Вы находитесь на странице: 1из 4

Methodology of the Monthly Index of Services

Interpolation

Interpolation (sometimes referred to as temporal disaggregation) is the process of producing a


time series at a higher frequency than is actually available, for example, a monthly series
from quarterly data. This is a common need across National Accounts, where the accounting
framework demands inputs that are not available as frequently as required.

In some cases, there may be an additional indicator series providing information about short-
term movements, and interpolation using such a series is known as benchmarking. If a
suitable indicator is not available, then an independent method of determining higher
frequency values needs to be used. It is possible to simply divide each data point by the
relevant number of time periods (e.g. divide each quarterly value by three to get three equal
monthly values) but this would not satisfy the primary condition of a good method of
interpolation - that the interpolated series moves in as smooth and continuous a way as
possible.

An important point to note is that interpolation will often be used in conjunction with some
form of forecasting/backcasting. There are two reasons behind this: 1) quarterly data are
supplied after the end of the quarter, and annual data after the end of the year. Thus, for
months between the end of the last quarter or year and the end of the next (or current) quarter
or year, no data are available; and 2) the interpolation method will need to be controlled at the
end of the series to minimise future revisions.

The interpolation function used in the IoS system is a robust method for interpolating a series
due to the reliable combination of the cubic spline and Holt-Winters forecasting. It produces
interpolated series that are smooth and without discontinuity and overcomes the end of series
problem, which affects the cubic spline when used on its own.

Interpolating the Index of Services

The current standard function for interpolation in ONS uses a cubic spline to interpolate a
smooth path through flows or index series. The process is quite complex and is explained in
detail in Annex C. This section gives a non-technical description of the procedure.

The function is called with five arguments (or parameters):

the first argument identifies the series to be interpolated. The function is run at the
periodicity of the output series, so this identifier requires a suffix specifying the
periodicity of the input series (annual or quarterly);
the second argument specifies the periodicity of the output series (this will be monthly for
the IoS);
the third argument specifies whether the interpolated series should sum (flows) or average
(index) to the original series;
the fourth argument is the required end date of the interpolated series, which can be
beyond the end of the original series due to the forecasting element of the function; and
the fifth argument is the type of Holt-Winters to be used in the forecast. This can be
additive, multiplicative or non-seasonal, to suit the nature of the data. For more
information about this, see the description of forecasting in Time Series Methods.

The method behind the function will depend on which optional arguments are introduced.
The primary process behind it, though, is the cubic spline. This is a very complex procedure,
but the basic concept is this:

it takes the quarterly (or annual) data and tries to run a smooth, continuous line through
all the points; and
this line is made up of polynomials of the third order (cubics). The coefficients of each
cubic are chosen such that the curve is as smooth as possible whilst still passing through
the quarterly data.

Aside from the selection of the optional arguments, there are some other points that are
usually considered when interpolating, which are presented here to help the reader understand
the process better:

Type of data

The method should only be used on flows and index series (like the IoS), not on stocks.

Forecasting

When using the spline function, it is always preferable to invoke a forecast of the original
series, even if estimates of future values are not required. This is because the cubic spline
would otherwise become unreliable at the end of the series due to lack of information (it has
past data points to use, but not future ones). Forecasting helps to control the cubic spline at
the end of the series and minimises revisions (see Figure 1).

Figure 1: Forecasting before interpolation


1.6
1.5
1.4
1.3
1.2
1.1
1
0.9
0.8
0.7
Jul-98 Jan-99 Jul-99 Jan-00 Jul-00 Jan-01

Quarterly data Without forecast With forecast


Level shifts (trend breaks)

Where there are sudden changes in the level of the series that have a known cause, the
interpolated series produced by fitting a smoothed curve may sometimes be inappropriate.
The cubic spline may overcompensate, producing an interpolated series with greater
curvature in the time periods either side of the break in order to ensure that the two segments
fit together at the break. This will lead to distortions in the interpolated series as the effect of
the break is spread across adjoining periods.

If such a trend break is distorting the interpolated series, a possible solution is to use an ad-
hoc temporary adjustment (as is sometimes the case in seasonal adjustment). Figure 2 shows
the effect a level shift can have on the path of the spline. Here the interpolated series falls
between the third and fourth quarters and rises between the fifth and sixth quarters, owing to
the distorting effect of the level shift occurring between the fourth and fifth quarters.

Figure 2: Interpolation through a trend break


250

200

150

100

50

0
Jan-95 Jul-95 Jan-96 Jul-96

Quarterly data Interpolated data

Figure 3: Interpolation through an outlier


145
140
135
130
125
120
115
110
105
Jan-99 Jul-99 Jan-00 Jul-00

Quarterly data Interpolated data


Extreme values (outliers)

Extreme values have a similar distorting effect to level shifts and should be treated in a
similar way, i.e. by making temporary ad-hoc adjustments. Figure 3 above shows a series
containing an outlier and the effect it can have on the interpolated series (the series is
distorted in periods to either side of the outlier). For information about how these distortions
(and those caused by level shifts) are dealt with in the IoS, see the section on Quality
Assurance.

Seasonality

The cubic spline can easily deal with seasonal data and has no problems in creating an
acceptable seasonal path for the higher frequency series. It is particularly important to
forecast a series prior to interpolation if it is seasonal.

Revisions

Each time the original series is updated with a new data point, it will cause revisions to the
interpolates in the previous 3 time periods.

Вам также может понравиться