Page 352 - Special Topic Session (STS) - Volume 3
P. 352
STS550 Angelia L. Grant et al.
models is that consumption patterns tend to depend on those in recent
periods, consistent with habit-forming preferences.
The main advantage of autoregressive models is that they perform well at
modelling persistence. On the other hand, the disadvantage is that they do
not use any other information, such as relative price movements. Each model
is estimated using ordinary least squares regressions.
The autoregressive models do not account for changes in relative prices,
which can drive important shifts in the share of each consumption component.
As such, the next models considered are regressions on the relative price of
the consumption component. The main advantage of relative price
regressions is that they take into account information about relative price
shifts. A disadvantage is that they do not include dynamics in the form of past
consumption shares.
The almost ideal demand system (AIDS) of Deaton and Muellbauer (1980)
takes into account that the consumption of a particular good or service
depends not only on its own price, but also the relative prices of other goods
and services which may be either complements or substitutes. It also takes
into account an income effect, with each of the shares depending on total
consumption expenditure. The main advantage of the almost ideal demand
system is its strong theoretical grounding. But this strong theoretical
grounding may mean that the model may be too restricted to fit the data well.
In addition, the model has a large number of parameters.
3. Forecast Combination Approach
There are two methods used to construct the forecast combinations. The
first method uses equal weights. The simple combination approach is often
found to outperform other more sophisticated combination schemes. The
second method weights the forecasts using past forecast performance. More
specifically, each of the models for each of the household consumption
components is weighted using a four-quarter rolling weight of the inverse of
the sum of the squared forecast error. The four-quarter rolling weight strikes
a balance between having relatively stable weights and weights that quickly
adapt when there is a change in performance across models.
The approach of combining forecasts based on past forecast performance
accounts for changes in modelling performance. That is, it captures the
benefits of different modelling approaches and accounts for the fact that
certain models can improve or diminish in performance over particular time
periods and at different forecasting horizons.
4. Forecasting Results
This section reports the out-of-sample forecasting results for each of the
models and for the forecast combinations. The out-of-sample forecasting is
341 | I S I W S C 2 0 1 9