Forecasting the GDP
Jason M. Kraus
ECO6433 O. Mikhail
The gross domestic product, or “GDP”, is probably the most studied economic time
series for all times. GDP is defined as the total value of all goods and services produced within that territory during a specified period (or, if not specified, annually, so that "the USA GDP" is the USAs annual product). GDP differs from gross national product (GNP) in excluding inter-country income transfers, in effect attributing to a territory the product generated within it rather than the incomes received in it. The GDP may be specified as „real‟ or „nominal.‟ Whereas nominal GDP refers to the total amount of money spent on GDP, real GDP refers to an effort to correct this number for the effects of inflation in order to estimate the sum of the actual quantity of goods and services making up GDP. The former is sometimes called "money GDP," while the latter is termed "constant-price" or "inflation-corrected" GDP -- or "GDP in base-year prices", where the base year is the reference year of the index
The GDP is an interesting statistic to study for several reasons. First, it is the single best indicator of the general state of the economy, as well as the underlying trend of the economy, such as whether it is expanding, contracting, recessionary or inflationary. Additionally, the GDP can be used to forecast trends within sectors of the economy, future employment levels, housing starts (Davis and Heathcote, 2003), corporate profitability (Kim, Miller, and Ozanne, 2003). Further, the GDP can be used in monetary policy. If the growth in the GDP is deemed to be inflationary, the money supply can be cut back to reign in inflation, and vice versa. Policymakers depend heavily on GDP statistics and forecasts to decide which course to take with economic, fiscal and investment policy.
Current literature is quite diverse. A significant number of the papers reviewed deal with the overall decline in GDP volatility which has been observed since 1984, not only in the US, but throughout the G7 as well (Ramey and Vine, 2005), excluding Japan. The bulk of the material focuses on statistical issues, such as modeling, cyclical aspects of the GDP, memory within the trend. One of the main thrusts of the research is identifying the business cycle and its determinants. The reason for this is if the cyclical aspects of the GDP were able to be fully decomposed and the determinants explained, then those factors that contribute to the inflationary and recessionary cycles of the economy could be mitigated.
There are two purposes of this paper. First to discuss some of the relevant issues and literature regarding GDP. These issues include:
? The decline in GDP volatility
? How GDP fluctuations affect areas of investment, example: housing
? Cyclical aspects of the GDP
Second, quarterly data since 1947 will be analyzed and a model will be developed based upon this data. The selected model will be discussed and the results that are uncovered from that analysis will be discussed. An 20 period look-ahead forecast will be generated. The strengths and weaknesses of the forecast will be discussed as well. Finally, a few recommendations for further research will be discussed as well.
Discussion: GDP Volatility
The consensus in the literature is that the overall volatility of the GDP declined
sharply after 1984. Kim, Nelson and Piger (2001) had four important findings
regarding this volatility:
1. The reduction in aggregate real GDP volatility is seen in the cyclical but not
the trend component.
2. The reduction in volatility is not confined to any one sector of the economy.
3. The reduction in volatility is seen in final sales as well as production.
4. The dynamics of inflation display structural breaks in persistence and
conditional volatility over a similar time frame as the reduction in volatility on
Irvine (2004) asserts that the reduction in sales persistence is attributed to two main
1. Improved corporate management – better inventory to sales ratios.
2. Improved monetary policy through interest rates creates stability in sales such
as the auto industry or in housing.
This phenomena has been observed throughout the G-7 countries, excluding Japan
and for varying periods and at varying levels (Ramey and Vine, 2005). This is an
important shift that remains as the focus of a great deal of research. It is important to
determine if this is truly a permanent change or if it is attributable to:
1. Good luck.
2. Good policy.
3. Structural change.
Discussion: GDP, Housing and the Business Cycle
Davis and Heathcote (2003) studied housing and the business cycle, and worked to
develop a multi-sector stochastic growth model designed to help understand the
dynamics of residential investment, which was based on the following facts:
1. Different sectors of the economy tend to move together, particularly asset
investment such as residential structures and business capital.
2. Residential investment is more than twice as volatile as business investment.
3. Residential investment tends to lead the business cycle while business
investment tends to lag.
These three facts present a challenge to modeling and the authors were able to
create a calibrated model that accounts for the first two of the three conditions.
In the end, the analysis of the model that was developed points to two variables that
contribute to the volatility of the housing market. First is the overall variability of the
construction sector, which is viewed as being highly volatile in itself. Second, the low
rate of depreciation of housing structures causes increased demand for new
structures even in periods of high relative productivity. The main failing of the model
Davis and Heathcote developed is the inability to reproduce the fact that housing
development leads the business cycle, which can be represented as corr(RESI, t-1
GDP) > corr(RESI,GDP). The model which best represented the data was not the ttt
lagged data, but rather the contemporaneous data, however the lead model does fit
the data better than the lag model. Housing is an important component of the GDP
and understanding its behavior is important to being able to accurately forecast GDP
behavior in the future.
