Pair Trading Strategy Journal

[amitkumar_ny]

Quote from Kevin_in_GA:

OK, interesting update.

I talked with Jared and asked him specifically about the time frame over which PTF determines the co-integration between a pair. He indicated that it is "based on the total period of data for the stocks."

That got me thinking ... does that mean that if I were to look at a single pair and get data for 1,2,3,4 ... 10 years, each time period saved as a separate backtest, that the co-integration would change?

I used ESRX:MDT (a pair that showed 0.99 co-integration over a 10 year period) and created separate backtest data as described above. Here are the results:

1 year co-integration: 0.34
2 year co-integration: 0.62
3 year co-integration: 0.54
5 year co-integration: 0.76
7 year co-integration: 0.98
10 year co-integration: 0.99

Raises an interesting question ... is it better to rely on long-term co-integration behavior, or short-term? My gut is telling me that short-term co-integration is what will be more important given the duration of most pair trades.

Thoughts?

This is great work. I think I agree with you. Co-integration should break down if something has changed fundamentally in the relationship. You can imagine that might happen between two stocks over 10 years. So it is more important that they are co-integrated in past 1 year or 2 years.

But I don't think the PTF real time co-integration is calculated like that. It seems much more fluid, changing every day. Almost as if using the same look back as correlation. If it was using 2 year data it won't walk 10% because the ratio is getting a little off.

If Jared can add the same chart as correlation over time that would be great. In the meanwhile maybe I run it on two computers, one using 2 year data and one using 3 month data...I don't know....

 

[NKNY]

Quote from ascheer7:

Has anyone else experienced problems running PTF 3.01?

When I run the back tester and then try and find pairs it runs much much slower than the 2.99 version.

I ran 2.99 on my old computer and it too 2 min and 20 seconds to run with 10 stocks to find the 45 pairs. The new version on a new computer took 34 min and 15 seconds to run the same data.

Just wondering if I did something wrong in the installation or setup.

PTF has not responded to my request for help.

Hi Guys, Jared has an upgrade to 3.01 that has corrected some of the bugs...its 3.03. Much more stable...

 

[ascheer7]

How do I locate PTF 3.03 on the PTF site. Jared sure does not make things easy to upgrade.

 

[Kevin_in_GA]

install from www.pairtradefinder.com/PairTradeFinderSetupV303.msi

 

[BD1]

Is there a list of bug fixes/changes for 3.03?

 

[Kevin_in_GA]

I haven't seen one yet.

Also got feedback from Jared that he is now thinking that shorter time-frames for cointegration make more sense for trading.

Kevin

 

[NKNY]

Quote from Kevin_in_GA:

I haven't seen one yet.

Also got feedback from Jared that he is now thinking that shorter time-frames for cointegration make more sense for trading.

Kevin

Kev how shor did he suggest. I've been doing test for five years and evtered a bunch of trades over the last few weeks. My resoning is that I am looking for pairs with solid historical cointegration.

Nick

 

[Kevin_in_GA]

Well, the shortest you can go with PTF is 1 year. If the average trade is 8-15 days, then a 252 day look-back (1 year) would probably be fine.

Personal thoughts here (currently unsupported by backtest data) is that long term cointegration is secondary to short-term.

Having both is good but I want to know that the pair has a high cointegration NOW rather than over a long time frame. Just some random thoughts that need to be tested before trading.

 

[ps18]

This is one of the best pair trade forum i have come across.
keep up the good work everyone.

just want share the idea here as well...

SHORT - EXPD
LONG - UTIW

any thoughts?

 

[Trader13]

Quote from Kevin_in_GA:

... Also got feedback from Jared that he is now thinking that shorter time-frames for cointegration make more sense for trading.

There are really two time factors for a cointegration test. The first time frame is simply the length of historical data used for the test. For example, last n days or years. The second time factor is the lag value used to conduct the test. In theory, this lag value (aka, half-life) should represent the half-life of the mean reversion cycle. The lag is either user-defined as a single value, or more advanced software may iterate and report results across a range of lag values. <i>Interpreting the lag value along with the significance statistic is very important. </i>There are resources on the web to learn more about this.

Now, if I were to abruptly wake you up in the middle of the night and ask you to explain the lag variable, you should be able to provide a complete answer without hesitation. Until you can do that, don't trade mean-reversion strategies.