Forecasting GDP and employment

LRI Discussion Paper No.1 Series of 2016

C. J. Castillo
Labor Research Institute
Labor Education and Research Network

December 2016

All models are wrong but some are useful.

George Box

1 Introduction
Economic forecasting is a useful technique for both individuals and organizations.
The presence of uncertainties motivates both individuals and organizations to deter-
mine plausible scenarios in the economy so as to aid in decisions making. Forecast
of prices, for instance, may encourage households to buy more of a certain good now
when higher inflation is expected to set in the future. The same price forecast may
also change the supply decision of firms. Also, because employment is also affected
by changes in price level, it is practical for workers especially trade unions to account
for expected inflation during collective bargaining negotiations. It is during making
decisions that are forward-looking that economic forecasting plays a very important
role. Despite the apparent usefulness of economic forecasting, the labor movement
in the Philippines does not seem to be keen on incorporating this technique in
its methodologies. This is understandable considering that econometric modeling,
which serves as the primary method for economic forecasting, remains exclusive to
those in the economics and statistics profession particularly in the academe, gov-
ernment and private sector.
This paper proposes two simple models to generate forecasts of growth rates in
output and employment. These two variables are of utmost importance to workers.
GDP growth not only measures the total income of an economy but also gives an
indication of the trends in employment as well as wages for workers. Meanwhile,
workers have always been advocating for job generation to achieve full employment,

Draft not for circulation.

1
Discussion Paper Series Labor Research Institute

thus, it is always important to keep track of the changes in the employment growth
rate. These two indicators also helps workers in choosing the best courses of actions
for campaigns and even in collective bargaining.
We proceed as follows: section 2 presents the empirical model; section 3 discusses
the data used in the model; section 4 presents the forecasting results; and section 5
concludes.

2 Empirical strategy
2.1 Forecasting methods
There are mainly two types of forecasting techniquesqualitative and quantitative
methods. Qualitative methods are typically subjective in nature and commonly
involves making judgments about a certain scenario where empirical data may not
be available. Expert judgment is very crucial in qualitative forecasting. The Delphi
Method is one of the most widely known qualitative methods which employes a panel
of experts who will then serve as primary resource for forecasting (Montgomery et al,
2015). Quantitative methods, meanwhile, are those that involve the use of empirical
data collected over time called time series. This data set is analyzed through the use
of a forecasting model. There are several time series models used in forecasting. We
are concerned with two of the most common time series models used in forecasting:
(1) autoregressive model, and; (2) vector autoregression model.
Autoregressive models are univariate models that have the following general form

yt = + 1 yt1 + 2 yt2 + . . . + p ytp + t (1)

where t is white noise, i.e. uncorrelated with time with a constant variance and
mean zero. This model is said to be an AR(p) model where p is the order of the
model. In an AR(p) model, the value of a variable y at time t is expressed as a linear
combination of its past p observations, a constant and an error term (Adhikari and
Agrawal, ND)1 .
In practice, some exogenous variables are often included to improve univari-
ate models. These models called ARM AX are typically used in forecasting. One
problem, however, with single equation models is that they fail to capture contem-
poraneous relationship between variables in a multivariate single equation model.
1
A general class of time series models combines both autoregressive and another class of models
called moving average (MA). These models are called ARM A. ARM A models have the following
form:
p q
X X
yt = c + t + i yti + j tj
i=1 j=1

One noticeable difference between AR(p) and M A(q) models is that the former is regressed against
a variables past values while the latter is regressed against past error.

2
Discussion Paper Series Labor Research Institute

Suppose we have modeled a certain variable yt as a function of its lagged values

and another variable xt and we know that yt is affected by xt , a single equation
model such as this one disregards the possibility that xt may also be affected by
yt . The inadequacy of single equation models gave rise to models that use system
of equations and an example is the vector autoregression (VAR) model. A VAR is
a model composed of n variables and n equations where each variable is explained
by its own lagged values and the current and past values of other variables. VAR
models have become increasingly popular in the past decades and they have become
part of policy analysis toolkit of a lot of institutions.
A V AR(p) model is given by
p
X
Yt = + i Yti + t (2)
i=1

where Yt is the vector of endogenous variables, A is a square matrix of coefficients,

and Ut is the matrix of innovations that are white noise.
VARs are commonly used for both assessing the impact of changes in a certain
variable on another variable as in policy evaluation and in generating forecasts.
Structural VARs are usually used in policy evaluation. These models incorporate
structural relationships among variables that represent how these variables react to
changes in other variables based on economic theory. Because of this, modelers are
able to trace the sources of fluctuations in variables which make structural VAR
very useful in evaluating economic scenarios and validating economic theories. For
forecasting purposes, however, reduced-form VARs, such as the model presented
above, are enough.
In the next section, we discuss the models, both univariate and multivariate,
that we use in forecasting.

