Page 353 - Special Topic Session (STS) - Volume 3

STS550 Angelia L. Grant et al.
based on 12-step ahead forecasts, which are computed every 2 quarters. The
forecast evaluation period is from 2006Q1 to 2016Q1. In assessing the forecast
performance, the benchmark model is the almost ideal demand system given
that it is the model currently used for forecasting the household final
3
consumption components.

Individual Models
Five models are estimated for each individual component of household
final consumption expenditure. They are the standard AR(2) models, AR(2)
models with linear time trends, relative price models, relative price models
with linear time trends and the almost ideal demand system (AIDS). The lag
length for the autoregressive models is chosen with a view to modelling the
persistence in the data, while maintaining a parsimonious specification. The
estimated AIDS does not include total consumption expenditure to ensure
that the model is not over-parameterised.
Table 1 reports the root mean squared forecast error (RMSFE) relative to
that of the almost ideal demand system for each of the models over different
forecast horizons. At the one-quarter-ahead forecasting horizon, both sets of
AR(2) models significantly outperform the almost ideal demand system for all
household consumption components. These models perform particularly well
for the components of food, cigarettes and tobacco, durables, other goods
and other services. For example, the RMSFE for other goods under the
standard AR(2) model is only 17 per cent of that of the almost ideal demand
system. The forecasting gains are smaller for the components of fuels and
lubricants and electricity and gas, but they continue to be better than the
benchmark.
In the case of the relative price models, the model with the linear time
trend generally performs much better than the model without the trend.
Further, even for the components where the model with the linear time trend
does not outperform the almost ideal demand system – alcohol and fuels and
lubricants – the performance is not substantially worse than AIDS. The relative
prices model with the linear time trend does particularly well at forecasting
other services, with the RMSFE being only 17 per cent of that of the almost
ideal demand system. The RMSFE for other services under the relative prices
model without a time trend is 57 per cent of that of the almost ideal demand
system.

3 Formal statistical tests could be performed to assess the statistical significance of the results.
However, given the relatively short evaluation period it would be difficult to obtain conclusive
results.

342 | I S I W S C 2 0 1 9