Steven A. Cohen’s Newest Bet: Do-It-Yourself Computer Traders
wsj.com
wsj.com
This seems like a pretty good idea. A billionaire who runs his own money and just spent a pile to build out his own quantitative team invests a "small amount" to see if there is something to the wisdom of the crowds there.
The only point to note is that quantopian has 85,000 people trying 400,000 algos, something about monkeys and Shakespeare seems applicable.
If my experience is any indications the funnel of ideas to profitable strategies gets very narrow very quickly. Something like for every 100 ideas, 40 get tested, 10 look profitable 5 make it into production and 2 actually work out. And then you have to factor in that the average life span for each algo can be measured in months.
Actually that sounds some what depressing:( If you are an algo trader, then you are always running.
Congratulations to the Quantopian team!! This is a big milestone and something you should be proud of!
This is so true and what eventually made me exit this line of work.
In most other forms of human endeavor, time works in your favor. If you're a dentist or plumber for example, as time passes, you gain more experience, build a customer base, can charge more for your accumulated skills, etc. Not so in algorithmic trading. The things that work well today slowly stop working, and you have no recourse. Always on the treadmill. I eventually decided that I didn't want the sustained worry of "what new thing must I invent today so that I can pay rent next year?"
Isn't that more or less a property of the technology industry as well? You have to continually reinvent yourself to remain employable, etc. Perhaps it's more acute in trading, but then again so (often) are the spoils larger.
They only focus on the trading algorithms with no support whatsoever for portfolio allocation, risk management, money management, etc.
This appeals to random-joe gamblers, and is the ideal approach if you're a hedge fund looking to copy ideas. People develop algorithms for you, but are blind to all the missing pieces (risk, money mgmt, etc).
Sorry what?
The algos are coded with python. How can you not include proper risk management, portfolio allocation etc with python?
And that is my point. Quantopian is aimed at inexperienced individuals who have never heard of Black-Litterman and won't notice what they're missing.
Some people want full control of the pipeline. I'm one of them so this doesn't bother me personally.
However I kind of agree with you as it's PITA sometimes that quantopian does not allow importing external or personal libraries so everything that is not included at the libraries they allow to import must be copy/pasted to each algorithm. I hope/think this will improve at some point though..
So if you want even simple risk management tools like tripwires, those have to be included explicitly in each algorithm. It's like programming in BASIC on a Commodore-64 all over again.
If they added the ability to build a pipeline I'd be perfectly happy. I'd even be happy with fixed (Say portfolio/risk/execution/algos) set of pipeline stages as long as you could write the code that ran at each step.
About "stages" I'm not sure I follow - the "algos" are just plain python code so it's up to you how you want to structure your code.
Python is fortunately quite far from commodore 64 Basic which I also happen to rememeber.. with all of it's gotos, pokes and peeks..
(sidenote: I'm not sure if you have ever created your own framework to trade with IB with their api or fix, if you have then you should know how big improvement something like quantopian is.. especially when their costs are basically zero)
So say you have 10 trading algorithms, they'd all send their buy/sell orders on to money management stage (an independent piece of code) which might size the order somehow.
Then, if the order passes money management it goes in to risk, where the risk management can OK it, or nix it altogether.
Having passed that, your order might go on to an allocation manager - which is now allocating resources between the 6 surviving algorithms. Allocation management will weight resources according to however you see fit - maybe Black-Litterman, and pass things along to execution/order management.
EMS/OMS is basically what it sounds like. You say "Buy 100 shares of AAPL", an OMS will go out and make sure they're filled. EMS does the same but tries to make sure it fills them at a good price.
None of these things exist, or are reasonably possible in Quantopian in its current state. Personally, I wouldn't trust a dollar to it.
Yes, I've built my own trading platform on IB. I have the scar tissue, to prove it.
NB: If Quantopian implemented this - even in a fairly clunky, fixed function way I'd consider using them in a heartbeat. Then, my only complaints would be in quality rather than kind.
My current Q "algo" or whatever we call it (it's a ensemble of multiple "algos") actually has almost every layer you describe (although not using the same methods you describe) and some extra layers you don't describe so it's quite strange for me to read how hard or impossible creating something like what I already have in place is.
As I said it's just normal code and python is fortunately quite clean language.
HN won't let me respond directly to this comment. This approach works fine in the small, but as you grow, do research and try new things out it becomes an unmanageable mess.
Same reason we have modularity in software development - to separate differing concerns.
I have quite a long history in creating complex systems so I think my code is quite clean and easy to work with (I have already removed and added methods multiple times). however there are the things that really suck:
1) I'm not able to package the framework to a library
2) copy-paste because of 1.
3) version management hell because of 1.
More importantly, a lot goes into developing an algorithm including clean data, back testing, infrastructure. Quantopian handles all that for you, so you can focus on developing a stellar algorithm. Of course YMMV.
In addition, a lot of strategies don't really work that well unless you have a 'critical mass' of capital with which to capture any meaningful profit.
Execution is also important. It's one thing to have an algorithm that works in theory and is back-tested. Execution is a different ball-game because you have to deal with factors like trading costs/brokerage fees, accounting, etc.
From Quantopian's FAQ: [1]
Q: Who owns my algorithms?
A: You own your algorithms. Everything you write is yours. Quantopian does not own your algorithms; you do.
Q: Are my algorithms secret?
A: Your algorithms are kept secret. We are committed to protecting your intellectual property and keeping it safe. Ideas are some of the most valuable assets anyone has. We take this responsibility to our members extremely seriously
Caveat: In their detailed terms: "If in the course of providing technical support or other maintenance of the Services it becomes necessary for Quantopian to view your private Content, such viewing will be restricted to the very specific technical purpose." [2]
I'll share my algorithm here:
https://www.quantopian.com/posts/xiv-slash-vxx-pair-trade-1
"Pair-trading VXX and XIV based on the StockTwits sentiments of the SPY at market open. The backtest did really well from 2011 to 2014 with 1700-1800% return in 3 years; and flat between 2014 to present-time... would love to see what people come up to reduce the drawdown's and improve the performance from 2014-2016".
The reason I share my algorithms is trading is one of the hardest ways to make easy money. Making money due to slippage, regime change and overfitting bias is difficult, a poor investment (most hedge funds don't beat index funds) but sharing and learning about statistics, machine-learning and big data is a better investment in self.
Also philosophically, most strategies have limited shelf-life, so it is better to learn how to fish than to hold onto the fish you've got.
Plus they don't require exclusivity for the algos so you can start your own hedge fund and trade your algos if you want to. IMHO It's a pretty good deal.
I spent a significant chunk of time doing a contest algo but was a bit hesitant to continue developing on their platform if it was unclear if it had a future.
This investment asserts that Quantopian is viable going forward and a very worthwhile investment (of time) for developers.
http://www.businessinsider.com/visium-asset-management-closu...