Brian Larsen wrote:
> On Mar 11, 1:27=EF=BF=BDpm, jameskuy...@[EMAIL PROTECTED]
wrote:
> > I've got a time series 807793 bins long, with missing data in all but
> > 48945 of those bins. Only 7392 of those bins have a non-zero event
> > count. Those bins have a total count of about 1 million events, which
> > tells you that events are highly clustered, at least at the time scale
> > of the bin size (5 minutes).
> >
> > I want to use autocorrelation analysis to investigate the clustering
> > of these events on longer time scales. The large amount of missing
> > data makes such analysis difficult, but the non-missing data is
> > clustered on time spans of 9 bins or so. Therefore, it seems to me
> > that with the right algorithm, it should be possible to estimate the
> > autocorrellation at lags of less than 9 bins. Does anyone know what
> > the right algorithm would be?
>
> Seems to me that this is an issue, I would use normal techniques on
> subsets of the data. There might be other ways but clusters of
> missing data are kinda like small data sets.
The individual clusters are too small to calculculate meaningful
autocorrelation values; I would need to know an appropriate way to
combine autocorrelation functions calculated from different sets of
varying lengths.
I've found an article <http://sankhya.isical.ac.in/search/
61a2/61a27036.pdf> which describes three estimators that can be used
for this purpose. I was hoping I could use code that had already been
written, but it should be pretty straightforward to write a program to
calculate those estimators.
|
