First, you may want to disaggregate both the numerator and denominator of your ratio because the ratio of annual data is not the sum or average of the ratio of quarterly data.

Second, you want quarterly indicators which are correlated with your annual data. You can first investigate this question by aggregating your quarterly indicators into annual and analysing the correlation, or fit in a regression, with your initial annual series. You can of course use more than one indicator.

Third on the choice of the particular technique between Chow-Lin and co, I would recommend Fernandez. This is for a practical reason: by construction, Fernandez will smooth the quarterly profile of the unexplained component the most. Therefore, the unexplained component of your disaggregation model will be the least likely to influence your econometrics later on.

Fourth, note that there are other techniques which can be richer in modelling the relationship between your initial annual time series and the indicators. Looking into state-space techniques is probably the most accessible ones if you work with eviews. For reference, look at Proetti's paper http://ec.europa.eu/eurostat/documents/3888793/5816173/KS_AN-03-035-EN.PDF/21c4417c-dbec-45ec-b440-fe8bf95661b7?version=1.0

Not a fan of any technique. You effectively assume away important quarterly effects by these techniques. Better to try to build the same data from the ground up, in case important short-run effects are there. After all, going quarterly explicitly says you think three month changes are interesting. They aren't if you smooth them.

@Peter: There is a misunderstanding here, Chow-Lin and the like actually do what you invite Abhishek to do, i.e. "try to build the same data from the ground up".

Only when you don't find any short term indicatorto help you do it are you forced to smooth your data, i.e. "assume away important quarterly effects".

Hi Aurelien,

This is an interesting and important topic that goes well beyond economics and finance, so thank you for the input!

No misunderstanding. I'm aware that there are several different methods to get the proxies; I don't agree that they are good. As a for instance, see https://journal.r-project.org/archive/2013-2/sax-steiner.pdf and notice that on page 85 there are several different approaches compared to actual data. None mimic the reality perfectly and they are all a little different.

Two takeaways: the proxies are dependent upon methodology; and the proxies are not a perfect substitute for the actual data, with occasional serious differences.

I know these proxy methods are used extensively in time series models and I would use them myself if I had to. But I would hold my nose. Moreover, I would feel uncomfortable with the final results.

Imagine if the data were taken once every 5 years, then annual data would be determined by these same methods. We lose something by doing this.

On the other hand, I'm also skeptical of extrapolating beyond my data when using a simple first-order difference equation like xt = xt-1 + random error. So I'm already uncomfortable believing that the extrapolation xt+1 = xt + random error is reasonable.

To me, time series data are difficult and treacherous and a lot of the manipulations we do with those data is questionable.