2.2 Forecasting GDP

A VAR model is used to forecast gross domestic product. The reduced form VAR
is as follows:
Yt = c + AYt1 + Ut (3)
where Yt is the vector of endogenous variables composed of GDP , government spend-
ing, investment spending, CP I and P SEI; c is the vector or constants; A is a 5 5
matrix of coefficients; and Ut is the vector of uncorrelated white noise disturbances.
A parsimonius model is preferred in forecasting (Allen and Fildes, 2002)2 . The
choice of the variables in the VAR reflects simplicity and at the same time, represen-
tation of major economic activities. The variable government spending represents
fiscal policy. PSEI represents business sentiments while CPI represents general price
levels as a result of transactions in the market.
2
Allen, P. and Fildes, R. 2002. Econometric Forecasting. In Armstrong, J.S. (ed) Principles of
Forecasting: A Handbook for Researchers and Practitioners.

3
Discussion Paper Series Labor Research Institute

2.3 Forecasting employment

The following autoregressive model (AR(4)) model is used to forecast employment:
4
X
Et = 0 + i Eti + t (4)
i=1
where Et is employment at time t and t is the error term.
Both these models are estimated with varying configurations and from which
respective forecasts are generated for comparison. The models that have the best
forecasts based on the appropriate forecast evaluation tool is used to generate fore-
casts.

3 Data
Quarterly GDP government spending data from 1981 to 2015 from the Philippine
Statistics Authority are used in forecasting output. These series were seasonally
adjusted and checked for stationarity and presence of unit root using augmented
dickeyfuller test. Stock index data from 1981 to 2015 are obtained from the Philip-
pine Stock Exchange. The data set is originally expressed as monthly figures which
were then converted into quarterly averages. Consumer price index data came from
the Bangko Sentral ng Pilipinas. CPI data were also expressed as monthly figures
and were also converted into quarterly averages.
Employment data from 1991 to 2013 came from the Current Labor Statistics.
The series is de-seasonalized and checked for stationarity. It is important to note
that the employment time series has break in 2014 because the disaster that struck
the Philippines in 2013 caused data collection problems. This made data prior
to 2014 incomparable with the preceding data. Thus, we use an expanding-window
forecasting technique where we estimate the model using data until 2013 and forecast
one-step ahead, then reestimate the same model now treating the forecasts as actual
data. We carry out this process until we come up with 20162017 forecasts.
Both models are estimated using Eviews 9.

4 Forecasts
Table 1 shows the forecast of GDP growth for 2016 and 2017.3 We obtained the
quarterly forecasts and re-seasonalized them using the seasonal factors in 2015. The
reported GDP growth forecasts are the conservative projections constituting the
lower hemisphere of 95 percent confidence interval constructed for the forecasts.
This is done to provide a narrower forecast band which is more useful in practical
sense. Forecasts show that the economy will maintain its growth in 2016 and 2017,
although the economy is expected to grow slower in 2017.
3
Confidence band widens overtime because of increasing degree of uncertainty.

4
Discussion Paper Series Labor Research Institute

Table 1: GDP growth forecast, 20162017

Period Low Average High

2016 Q1 6.06 6.84 7.62
2016 Q2 5.50 6.55 7.61
2016 Q3 5.67 6.96 8.25
2016 Q4 4.80 6.40 8.00
2017 Q1 5.28 6.37 7.47
2017 Q2 4.82 5.89 6.95
2017 Q3 4.59 5.62 6.66
2017 Q4 4.83 5.86 6.89
Source: Authors estimates.

Table 2: Employment growth forecast, 20162017

Period Low Average High

2016 Q1 (1.6) 0.4 2.4
2016 Q2 (0.8) 1.8 4.4
2016 Q3 (0.7) 2.5 5.6
2016 Q4 (0.8) 2.8 6.5
2017 Q1 0.8 2.5 4.2
2017 Q2 1.02 2.3 3.6
2017 Q3 1.1 2.0 3.0
2017 Q4 1.1 1.8 2.5
Source: Authors estimates.

Table 2 presents the forecast of employment growth. A growing economy is

expected to produce employment. In the past, employment growth tends to be half
of GDP growth. Forecast shows that there will be growth in employment along
with GDP, however, there is also possibility of employment contraction as shown by
negative figures for 2016. This will be offset by positive employment growth in the
following year.
Growing economy and employment are good for the labor movement provided
that employment grows faster than the labor movement, otherwise, unemployment
may worsen.

5 Conclusion
In this paper, we developed forecasting models for forecasting GDP growth and
employment growth. We used a VAR model with 5 variables to forecast GDP

5
Discussion Paper Series Labor Research Institute

growth while a autoregressive model is used to forecast employment. We used

deseasonalized quarterly data in generating the forecast and forecast bands. We
used the seasonal factors obtained in seasonal adjustment procedure to generate
final forecasts.
In this exercise, we find the GDP and employment are both expected to grow
in the short run. This should be considered positive development for the labor
movement and the government must support this trend with the appropriate fiscal
policy.
We encourage the labor movement especially the trade unions to incorporate
the forecasts generated by the model we developed into their analyses, education
courses and use these figures during dialogues with the government and employers.

6 References