How should intraday cointegration be calculated?For my purposes I never hold positions over night, so I can basically start and stop the cointegration at the beginning and end of each trading day. Is this the normal method, or should I concatenate the intraday series together and disregard the overnight changes? This seems like a bad idea as it would introduce huge jumps at the end of each day.

cro, over what period do you normally estimate your cointegrating regression?

A week, or month using 1-second quote data, not trade data, so the bid-ask-bounce effect is greatly eliminated.QuoteOriginally posted by: JamesH83cro, over what period do you normally estimate your cointegrating regression?

No takers? I'll state the problem more generally. Please be gentle, I am fairly new to this.It seems that cointegration is typically applied to time series where there are no gaps, e.g., daily observations. Do most methods ignore the weekends, or have any way of accounting for weekends?I am attempting to fit find cointegration vectors for very high frequency data and have been successfull with 1-day periods, but when I estimate the vectors for another day I get different vectors and causality flows.What I'd like to do is estimate a single set of cointegration vectors using several days or weeks worth of data, but I only have observations from 9:30 to 4:30, do I just concatenate the series together and ignore the jumps, or somehow modify my cointegration code to account for it? Has this been done before? Any papers or references would be greatly appreciated.

do you think the intraday cointegration relationships are stable enough to estimate over such a long window? Do you weight your regression?

cro i wouldn't just concantenate the series. i working on something similar, let me have a think and i will get back to you.may i ask what series you are using?J

QuoteOriginally posted by: JamesH83do you think the intraday cointegration relationships are stable enough to estimate over such a long window? Do you weight your regression?I'm not sure that they are stable or not, but it is my plan to find out. I have both weighted and unweighted regression outputs. I don't see why 1-week to 1-month would be too long. If you go much shorter you run into a problem of when to re-estimate your model every few days, it might be possible that it is really necessary. not sure yet.My series are ETFs and 4 to 5 of their top component stocks, sampled every 1-second.

using mids?

QuoteOriginally posted by: JamesH83using mids?Been trying both.. results seem about the same. Bids, asks, and mids.

I am working on exactly the same problem. I think concatenate is not such a good idea. Which cointegration method are you using ? Engle-Granger or Johansen ?

Johansen.. one curious result is that I am getting quite large differences of the optimal lag whether I use AIC or BIC to select it. AIC always seems to select a lag roughly twice as long as BIC.QuoteOriginally posted by: DogonMatrixI am working on exactly the same problem. I think concatenate is not such a good idea. Which cointegration method are you using ? Engle-Granger or Johansen ?

QuoteOriginally posted by: crowlogicJohansen.. one curious result is that I am getting quite large differences of the optimal lag whether I use AIC or BIC to select it. AIC always seems to select a lag roughly twice as long as BIC.QuoteOriginally posted by: DogonMatrixI am working on exactly the same problem. I think concatenate is not such a good idea. Which cointegration method are you using ? Engle-Granger or Johansen ?I am not sure that' s actually surprising , BIC will always penalize more for the loss of degree of freedom

I would prefer using SIC or BIC over AIC. Most of the times they give either same or more reliable results than AIC.For the original question, you can add Stochastic Structural Breaks in your model to deal with jumps. It would take sometime for me to go over my papers and find something useful for you but if you google it there are many papers available on the internet.

Can you expand a bit on SIC? The code I have only uses AIC or BIC.. and what information I can find sometimes calls BIC "SBIC".Also, re: Stochastic structural breaks. I don't think that is needed here, when estimating intraday conintegration the breaks are deterministic and known.. at the beginning an end of each day. I've observed that some series are only together very tightly in the short term and only loosly over the long term. Can cointegration still be used in this case?QuoteOriginally posted by: quantumarI would prefer using SIC or BIC over AIC. Most of the times they give either same or more reliable results than AIC.For the original question, you can add Stochastic Structural Breaks in your model to deal with jumps. It would take sometime for me to go over my papers and find something useful for you but if you google it there are many papers available on the internet.

QuoteOriginally posted by: crowlogicCan you expand a bit on SIC? The code I have only uses AIC or BIC.. and what information I can find sometimes calls BIC "SBIC".Also, re: Stochastic structural breaks. I don't think that is needed here, when estimating intraday conintegration the breaks are deterministic and known.. at the beginning an end of each day. I've observed that some series are only together very tightly in the short term and only loosly over the long term. Can cointegration still be used in this case?QuoteOriginally posted by: quantumarI would prefer using SIC or BIC over AIC. Most of the times they give either same or more reliable results than AIC.For the original question, you can add Stochastic Structural Breaks in your model to deal with jumps. It would take sometime for me to go over my papers and find something useful for you but if you google it there are many papers available on the internet.SIC and BIC are the same thing: Schwartz information criterion is sometimes called Bayesian information criteria. The thinh about concatenating the series is that at the point where you go from day t to day t+1, and if the 2 contiguous points are not really adjacent time periods (one series ends at 4pm and the other starts at 7am), you will have a strange return, that may overwhelm the other returns(which may be hourly,1mn, etc returns)