Hi,I need a clarification as to the applicability of cointegration in ultra high frequency data set. I am talking about 1 mins or less sampling frequency. is cointegration applicable at this level given the non normality and heteroskedasticity of the dataset at this high frequency.Thank you.

What are you trying to do? Can you explain in normal terms?

Thanks for your response: I am doing a research in high frequency data analysis. I have succesfully applied cointegration to equity data using weekly closing prices. Now I want to narrow it down to intraday frequency such as 1 minute sampling frequency.my concern: there is problem of non-normality at 1 min sampling frequency which is very minimal at weekly closing prices. the question is can I still go ahead and apply cointegration at this very high frequency given the problem of non-normality of the log-return.Thanks alot.

other questions to contemplate -at the 1min interval, are the time series dominated by noise, bid-ask bounce, or by the "real" prices?if the effects of these contaminations are too big, does it make sense to apply co-integration (or any other stats. techniqutes) on these data?what data cleaning, preparation steps are needed before we can apply the procedures?I personally tried it on FX rate data but no much success.

avogadroh,Yes, co-integration exists at 1 minute intervals. I am currently trading a MR strategy at this frequency and have been for the last 2+ years.I do not consider 1 minute intervals to be ulta high frequency and the HF guys I know would agree with me.hli7,I agree with you about preparing data. Also maybe your model needs some adjustments so don't give up keep researching.$.02

Marine,thanks for the tip :-)My most profitable strategy is also MR on intra-day trading, multiple assets.cheers!

QuoteOriginally posted by: hli7at the 1min interval, are the time series dominated by noise, bid-ask bounce, or by the "real" prices?At ultra high frequency, information is dominated by the minimum price increment. Especially in equities, your new information will be mostly ticks at the bid and ask. And you will get some cancellation and placement of visible limit orders.But these are real prices. In theory, the bid will take the same amount of time to move 1 cent whether it is the minimum increment, or there are smaller increments. So in theory, you can just call this amount of time it takes the bid to move the minimum sampling interval, and sample the same pattern even if at irregular increments.The minimum increment is more of a barrier to making money than to prediction. If you buy at the offer and sell at the bid, your prediction has to be bigger than the minimum increment to make money. If you enter or exit with a limit order, you will get in and much more slowly, and so your minimum frequency has become much longer. In either case you can make predictions in a trading increment-dominated environment, but have a hard time profiting from them. If the increments were smaller, attempts to capture the correlation would fill it with noise, and your predictions would not be any better.So with big trading increments you can see correlation but you can't catch it. With smaller trading increments, the efforts of traders to catch what they see fills the ticker with noise, so that your inputs and your comprehension of the action are not automatically better.

Marine,Thank you very much for sharing information.QuoteOriginally posted by:MarineI agree with you about preparing data.Would you be able to make a suggestion what sort of data preparation would be suitable for cointegration models?Best Regards

QuoteWould you be able to make a suggestion what sort of data preparation would be suitable for cointegration models [at Ultra High Frequency] ?Removal of bid-ask bounce and other microstructure effects.

Quote[at Ultra High Frequency]Actually I was interested not only in UHF part.QuoteRemoval of bid-ask bounce and other microstructure effects.Some time ago I came across a paper about data preparation by applying Hodrick-Prescott filter but I am looking for alternatives to have a broader scope.

QuoteOriginally posted by: m4dkaQuoteRemoval of bid-ask bounce and other microstructure effects.Some time ago I came across a paper about data preparation by applying Hodrick-Prescott filter but I am looking for alternatives to have a broader scope.Geez, just take the average of the bid and the ask. Or the average of the depth-weighted bid and ask. Or the average of the predicted bid and ask, which you can predict from depth, book pressure, and anything except that other stock.Hodrick-Prescott is not designed for a known bid-ask. And you cannot "prepare" live data with it.