I'm currently involved with a project to integrate RiskMetrics into our business and one issue we've identified is the treatment of market data timing across time zones. This can have the effect of lowering correlations between securities we can otherwise observe being highly correlated. We've tried a couple of approaches to deal with the issue but none have been satisfactory in our opinion. Just wondering if anyone else has dealt with this issue and what solutions have been proposed. We have gotten to the point of considering uploading our own data series, with all the complications that come with that.
$\begingroup$
$\endgroup$
3
-
1$\begingroup$ The two approaches are to (1) store all timestamps in UTC, or (2) to store all timestamps in the local time of that exchange and to maintain a database of UTC offsets for exchange/date pairs. $\endgroup$– chrisaycockSep 28, 2013 at 16:04
-
1$\begingroup$ Be sure to look at this answer: quant.stackexchange.com/questions/7650/data-synchronization $\endgroup$– vanguard2kSep 30, 2013 at 7:39
-
$\begingroup$ Thank you for your answer. Unfortunately I'm not sure if I can give precise examples since that might be disclosing proprietary information but it definitely is an issue that we've been struggling with, especially when there are proxies that could be used to estimate relationships (thinking NX Nikkei futures and S&P 500). $\endgroup$– Matthew McMasterOct 3, 2013 at 1:52
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
as vanguard2k points out the prolem is dealt with e.g. in Scaling portfolio volatility and calculating risk contributions in the presence of serial cross-correlations and references therein.
It turns out that correlations are lowered while lag one cross-correlations increase. E.g. you can probably see a correlation of Japan today to US yesterday due to the different closing times. The approach desribed above tries to incorporate this effect in volatility estmates for holding periods of more than one day.
This approach usually yields similar estimates as using weekly data (where closing-time differences still play a role but a much smaller one). But in my mind using daily data (and handling the effects described above) is superior in ex-ante risk analysis and portfolio optimization (more reactive, more data points).
In my experience you can not solve this problem by shifting data by full days. Shifting by full days however can be necessary treating funds whose pubslished NAV corresponds to a market date in the past (T-1 e.g).
-
$\begingroup$ Thank you Richard for the article, I'll definitely read through. We've been trying a couple of approaches within the constraints of the RM system. $\endgroup$ Oct 3, 2013 at 1:54
-
$\begingroup$ Does the system use a covariance matrix as input? If this is the case the the Newey-West estimator can help. By the way: the S&P versus Nikkei situation is the prime example for doing all this (often used when a lead-lag effect is explained). $\endgroup$– Richi WaOct 3, 2013 at 16:04
-
$\begingroup$ Yeah, we've experimented with NW, although RM does not support that in their system. We had been trying to do some type of multi-day return with overlaps as a possible fix (seems to work ok for Nikkei/S&P) but there's still some odd time differences between currency rates (most currencies taken at a different time than the Dollar index -- DX futures) $\endgroup$ Oct 4, 2013 at 19:58
-
$\begingroup$ @MatthewMcMaster I have seen quote sever problems with multi-day returns if it is done in a rollion fashion - I assume you don't do this. The problem with US and Japan should be well covered with NW in my opiinon. In the preprint you see a US/Europe dominated portfolio and how the usual covariance underestimates the risk coming from Asia. If you apply NW (or something similar) then the true risk contribution can bee seen. $\endgroup$– Richi WaOct 7, 2013 at 10:48