess I'll just answer my own question: I guess you can't write anything that can account for some new disruptive technology appearing.

>>1806841
Don't know if you've seen this but it's a form of what you were describing https://twitter.com/deepdrumpf?lang=en

>> No.1807033
File: 275 KB, 1280x720, th.png [View same] [iqdb] [saucenao] [google]
1807033

>>1806756
good idea anon

>> No.1807036

waste of time

>> No.1807073

>>1806820
I set up a github. https://github.com/trinitron83/Trademan
If you want to work together send an email talking about something you've worked on so I know it's you to this: coniphicac@housat.com

>> No.1807084

>>1807028
you can't train a neural network if you have changing inputs simply put.

you need finite known inputs for it to make predictions be able to distill the data so that cause and effect delays are accounted for in real systems (neural networks can't do this they will just give crap result with no correlation) the news and developments in the world that move the prices long term can not be turned into finite scalar inputs and even if you could do it the data is insufficient to train a decent network.

that is why i said it's not doable.

>>1807073
what the fuck do you think data sciences are still in baby shoes where you have to write custom software for stuff like this? it's about 3-5 clicks with a mouse to set an experiment up but getting, sanitizing and transforming the data is insanely hard for something like this.

>> No.1807087

>>1807084
so that cause and effect delays are accounted for in real systems
Could you expand upon this?
>but getting, sanitizing and transforming the data is insanely hard for something like this.
Failing can be a better learning experience than succeeding. You seem to know this as a fact, it might be fun to find out why.

>> No.1807090

>>1807087
Forgot my greentext
>

>> No.1807105

>>1807090
i remember an example where they tried tried to find correlations to certain settings and measurements in an oil refinery and output quality. turned out a lot of the settings had a delayed effect in the system and until they would find the offset for every single thing they couldn't even begin to make sense of the correlations.

>> No.1807122

>>1807105
Ah. I see. Damn. So I guess in short it's kind of like predicting the movements of a bunch of molecules in a hurricane based off past measurements of the movements of those same molecules?

>> No.1807141

>>1807073
Brain, you have to be 18+ to post here. Learn programming and math first before trying to build huge projects.

>> No.1807151

>>1807141
Heh. hehehehehe. thanks for the tip

>> No.1807160

>>1806756
You'd be better off learning machine learning/AI and getting a high paying tech job.
Hedge funds have teams of Phd's working on AI trading systems that are already in operation. You can't compete against that.

>> No.1807383

>>1806792
you need capital. most quants at my firm are doing some basic shit but with good execution it's enough.

>> No.1807386

>>1806822
actual quants are not using ML for predicting prices. people use ML for things like portfolio combination weights and execution strategies.

>> No.1807640

>>1807386
there have been numerous publications and experiments on the subject but like i said all success is very short term. and probably has to do with the ml experiments getting a feel of the automated trading.

>> No.1807875

>>1807383
>>1807386
>>1807640
Good stuff mang. Thanks for the info.