Hi is anyone using co-integration as a filter for tradable pairs. I noticed that Catalyst Corner gives a lot of pairs are cointegrated but when you go to test them yourself it's rare that they are actually cointegrated. How is everyone else checking cointegration? Is it right that it is the main filter? GLD/GDX for example seems to be a very profitable pair however is not cointegrated. Pairs that I HAVE found to be cointegrated THUS far are only ETF's - I am going to go check each pair in each sector for cointegration and hopefully have more to work from. PTF on its own just doesnt do it for me.
You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
Pair Trading Strategy Journal
- Thread starter jonnysharp
- Start date
Their glossary only says, "What Does Cointegration Mean? Cointegration is the statistical confidence that a pair will revert to the mean"Quote from clarodina:
How does catalyst corner calculate the cointegration percentage? Anyone knows? which type of cointegration does the web page uses?
Or is the formula proprietry?
Their demo says something about likelihood of reverting to the mean.
I've haven't asked them, but I suspect it is something simple like the ratio of coefficient in the Engle-Granger cointegration.
I keep reading academic papers hoping to find someone relating the coefficients in the Error Correction Model based on the Johansen procedure with a rate of convergence/reversion. Given my half-remembered ODE coursework, one should at least be able to say that from the current perturbation, here is an exponential decay function to the steady state if there are no further perturbations.
The fact that no one has provided makes me suspect that it would have no bearing on real market data. It doesn't seem that hard to find out though if I just had the time.
I can't speak to Catalyst Corner's Cointegration Percentage since I don't actually know what it is.Quote from mpat89:
Hi is anyone using co-integration as a filter for tradable pairs. I noticed that Catalyst Corner gives a lot of pairs are cointegrated but when you go to test them yourself it's rare that they are actually cointegrated. How is everyone else checking cointegration? Is it right that it is the main filter? GLD/GDX for example seems to be a very profitable pair however is not cointegrated. Pairs that I HAVE found to be cointegrated THUS far are only ETF's - I am going to go check each pair in each sector for cointegration and hopefully have more to work from. PTF on its own just doesnt do it for me.
I just ran GLD vs GOLD
Catalyst Corner comes up with cointegration 1.00%. When I run GLD vs MSFT, I get a whopping 90%.
The prices for GLD vs GOLD are 80-90% correlated over the last 5 years (diverging currently.)
The log price is 70-90% correlated, and the return is 63-72% correlated.
Not correlated enough on the return basis to be an HFT trade, but the current divergence in price has some cause which might be tradeable.
Perhaps their Cointegration % is a shorter-term construct?
Hello, I am new to this forum and industry.
From everything I have read so far co-integration is key to finding profitable trades; however, I am having the same problem you guys seem to be having... How is catalyst corner coming up with there co-integration #s?
I was also wondering if any of you have experience with stat arb in the FX market? For equities I take end of day (or close to) prices and run my analysis; but, with FX the market never closes so by the time I run the analysis the prices have already changed. Does anyone know how to work around this/incorporate this?
Thanks. I am looking forward to reading more of your ideas and adding my 2 cents (probably not even worth that much).
From everything I have read so far co-integration is key to finding profitable trades; however, I am having the same problem you guys seem to be having... How is catalyst corner coming up with there co-integration #s?
I was also wondering if any of you have experience with stat arb in the FX market? For equities I take end of day (or close to) prices and run my analysis; but, with FX the market never closes so by the time I run the analysis the prices have already changed. Does anyone know how to work around this/incorporate this?
Thanks. I am looking forward to reading more of your ideas and adding my 2 cents (probably not even worth that much).
I have developed some code to check through the stock market for cointegrated pairs using end of day data. The number of cointegrated pairs I have is absolutely astonishing (so far a small percentage of stocks have been tested but I already have too many pairs to know what to do with), and using a simple bollinger band strategy backtested using probacktest from prorealtime, the profitability of these pairs is astounding too.
What filters do people use? I was thinking of incorporating half life for the mean reverting property and filtering out those greater than 30 days. Ideally I would need to back test each pair and presumably calculate a sharpe ratio and filter out further pairs on this. Does anyone have experience back testing? I assume I would have to step through the data as the hedge ratio would constantly change as the time series steps forward. Seems like I'm going down the route of a massive project with a huge learning process along the way. I am surprised that I was even able to get as far as create code to check the markets for cointegrated pairs so I suppose with a lot of work I should be able to get to the required stage. It seems like the best way to run pairs trading is through mathematical filters on the pairsto give a tradable universe alongside some sort of trade management program connecting to the broker API to do the actual trading, automated. Seems there are enough pairs out there to split capital a very large number of times and if done right the strategy could be a very good money maker with minimal effort required once the hard work has been done. Keen to here the input of others who have gone down this route.
edit: another thing i was wondering was the market cap of traded pairs, i was thinking it would be wise to filter the pairs to only includes these where the market cap exceeds a threshold, say $5bil, but does this actually add value?
What filters do people use? I was thinking of incorporating half life for the mean reverting property and filtering out those greater than 30 days. Ideally I would need to back test each pair and presumably calculate a sharpe ratio and filter out further pairs on this. Does anyone have experience back testing? I assume I would have to step through the data as the hedge ratio would constantly change as the time series steps forward. Seems like I'm going down the route of a massive project with a huge learning process along the way. I am surprised that I was even able to get as far as create code to check the markets for cointegrated pairs so I suppose with a lot of work I should be able to get to the required stage. It seems like the best way to run pairs trading is through mathematical filters on the pairsto give a tradable universe alongside some sort of trade management program connecting to the broker API to do the actual trading, automated. Seems there are enough pairs out there to split capital a very large number of times and if done right the strategy could be a very good money maker with minimal effort required once the hard work has been done. Keen to here the input of others who have gone down this route.
edit: another thing i was wondering was the market cap of traded pairs, i was thinking it would be wise to filter the pairs to only includes these where the market cap exceeds a threshold, say $5bil, but does this actually add value?
I too once felt this way, the problem is that you might be running PNL backtests on stocks that you already know had exhibited a particular behavior. For example if you find cointegrated pairs now, if you run PNL analysis, ofcourse they are going to come up as performing well. The same time period you are running tests for is the same time the behavior is identified.
Being cointegrated for a certain time frame essentially means that the pair was mean reverting for that time, so if you test it, ofcourse your gonna be profitable. The question really is, could you have identify and traded thos pairs if it was real time, and not looking back. Did you run forward tests?
Being cointegrated for a certain time frame essentially means that the pair was mean reverting for that time, so if you test it, ofcourse your gonna be profitable. The question really is, could you have identify and traded thos pairs if it was real time, and not looking back. Did you run forward tests?
It sure is changing. What if my criteria was high performing stocks over that last 12 months, and I went and backested these stocks over the same 12 months...that wouldnt make any sense. I would be trading on information I did not have at the time. 12 months ago these stocks did not exhibit the behavior I was looking for, hence the trade would have never taken place.
What you should be doing is this, find cointegrated or correlated pairs or pairs that your psychic told you about, or any other metric you want, find that metric up to 12/31/2010. Those stocks that you get at 12/31/2010, back test from that point until now. This will avoid using information you would not have had at the time of your trades. You absolutely can not trade on information that you would NOT have had. Hope this helps you.
What you should be doing is this, find cointegrated or correlated pairs or pairs that your psychic told you about, or any other metric you want, find that metric up to 12/31/2010. Those stocks that you get at 12/31/2010, back test from that point until now. This will avoid using information you would not have had at the time of your trades. You absolutely can not trade on information that you would NOT have had. Hope this helps you.