Time series temporal disaggregation (or: going from low frequency to higher frequency)

Brian

Sorry for not answering your questions sooner.  I appologise if this reply
is a bit off topic for this mailing list.

I have attached two pdf's which I hope explain in some way how we used the
Chow-Lin interpoalation/distribution  methods.  In the Central Bank of
Ireland we were interested  in modelling various aspects of the Irish
Economy but quarterly national accounts were not available in Ireland until
the late 90's.  In the early 80's I wrote  Chow-Lin routines first in TROLL
and then in Gauss and calculated several sets of quarterly national accounts
which were used in modelling various aspects of the economy.

Manual.pdf describes a RATS program used to produce a set of national
accounts for the Irish macro model component of the ECB systrm of
macro-models.  The implementation assumes a certain relationship between the
annual variable and the quarterly indicators.  If the model is valid then
the estimates are unbiased.  As the model is probably not valid some bias
certainly exists.  However analysis using the derived data has generally
produced reasonable useful results.

The methodology used differs from the original Chow-Lin methodology.  Given
the assumed model between the unobserved quarterly model and the indicators
one can calculate the distribution of the annual data and estimate the
parameters using maximum likelihood.

Manual2.pdf describes an extension of the methodology where quarterly data
are available for some of the period and the likelhodd estimation is based
on the distribution of the quarterly data where available and on the annual
observations otherwise.  I think that tis method has been used in the
Central Bank but I do not know the extent of that use as I retired from the
Central Bank about 5 years ago.

The method can be used to decompose a series by using a constant and/or
trend as indicators.  In most cases of interest there is some indicator.

In some cases one may set up some form of penalty function to be minimised
and assume that the quarterly series follows some form of time series.

A long time ago we also experimented a little with Kalman Filters with
limited sucess. These methods might be easier with current computer
facilities.

I

2009/10/30 Brian G. Peterson <brian at braverock.com>