191 karma · joined March 10, 2014
About me: HTTPS://alexizydorczyk.com
It's a fantastic tool. Right now, I'm mostly creating individual components (or groups of components rather that perform one functionality) and exporting them to a larger project/app.
Do you think things such a drag and drop functionality will be implemented? Or is the view that most would prefer not to have this?
How do you see this competing with / differentiating itself from RStudio? The python support?
In the statistics department I was recently at, R was the lingua franca and Rstudio was how we wrote it. I have not found a better product yet for ad hoc exploration than Rstudio but will certainly give this a go.
I think an interesting avenue to explore with that product would be more templates/easy-to-start documents/quick tutorials.
LaTeX is the standard in some parts of academia (math, CS, stats) and not others (my brother, a biochemistry student at a public research university with 30k+ students, say nobody has even heard of it there) - I think the amount of market that is still using Word for academic papers (often because they do not know of LaTeX or think they don't have time to learn it) is very significant. ShareLaTeX is definitely a step in making LaTeX more accessible for that segment but there's definitely a long road ahead too.
We're a group of researchers / students working on an analytics platform. The backend of our platform is built in Flask and C++/R.
We're seeking a developer(s) to develop a standalone front end of the app (that makes API calls to the Flask backend).
We require:
- Knowledge of a modern Javascript framework / some founded opinion on which should be used.
- Knowledge of d3.js / Knowledge of data visualization in Javascript
- Experience building dashboard/admin style web apps
- Experience with Flask would be a bonus
We're "technical" in the sense that we have extensive experience with scientific programming / algorithm development / machine learning. We have a basic understanding of Flask. We want to work with someone someone who understands that some parts of the codebase cannot be shared - hence the need for an independent "front end" app that makes API calls. We understand that this may not be an orthodox way to work so patience/understanding are a must.
If this sounds like something you'd be interested in, email us at: btydee@gmail.com
As other commenters pointed out, the fact that you charge by minute not hour does in fact make a big difference in price, particularly for those of us that need to run intensive but sporadic/short tasks.
A few question about your points that I am trying to find in documentation right now but perhaps you can save me the trouble if you happen to see this first:
> Domino makes it really easy to start and manage multiple runs in parallel (think a modern, easy-to-use cluster). If you're doing all this directly with AWS, you're quickly running into pain points managing all your instances and images.
How so? Does Domino allow you spin up more cores at will from R? That would be awesome.
> - Domino lets you deploy your analyses as self-service web UI tools, or deploy them to API endpoints. Doing this on your own would involve building an entire web stack around your analysis.
This is awesome and definitely useful if you are doing work for clients and do not want to be bothered with spending too much time building production grade stuff. In some sense is this like yhathq.com? (I understand you guys do more than they do in the sense you provide all these other features).
Presumably, Domino does not take unparallelized R code and transform it into parallelized code - so if you have to use the parallels R package (or some equivalent) anyway to get it to run on multiple cores, what really is the value add? Am I missing something?
It would be helpful to know what specific deep learning algorithm (convolutional, deep belief?). Or at the very least, what / who's implementation of neural nets did you use in your model and how it compares performance wise to the more conventional tools in NLP (when you give them the same original features to start with).
The real challenge is having the creativity to come up with a question worth asking. From Yelp's perspective, I imagine the most value they get from hosting this competition is the (potentially monetizable) ideas for what types of business problems could be solved using their treasure trove of data.
I'm sure they could have their data science team build an incredibly sophisticated model for some specific prediction task (probably more sophisticated than what any student could reasonably submit) - but if that specific prediction task turns out difficult or not monetizable, that's a huge investment lost. Having many students look at many different ideas is more likely to result in finding the right question.
I think how successful a model is depends on your purpose too - if you have a model that predicts only 60% of games correctly, while technically you're not very "good", you're doing much better than those gambling on the sport (usually the favorite in sports betting wins at a rate of 52-54% I think) or probably conventional 'sports analysts' (no data to support the second point, that's an educated guess based on the gambling statistic).
I disagree with you premise that R is not popular among machine learners and the implication that has for the "bleeding edge" comment. Perhaps it is less popular with ML people coming from a CS background - but it seems to be the most widespread language for ML (or "statistical learning") among statistics people. And that fact is not to be discounted - perhaps the preeminent book on machine learning (or at the very least one of the most popular) - "The Elements of Statistical Learning" - is written by statisticians (and in fact uses R exclusively!). The Journal of Statistical software, which features papers detailing many ML libraries, has far larger coverage of R packages than any other language: http://www.jstatsoft.org/ . I would suggest this is the evidence for the "bleeding edge" comment. Other languages' packages, say Python's ML libraries (scikit-learn, pyBrain, etc.), do note even come close to the breadth of capability in this space.
Coming from a statistics background,- I would venture to ask.. what's wrong with "pushbutton" libraries? And why are they only useful to people who do not know/understand much about ML/statistics?
