Intraday Pair Trading/Stat Arb
- Thread starter Shanb
- Start date
I just started demo testing a fx correlation system found on forex factory. I would suggest reading the thread if you are interested.
When do you take profit for divergent approach, ratio chart touching bbands could become a trending channel?
I'm talking stocks...
Stats Arb and Reversion To Mean used to be very profitable...
10 years ago you would make "easy" money from:
(a) the reversion if your quant analysis was good
(b) the spread
It was about 50/50.
Today, it's just the spread...
The quant analysis, etc... just keeps you hedged enough...
So you can maintain a market neutral Portfolio...
While you grind a lot of small trades...
The volume and $0.01-0.02 you capture from spread does add up.
It only works if you're Automated and can do volume...
But your profits go up 200-300% during high volatility...
Because stock spreads can widen dramatically.
To do this is Forex, you better be trading 24 hours...
Because the biggest opportunities are in off hours relative to USA.
In both Forex and Futures...
The spread in not enough...
You must have real-time squawk news...
And be sophisticated enough to momentum trade.
There is no easy way to make money trading in 2011...
Unless you put up a web site like Bone and sell "services".
I kindof skipped over a bunch and went to the end of the thread but I know a bunch about ETFs so I'll just post.
What is your firm's data, hardware and execution like? (This isn't a knock against you but if you are asking these questions I guess its not that good.)
Do you have TCP or UDP data? (Please don't take UDP as a new "must have" buzz word for algo trading)
Do you have risk check or true DMA? If you have risk is it front-end or back end? Is it local (on machine) or remote (risk server)? How long does it take to check?
What are your rates like? Who is your prime and what are your borrows (short list) like?
Do you have enough capital to sustain a $5-$20k drawdown during testing?
What is your firm's data, hardware and execution like? (This isn't a knock against you but if you are asking these questions I guess its not that good.)
Do you have TCP or UDP data? (Please don't take UDP as a new "must have" buzz word for algo trading)
Do you have risk check or true DMA? If you have risk is it front-end or back end? Is it local (on machine) or remote (risk server)? How long does it take to check?
What are your rates like? Who is your prime and what are your borrows (short list) like?
Do you have enough capital to sustain a $5-$20k drawdown during testing?
Ummm... there has never been an easy way to trade and make money. If you thought is was so "easy", then you are foolish and naive to have believed that the ATM would last.
And I have 70 clients, all who are ET Members, who make money spread trading futures pairs, butterflys, and condors. And they provide excellent references.
In terms of futures spreads, this is another naive and uninformed opinion.
Hi Shan, a few comments on your project from my side:
1) Edge
There are a myriads of hedge funds out there that have erased any possible edge with such simplistic approaches as described by Ernie Chan. He also emphasizes in his book that they are only a starting-point for further research. So these simple models do not work anymore, either you find your edge in illiquid stuff or you have to come up with something smarter. You might go into the shorter-term direction, but don't forget that factors like fees and execution come into play then where you have no chance against other players.
2) Software
As you mentioned there are two possible ways to trade stat-arb / pairs. Either you do it like your group, with a focus on longer-term pairs trading, incorporating additional inputs like fundamentals etc. Then you might be fine with using Excel.
However, if you want to go the route you mentioned - intraday stat-arb - then I think there's no other way than to learn a programming language and do it all on your own (as others already mentioned)
You mentioned Tradestation: It is probably the worst of the retail softwares out there for stuff like that. It has no/ only poor portfolio testing capabilities, all you can probably do is to conduct some simple pairs trading tests, but that's it.
Consider this: Using only the 500 stocks from the S&P500 gives 124750 possible pairs. In no way can you test this with Tradestation. And don't forget, we are only speaking of pairs up to now.
Also in Excel this would be a pain in the ass, even when using VBA (although it's possible).
So either you focus on a few selected pairs / relationships, then you might be happy with Excel or any retail trading software out there. Or you learn Matlab/R/Python/whatever, but then it's definitely more than a "side project" as you mentioned.
