Pairs Trading

Do we need any more ads for BT?

This whole thread, the ads, Don's posts, rerunning an "interview."

I would not be at all surprised if BT was the actual owner of this website.

As to pairs trading, it's a great business model

1 ) Charge for boot camp

2) Charge for "mentoring"

3) Charge for correspondence course

4) Fuel traders with hopes & dreams (see pairnomics website)

5) Encourage high volume strategy that generates lots of commish $

6) Charge for capital usage

7) Sell pair-trading technical info

The only thing nobody mentioned is the difference between "pair-trading" and long/short equity trading. Of course that distinction seems lost to most pair traders.

Look at a long/short equity fund that really knows their stuff and you will find out the difference between that and pair-trading.

Again, what a GREAT business model.....
 
Quote from itmediaco:

segv... I took a cursory glance at your code and basically, please correct me if I'm wrong, it just sorts a universe of stocks by highest correlation and stops after 200 pairs are made.

You could generate 1000 random walks and just as easily find correlations in those samples... but the resulting pair is still random. i.e., There isn't any new information gained in high correlation pairs, and your profit potential is much lower because there isn't enough spread.
More...

Vidyamurthy had speculated in his book "Pairs Trading: Quantitative Methods and Analysis", that stocks with correlated returns might be a good place to start looking for residuals that satisfy the conditions for co-integration. He advocates a metric he termed the "distance vector", which is the absolute value of the correlation coefficient between returns. The code takes the list of instruments from the instrument list calculates the absolute value of the correlation coefficient of returns, and creates a list of unique "pair" instruments sorted from highest to lowest coefficient. This is the starting point for building the basket based on Vidyamurthy's method. The next step would be to test each of the pair instruments for co-integration and trade-ability. Continuing with Vidyamurthy's method, that can be accomplished by regressing the time series of each instrument in the pair and testing the residual for stationarity (Engle-Granger test). Alternatively, you could use the Johansen procedure to build synthetic instruments, and test the resulting time series for stationarity. The question is whether or not any of this kind of analysis has any kind of real information content, which brings us back to your point, and back to my point. Your point is exactly the point I have been attempting to highlight, and that is that "technical analysis" applied to random data is just random data. You have to have some sort of theoretical, scientific framework to work with. The problem is that everyone else has a similar framework, and equal access to the data in question. The universe we are talking about is very small. At the end of all of this analysis you will find a handful of instruments that satisfy the conditions for co-integration and trade-ability, and everyone will already be trading the same instruments. To make matters worse, when you start to stress-test these kinds of baskets, you find that diversification and market neutrality fail catastrophically under a variety of extreme conditions. You are effectively "short the tails", to put it in trader talk, since you are long long the correlation. Short gamma, short volatility, long correlation, and hardly a magic bullet.
 
For all the long informative yet complex post in this thread, do any of you even know the real roots of pair trading? It's quite simple, no real math needed to understand.

It's like 40 years old, if not more.

These stocks are not correlated because of some magical market number theory, you know.
 
Quote from segv:

Vidyamurthy had speculated in his book "Pairs Trading: Quantitative Methods and Analysis", that stocks with correlated returns might be a good place to start looking for residuals that satisfy the conditions for co-integration. He advocates a metric he termed the "distance vector", which is the absolute value of the correlation coefficient between returns. The code takes the list of instruments from the instrument list calculates the absolute value of the correlation coefficient of returns, and creates a list of unique "pair" instruments sorted from highest to lowest coefficient. This is the starting point for building the basket based on Vidyamurthy's method. The next step would be to test each of the pair instruments for co-integration and trade-ability. Continuing with Vidyamurthy's method, that can be accomplished by regressing the time series of each instrument in the pair and testing the residual for stationarity (Engle-Granger test). Alternatively, you could use the Johansen procedure to build synthetic instruments, and test the resulting time series for stationarity. The question is whether or not any of this kind of analysis has any kind of real information content, which brings us back to your point, and back to my point. Your point is exactly the point I have been attempting to highlight, and that is that "technical analysis" applied to random data is just random data. You have to have some sort of theoretical, scientific framework to work with. The problem is that everyone else has a similar framework, and equal access to the data in question. The universe we are talking about is very small. At the end of all of this analysis you will find a handful of instruments that satisfy the conditions for co-integration and trade-ability, and everyone will already be trading the same instruments. To make matters worse, when you start to stress-test these kinds of baskets, you find that diversification and market neutrality fail catastrophically under a variety of extreme conditions. You are effectively "short the tails", to put it in trader talk, since you are long long the correlation. Short gamma, short volatility, long correlation, and hardly a magic bullet.
More...