Say you have a dataset and you believe ,say, a Random Forest would be well suited to predicting some response. Let's assume you have a very good understanding of Random Forests. Why would you not want a push-button library? Why would you WANT to recode the thing yourself? It's not as if your Random Forest will be any better or "more correct" than the one on CRAN or whatever language's repository you are using. I would argue it's more likely to have mistakes (coding something like this from scratch is no small task and the ones on CRAN have often gone through many iterations, improvements, and reviews from many experts in the field). And let's say you have some domain knowledge or some informed belief about how the Random Forest needs to be adapted to your particular problem (a different loss function or something) - you can easily edit the R package source code to do that - no need to rebuild the car just to give it a new paint job...
As you mention, the vast selection of libraries (and the continual development of these libraries) make even the python module universe look small (in reference to statistical methods, statistical algorithms/machine learning, plotting).
Further, not only are there a lot of libraries.. but many of them are bleeding edge written by the same people who originally invented the given algorithm. Most similar libraries in other languages (like Python) have incomplete ports of these libraries to them often not written by the same people. Not to mention, the documentation behind these algorithms is often excellent - besides the usual man files, there's "vignettes" (detailed instruction manual) as well as an entire journal that now has become a place to publish papers detailing the usage and implementation of many of these algorithms.
So I'm curious, why leave?
-When for instance is it appropriate to use dendrograms (hierarchal clustering) versus non-hierarchal approaches (ie. k-means clustering)?
-What does it actually mean for two observations to be clustered close together in a dendrogram (and would they also be clustered together if you uses a non-hierarchal approach)?
Speaking of parsing, I was made aware of http://www.tagnifi.com/ I think elsewhere in this thread (I have no affiliation with them whatsoever). They appear to parse XBRL - probably in much a similar way as you do (although, I think they offer just raw data rather than the screening tools you do, so slightly different purpose/target). Just thought I'd share in case you found in interesting/useful.
News articles and blog posts are amplified opinions of a small set of people, they are a poor substitute for direct measurement of many many people.
You might be correct (sounds very reasonable) but I guess the question is why do I necessarily care about directly measuring many many people? What if loud, amplified opinions are actually better indicator of market momentum?. Tetlock's article suggests that those amplified opinions actually do contain useful information. Again, I have yet to find any conclusive evidence that measuring people's chatter actually ... matters.
Derwent licensed the technology and subsequently ran into trouble because the tech and data was not sound.. though Bollen's initial theory was correct.
Right - I think there are more than a couple complaints about his paper. Why do you say his initial theory was correct though? If I recall correctly, his initial theory was that social chatter contained market information that could be extracted using neural networks and used to predict short term market returns (like a few days forward). I haven't really seen anything along the same lines (especially using his approach of neural networks) anywhere - although the articles you link to might be considered similar.
I am just reading through the articles you posted from Deltix now, sound interesting. One problem is that because of the relatively novelty of services like PsychSignal, the data does not go back as far as traditional indicators (this was a problem with Bollen's paper as well). When you say you have impressive alpha, sharpe, or whatever for 6 months.. that does not really prove much, especially in the eyes of academics. I'd love to replicate or work with this data presented in Deltix - but given the data is proprietary, that's difficult to do.
side note: This was just mentioned in the Wall Street Journal yesterday, http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1807265 claims to find signal in blog posts. That's not to say measuring many people's chatter wouldn't work as well or better though. Also just heard of paper today, so no guarantees about quality but maybe someone finds it interesting.
The only concern - and maybe you have a solution - I have is that to my knowledge, XBRL has only been required on 10K's and 10Q's since around 2009 (I may be incorrect, 2010?). As someone involved in research with financial data, there's really no avoiding using pre 2009 data. Parsing that from raw text files is even harder than XBRL - is TagniFi able to do that as well? (I've perused your website but haven't found the answer). Unfortunately, without that capability, I'm forced to rely on services like Fact Set at least partially.
I am interested in your claim that the data has alpha and smart money is trading on it. I have been looking for any peer reviewed, academic articles actually documenting that alpha (ie. by anyone except the people trying to sell you the data). One such paper I found was Bollen et al's http://arxiv.org/abs/1010.3003. From my understanding, there are several problems with the methodology (I could find a rebuttal paper I read if you're interested) and the methodology was used in Derwent Capital - although the fund only traded it for 1 month: http://venturebeat.com/2012/05/28/twitter-fueled-hedge-fund-...
Care to point me to anything else like that?
That said, your startup looks cool and could be very useful if you're able to extract data all historical financial data from annual and quarterly filings (something that is very difficult right now) but it's worth noting that there any many startups in the space of making investing easier for retail investors (many of whom frankly aren't passionate about or interested in anything to do with investing beyond growing their personal funds) probably the most notable ones are maybe Wealthfront https://www.wealthfront.com/, Robinhood https://www.robinhood.com/, and most related to your startup, Quandl.