3) Data
One often overlooked factor when backtesting stat-arb like systems is data: As such systems will be of mean-reversion type most of the time, every data error, i.e. a wrong outlier - will present you a profitable mean-reversion opportunity that didn't really exist. This bias will always works against you, i.e. improve your backtest based on wrong assumptions. So you better get high quality data or apply appropriate data filtering.
The second point is that you have to be aware of the survivorship bias (I think I mentioned this point to you already somewhere else)
1) Edge
There are a myriads of hedge funds out there that have erased any possible edge with such simplistic approaches as described by Ernie Chan. He also emphasizes in his book that they are only a starting-point for further research. So these simple models do not work anymore, either you find your edge in illiquid stuff or you have to come up with something smarter. You might go into the shorter-term direction, but don't forget that factors like fees and execution come into play then where you have no chance against other players.
2) Software
As you mentioned there are two possible ways to trade stat-arb / pairs. Either you do it like your group, with a focus on longer-term pairs trading, incorporating additional inputs like fundamentals etc. Then you might be fine with using Excel.
However, if you want to go the route you mentioned - intraday stat-arb - then I think there's no other way than to learn a programming language and do it all on your own (as others already mentioned)
You mentioned Tradestation: It is probably the worst of the retail softwares out there for stuff like that. It has no/ only poor portfolio testing capabilities, all you can probably do is to conduct some simple pairs trading tests, but that's it.
Consider this: Using only the 500 stocks from the S&P500 gives 124750 possible pairs. In no way can you test this with Tradestation. And don't forget, we are only speaking of pairs up to now.
Also in Excel this would be a pain in the ass, even when using VBA (although it's possible).
So either you focus on a few selected pairs / relationships, then you might be happy with Excel or any retail trading software out there. Or you learn Matlab/R/Python/whatever, but then it's definitely more than a "side project" as you mentioned.
3) Data
One often overlooked factor when backtesting stat-arb like systems is data: As such systems will be of mean-reversion type most of the time, every data error, i.e. a wrong outlier - will present you a profitable mean-reversion opportunity that didn't really exist. This bias will always works against you, i.e. improve your backtest based on wrong assumptions. So you better get high quality data or apply appropriate data filtering.
The second point is that you have to be aware of the survivorship bias (I think I mentioned this point to you already somewhere else)
Thanks for the tips...after realizing that this would require alot more work than anticipated I am slowly working on some stuff. Mostly I am working on refining a system that I'm already running.
I am definitely aware that the simple strats have been arbed away by the larger players. Even then they operate on a different TF and with different liquidity constraints that I have.
My original idea started with a conceptual idea of industries and underlying sectors. Like I mentioned before...for example I would hypothesize that there is a underlying relationship between luxury retailers and the general consumer discretionary sector. I would stand to say that I often see sub-sectors get out of wack with broader sectors and there is an "appearance" of a narrowing spread. Without programming knowledge the only way I would be able to do this is to use specialty ETF's vs broad ETF's. This may be suboptimal and testing would have to confirm this.
I may be stuck with doing individual stock pairs and even then it would have to flow from a logically consistent idea as to why these things should even work. That would cut my universe of stocks down considerably. I do have some good pair guys in my office and need to sit down and get some specifics of what they are looking at as they are doing individual stock pairs.
That last tip is key here, a programmer friend of mine has mentioned this to me. Clean data is very important, will definitely keep that in mind!
Also I read through some of the responses and couldn't even understand what some of the questions were. Talking about the ultra-specific quant and exceuction stuff. Although I realize this is important, Ive seen many people make a killing with relatively simple ideas and execution. You would think this would be arbed away, but experience has taught me this is not the case...specifics do matter though.
I asked some of the pair guys in my office what they were using for some of their work and almost LOL. This guys are pulling in 6-figures a month and they wouldn't know if some of you guys were speaking English!
I asked some of the pair guys in my office what they were using for some of their work and almost LOL. This guys are pulling in 6-figures a month and they wouldn't know if some of you guys were speaking English!