Discussion: Cyclical Aspects of the GDP
thSince close to the beginning of the 20 century, economic statistics, such as GDP,
have been collected and analyzed. The underlying reason for the mountains of
research which have been generated in this effort is to understand the business cycle,
that is, the cyclical expansion and contraction of the economy. The purpose and
hope of that effort is to be able to predict and mitigate the effects of the business
cycle, hopefully managing the economy into sustained, non-inflationary growth.
Currently, economists are forced to rely on real time economic data in order to
identify turning points within the economy as they happen, rather than accurately
forecast these changes.
Chauvet and Piger (2002) found that the qualities that characterize the turning points
from positive/negative deviations from the trend are quite different from each other:
1. Knowledge of which regime (expansion/contraction) the economy is in can
affect the quality of the forecast of future economic activity.
2. The relationship between economic variables in expanding economies is
different than those in contracting economies. For example, the relationship
between initial claims for unemployment claims and employment growth is
stronger in recessions than in expansive periods.
3. Evidence indicates that production output gains during expansion tend to be
permanent while negative deviations during recessions tend to be temporary. These are examples of guideposts which can be used to improve the quality
(closeness of fit) of GDP forecasts. As previously stated, having a better model
should lead to better policymaking, which should, in turn, mitigate the positive and
negative deviations from the trend, and maintain the proper direction and magnitude
of growth in the US GDP.
Quarterly data from 1947:01 to 2005:01 was analyzed for this paper. The data was
provided by the Bureau of Economic Analysis (BEA) website as measured in billions
of chained year 2000 dollars. Below are charted the raw GDP historical data and the
log of this data:
GDP: History (1947.01-2005.01) log(GDP): History (1947.01-2005.01)
0 GDP History (billions chained 2000 dollars)7.0log(GDP) History (billions chained 2000 dollars)505560657075808590950005505560657075808590950005
Figures 1 and 2: Historical GDP and log (Historical GDP)
Immediately evident from visual inspection of these two figures are two points. First,
the GDP series appears to be an exponential one, evidenced by the near linear
nature of the log (GDP). Second, the reduction in volatility within the GDP series,
previously discussed, is especially evident in the log(GDP) data from the mid-1980‟s
on. Selection of the best model is the next task.
Analysis: Model Selection
For the given series, initially, the linear, quadratic and log linear models were tested
against the data using Eviews 3.1. The AIC results of the regression analysis point to
a clear leader: the exponential or log linear model. The table below summarizes the
AIC information from the regression analysis:
Table 1: Linear, Quadratic and Exponential AIC Regression Analysis
This analysis backs up the visual inspection and indicates the exponential model is
the model that should be pursued.
Next, the data was tested for quarterly seasonality. The regression analysis, for
seasonality indicated that seasonality did not play a large role, and should not be
pursued in selecting a model. The AIC results of the regression analysis are listed
GDP + Seasonal
Table 2: GDP + Seasonality AIC Data As the table above indicates, adding quarterly seasonality to the analysis does not
improve the model fit. The simple exponential model is slightly better.
Next, the data was tested to see if the lagged data (Y or Y) provided any better fit. t-1t-2
As the results below show, the lagged data did not fit as well as the exponential. The
exponential regression could not be performed on the lagged data because the result
of some of the Y - Y and Y- Y are negative. tt-1t t-2
GDP + Lagged Yt-1
GDP + Lagged Yt-2
Table 3: Lagged GDP AIC Data
The next step was to test to see if the time series had any autoregressive or moving
average components to it. An extensive series of regressions were run against the
log(GDP) statistic, since that model had the best AIC results, and the AIC results for
each combination of ARMA were tabulated. The results of this analysis are tabulated
Table 4: ARMA Analysis Results
The data above shows that the log(GDP) ARMA(4,4) data is best fit. However, the
AR (4) regression AIC is very close to the ARMA (4,4) AIC. Since purely
autoregressive models are easier to study and manipulate than models that have a
moving average component, and the „goodness of fit‟ doesn‟t seem to be affected by
selecting the AR(4) model, the AR(4) model will become the chosen model for this
series. Although the previous regression output indicated that the exponential form
was the best fit model, the above analysis demonstrates the AR (4) on the log(GDP)
data is the best fit. The charts below illustrate this:
Log(GDP) AR (4): History (1947.01-2005.01)
Figure 3: Actual, Fitted, Residual for log (GDP) AR (4)
Figure 4: Actual, Fitted, Residual for log (GDP)
Clearly the, as the above charts demonstrate, the AR (4) is a better fit than the log-
Analysis: 20 Period Ahead Forecast
Using the exponential form, the GDP was forecast 8 periods ahead. The equation
generated by the regression analysis is as follows:
The graphical representation of this equation, is given in figure 5 below: