Attempting to front run a stock's move based on the mining of internet sentiment via some cutting edge Neuro-Linguistic Programming voodoo seems ripe for nefarious exploitation to me. Pretty easy to trick fuck something like that. Look at what the Russians did with Tweets leading up to the 2016 election. Huge Achilles heal on this stuff.
Natural Language Processing
- Thread starter 2weels
- Start date
There is no nefarious, magical, deceitful behavior in this at all. It's all about getting your order in before the big boys do so. And that in a 100% ethical and legal way.
You misunderstood my point. The point is the "sentiment" can be spoofed by a skilled coder. The stock runs up on nothing, and the "nefarious" step in and short it. Traders thought they had an edge... and get F'd.
Last edited:
I have listened to a few of the TiML (This Week in Machine Learning) podcast episodes where they talk to cutting edge machine learning experts in different fields. The consensus seems to be among the ML community, by industry leading experts who have the finger on the pulse, that the ML algorithms and data mining methods are definitely being used in the financial space - though not even their buddies in that space are talking about it.
Like any 'edge' in the financial space, it makes sense. You aren't going to blog about your exact technique / method for beating the markets if you actually are beating the markets, thus not going to be much application material for NLP and trading online by those doing it. I have found Turing Finance a good website for understanding how neural networks etc can be used for quant models, but not sure if they have any NLP ideas on there. Maybe read The Unreasonable Effectiveness of RNNs blog to get inspired to test your own ideas. Your tweet scraping natural disaster algo certainly sounds like you have the goods to come up with an edge using these tools.
After getting my ass handed to me trying to take my Eurex / Treasury discretionary trading and make some ticks in the Nikkei - i got inspired and have been ticking off coursera ML courses this last year. A lot of the algorithms definitely seem to have applications in trading. Though from googling just a little, I get the impression that A LOT of guys are trying to exploit financial markets using the predictive property of supervised learning / neural networks which makes it a crowded space. I think in one podcast an expert suggested unsupervised learning and reinforcement learning as the areas financial titans are currently researching for trading / investing.
My 2c anyway.
Like any 'edge' in the financial space, it makes sense. You aren't going to blog about your exact technique / method for beating the markets if you actually are beating the markets, thus not going to be much application material for NLP and trading online by those doing it. I have found Turing Finance a good website for understanding how neural networks etc can be used for quant models, but not sure if they have any NLP ideas on there. Maybe read The Unreasonable Effectiveness of RNNs blog to get inspired to test your own ideas. Your tweet scraping natural disaster algo certainly sounds like you have the goods to come up with an edge using these tools.
After getting my ass handed to me trying to take my Eurex / Treasury discretionary trading and make some ticks in the Nikkei - i got inspired and have been ticking off coursera ML courses this last year. A lot of the algorithms definitely seem to have applications in trading. Though from googling just a little, I get the impression that A LOT of guys are trying to exploit financial markets using the predictive property of supervised learning / neural networks which makes it a crowded space. I think in one podcast an expert suggested unsupervised learning and reinforcement learning as the areas financial titans are currently researching for trading / investing.
My 2c anyway.
Graham, wow, great resources, thanks. Yes, it's inevitable that ML is going to dominate trading. Look at the recent changes at Blackrock. I'm not too worried about many people using ML in trading. Look how many have used TA for years and still some do well. There are unlimited parameters that you can train ML on (including TA) so there will always be an unexploited angle.
That amazing disaster detection algo was by dumpdapump, not me. Incredibly good stuff.
Here is a good overview of machine learning progress, written by an undergrad ee student at my alma matter, very understandable and a big help to me.
https://adeshpande3.github.io/adesh...p-Learning-Papers-You-Need-To-Know-About.html
I think if you check out these two sources for ML by Hinton you may agree that unsupervised is the way to go, but it's not the case that supervised can't be helpful. Just remember that this is Hinton's personal approach for many years, one of many.
(coursera) https://www.coursera.org/learn/neural-networks
(google)
Andrew Ng's coursera course and youtube videos are also great.
(coursera) https://www.coursera.org/learn/machine-learning?utm_source=gg&utm_medium=sem&campaignid=693373197&adgroupid=36745103675&device=c&keyword=andrew ng coursera&matchtype=e&network=g&devicemodel=&adpostion=1t1&creativeid=156061453600&hide_mobile_promo&gclid=Cj0KCQjwvr3KBRD_ARIsANSQYJofuxuRBcnWIOrUQCEBu0ICrDrzgrGj1OtkBIPEb9xZ16EYdQ4qWTsaAjs4EALw_wcB
This is a quite good intro to ML:
(audacity/google/stanford)
Tucker Balch of Georgia Tech has a good coursera course on algorithmic trading. I wish he'd post part 2
.(coursera) https://www.coursera.org/learn/computational-investing
If you know of other coursera courses besides the above three (or other online courses) on ML please let me know.
Glad to make your acquaintance. I have never done any trading or ML so this is a new adventure and I look forward to good interactions with others on it here. I'm also interested in TA and the other approaches used by folks here but the ML/AI approach is more intellectually engaging for my background. Besides all that, I need to make about a few million bucks this year so I can buy the yacht I want to travel around when I retire.
Jim
This site pretty much has anything you need on ML:
http://www.jeremydjacksonphd.com/quantitative-finance-resources/
http://www.jeremydjacksonphd.com/category/deep-learning/
And these coursers are held in high regard (though Andrew Ngs course I think covers everything. Also look for Andrew Ng's youtube lecture series)
https://work.caltech.edu/telecourse.html
https://www.coursera.org/learn/neural-networks
Good luck with your predictive neural nets and keep us posted.
http://www.jeremydjacksonphd.com/quantitative-finance-resources/
http://www.jeremydjacksonphd.com/category/deep-learning/
And these coursers are held in high regard (though Andrew Ngs course I think covers everything. Also look for Andrew Ng's youtube lecture series)
https://work.caltech.edu/telecourse.html
https://www.coursera.org/learn/neural-networks
Good luck with your predictive neural nets and keep us posted.
Excellent resources. Thanks very much graham, will do.
OK, I've thought about this quit a bit, and my best response is, it seems like looking at the price and volume at time X is great for understanding time X but we want to know about a future time Y. The current price and volume at time X can be fed into any algo, hypothesis, intuition, or whatever and that is great and permits lots of profit. My thought about sentiment analysis, NLP, etc., is for them to be parts of an automatic decision system for trading. Either way, gauging sentiment by looking at the chart, or by NLP and sentiment analysis, are two ways to develop a system. It would seem there are advantages to basing a system on the price/volume data, and different advantages to NLP, sentiment, ML, etc approaches to system. Are the latter better than the former? I don't know yet. I don't really understand either yet. But I do know about ML, neural networks, etc from my day job. Maybe when I learn to read the charts like you can I won't feel the need for all the other tricks....time will tell. I'm trying to learn both TA (and general trading approaches...if that's a word) and AI/ML etc.
Thanks
Jim
The standard is:
x=time;
y=price;
the traditional index for time is:
time.historical=t[x];
time.current=t[0];
time.future=t[-x];