Page 369 - Special Topic Session (STS) - Volume 3
P. 369
STS550 Kyle Hood et al.
specifications) using a set of model-averaging algorithms. We have two goals:
The first is to identify specific methods that can be used to accurately impute
individual detailed PCE services components for the advance GDP estimate,
while the second is to look for patterns in the revisions implied by the different
classes of model-averaging techniques. These patterns provide useful
evidence for the type of model-averaging techniques that are appropriate in
similar situations.
The two nowcasting models that we use differ not so much in form, but in
the type of information used. In the GB model, quarterly indicators are derived
from monthly source data. However, instead of the current method which uses
the growth rate of the quarterly indicator as the estimated growth rate, we
allow for a long-term trend, lags of the dependent variable, and lags of the
indicator. The second model that we use, BF, discards the indicator
constructed in this way, using rather a common factor derived from all
indicators from the GB model. This collection of indicators yields two factors
that appear to accurately capture the movement of many of the detailed PCE
services component series.
The reason that model averaging is chosen to combine this information
rather than using the more traditional technique of augmenting the model
with additional data is that in this application we have only about 34 time
periods. The short time-series dimension that we are working with is not
compatible with a large number of right-hand-side variables. Model-
averaging techniques are designed in part to combat this issue.
We estimate and average 12 versions (6 GB and 6 BF) of the models which
differ by how many lags of the dependent and independent variables are
included. Five model-averaging techniques are then considered: Two simple
techniques (equally-weighted average and median), two information-criteria-
based (IC-based) averaging techniques, and Bates-Granger (BG) averaging
with leave-one-out cross-validation (LOO-CV). All models and averages are
computed on an estimation sample (sometimes called a “training sample”) and
compared using pseudo-out-of-sample data from the end of the period.
2. Methodology
This section contains a more detailed description of methods. We start by
discussing the GB and BF models, then discuss the model-averaging
techniques, and finally we detail which GB and BF specifications are averaged
using these techniques and describe other details of the algorithm.
Nowcasting models
We consider two types of nowcasting models. In the GB class of models,
one or more indicators, lags of these indicators, and lags of the dependent
variable are used to nowcast the target variable in a regression framework.
Indicators are typically available at a higher frequency than the target variable,
358 | I S I W S C 2 0 1 9