I'm not an OpenQuant user so I had to look up GetReturnSeries(). Now it makes sense; differences in returns is stationary having stable statistical properties. Whereas, a price series alone has no finite variance.

The exception is if two series are perfectly cointegrated, in which case, the ratio is no longer random, but a stationary process. That rarely happens in the real world...

There is no advantage in pairs trading! Why doesn't everyone see that by now?
 
Quote from itmediaco:

There is no advantage in pairs trading! Does everyone understand that by now? [/B]
More...
If the logical structure imposed by pairs trading gives you the confidence to trade, and you do so profitably, then it most certainly offers an advantage.
 
Quote from Yisterwald:

If the logical structure imposed by pairs trading gives you the confidence to trade, and you do so profitably, then it most certainly offers an advantage.
More...
Self-confidence, knowledge, experience, discipline and concentration are necessary elements of a trader's personality. It just bothers me when companies like PairCo, daytradepairs.com, et al. sell this BS to naive traders.
 
Quote from itmediaco:

Self-confidence, knowledge, experience, discipline and concentration are necessary elements of a trader's personality. It just bothers me when companies like PairCo, daytradepairs.com, et al. sell this BS to naive traders.
More...

Now that I'm being called out as selling "this BS to naive traders", I must respond.

I have put hundreds of hours into developing the IPTS. It is something I am very proud of and I use it every day. For every person like yourself that insults and puts down pair trading, there are just as many people out there trading this style and making money. Nothing makes me happier than getting an email from a user telling me how they are finally a profitable trader and its all due to pair trading. Check the comments on Ebay for the IPTS, they are all positive.

The original poster of this thread asked "Can it be done on a intra day basis"? Have you done any analysis to prove it doesn't work? And not just theory mentioned in some previous posts, but actual test with actual data. I've done actual tests: its called trading.

Nobody is saying that pair trading is the Holy Grail of trading. Trading still involves "Self-confidence, knowledge, experience, discipline and concentration" that you so rightfully point out.

I guess its going to come down to "agreeing to disagree". You can keep saying that it doesn't work, but the traders that actually try it might find a profitable trading technique.

Jeff
 
Mr Pairs, we have someone claiming they make 300-500 a day trading your system with only 500 share lots. No indication how many trades needed to make that.

Since most of the stocks presumably have high liquidity, you're talking about a $5000 a day cash machine (@ 5000 shares), even for the paupers on ET.

I find that incredible. To net that much would imply a huge edge, which could easily be shown in backtest, and not with a few hand picked winners a month.

Im not about to subscribe to tradestation to test this, but I'd love to hear from someone that bought your system and did so.


PS I don't doubt that pairs trading is useful, and I do a special subset of it myself, with success. I do think you need more than simple correlation and divergence though, or the black swan will take a dump on your head eventually.
 
Jeff, Does your system use cointegration or correlation? Log returns or non-stationary price series?

In my tests, historically cointegrated series (Granger-causality test, error correction model) consistently broke down out of sample. Secondly, choosing an alpha to input to the regression is far from an optimal procedure... it's mostly heuristic. Third, significant correlation doesn't determine whether the pair will mean revert or not to some common ratio. A confidence interval of the correlation of 0 means the pair is not suitable trading. BUT a high correlation (positive or negative) means nothing because you can't conclude anything new. Thus, your trading decision is independent from those observations!

I also tried building a portfolio of spreads, in which I traded one leg at a time and the final residual is hedged with futures/options. But that's not "pure" pairs trading (hence, my conclusion).

Granger won a Nobel Prize for his discovery of cointegration. It's very powerful if used properly, and I think using it for pairs trading is the wrong approach. Described below is a paper explaining one effective use:

The paper disentangles the different effects of long-term relations on optimal asset allocation with different planning horizons: error-correction mainly affects tactical asset allocation, while cointegration affects strategic asset allocation. The paper also presents results on the effects of incorporating an incorrect number of error-correction mechanisms in financial decision models. Mis-specifying the number of cointegrating relations in a scenario generator can induce either inefficient or overly risky financial management decisions.

http://www.feweb.vu.nl/fb/papers/almcoint.pdf
 
I was thinking about this some more, and remembered that cointegration can exist with or without Granger causality. i.e., contemporaneous cointegration is possible, as well as a leading-lagging effect.

Jeff, which relationship does your system use? Convergence or lead-lag? If the latter, how do you quantify whether the leading leg and lagging leg has 'moved enough' (if price moved at all)?
 
You must log in or register to reply here.
Back
Top