generation of a time series

[Paul Mueller]

Hi everybody,

I need to generate a time series from a set of micro-level spell data. My original data are recorded on a daily basis which means I could actually generate a series of daily values. However I think that would be too fine-grained, because I analyse a period of at least 10 years and my data would probably become too big. On the other hand the general convention to use an annual time series doesn't seem reasonable either simply because my data allow for a much more detailed analysis.

With that said my question is: How can I justify the frequency of my time series granted that a higher frequency doesn't imply the effort of additional measurements? And what are the implications for criteria such as significance and accuracy of measurement?

Thanks in advance!

[Andrew Musau]

Hi

The choice of frequency of your time series depends on what you are studying - and to what extent frequency affects this. In many macroeconomic applications, analyses are performed using data of higher frequencies (daily, weekly, monthly, etc.). For example, in Granger causality tests, annual data is sometimes not be suitable since a year is too long to assume that a change in one variable may influence other variables in later years. Using monthly data, the assumption is that changes in one variable only influence other variables in later months. If the true transmission (causality) mechanism is monthly (or weekly/ daily) - then using annual data will not capture this - and therefore in this case causality is suppressed in the data.

Now, there may be cases where you can justify using annual data - but justifying not using the data on the basis of your dataset being too big may not be a strong argument since modern software packages are designed to handle such data - and daily data for 10 years is 365 x10 = 3650 observations - not too big for a macroeconomic dataset! (The dataset that I am working on now has 6800 observations). As a robustness check, you may want to run the same analysis using different frequencies (e.g. both monthly and weekly). If you can show that the results do not change much depending on the frequency that you choose (except possibly in the case of annual data), then you make a stronger argument.



.

[Nick Cox]

sinceyouasked: First of all as a matter of list etiquette please note our strong preference for full real names here. This is explained in the FAQ Advice.

I don't see how we can answer your question independently of what you intend to do with the time series.

Incidentally, I don't see that you have specified that you are using economic data. Andrew Musau's reply seems to assume that, although many of his points apply generally.

[Paul Mueller]

Hi,

I just applied for a change of my user name.

For the rest: You already helped me a lot. The fact that the frequency of my time series should depend on the nature of the variables seems simple but I actually haven't thought of it before.