Statistical Modeling

AC/UNU Millennium Project Futures Research Methodology V2.
0
Statistical Modeling 1
10.

STATISTICAL MODELING:
FROM TIME SERIES TO SIMULATION

By
The Futures Group International
1

I. History of the Method
II. Description of the Method and How To Do It
Time-Series Analysis
Curve Fitting
Averaging Methods
Exponential Smoothing
Explanatory or Casual Analysis
Multiple-Regression Analysis
Forward-Stepwise Regression Analysis
Polynomial-Regression Analysis
Autocorrelation Analysis
Simultaneous Equations
Input-Output
Simulation Modeling
III. Strengths and Weaknesses of the Method
Major Strengths of Regression
Major Limitations of Regression
IV. Frontiers of the Method
V. Who Is Doing It
Bibliography

1 The Futures Group International, http:/futuresgroup.com
AC/UNU Millennium Project Futures Research Methodology V2.0
Acknowledgment

The managing editor wish to thank the reviewers of this paper who made many important
suggestions and contributions: Peter Bishop of the University of Houston, Graham Molitor
president of Public Policy Forecasting Inc. and vice-president of the World Future Society,
Stephen Sapirie of the World Health Organization. And finally, special thanks to Elizabeth
Florescu and Neda Zawahri for project support, Barry Bluestein for research and computer
operations, and Sheila Harty for editing. Thanks to all for your contributions.

I. HISTORY OF THE METHOD

Statistical modeling assumes that information contained in historical data can be extracted,
analyzed, and reduced to one or more equations that can be used to replicate historical patterns.
The techniques use the mathematics of statistics to deepen understanding of causality in complex
systems.

An equation or set of equations can replicate the history timeline of a variable, for example,
population. In this manner, when a prior date is used in an equation describing population, the
answer generated is very close to the actual population existing at that time. To forecast "future
population," simply substitute a future date in the same equation. This method, however, carries
some serious implications:

All information needed to forecast desired aspects of the future is contained in selective
historical data;

The model structure validly replicates the real life structure of the system that gave rise to
the historical data;

The ongoing structure of the system that gave rise to the historical data will be
unchanging; and

Deterministic (that is, single-point) forecasts are useful.

These implications clearly cause concern. New forecasting methods have been introduced to
circumvent some of these issues (see Trend Impact Analysis). Statistical forecasting methods are
exceedingly useful and important to futures research. They can deepen understanding of the
forces that shaped history. They can also provide a surprise-free baseline forecast (for example,
"suppose things keep going as they have in the past . . .") for thinking about what could change
in the future.

The methods of statistical modeling include
2
:

time series analysis;
regression analysis;
multi-equation models (for example, in econometrics); and
simulation modeling.

Time-series analysis refers to the mathematical methods used to fit trend data. These methods
can be simple or complex. The simpler ones involve plotting a curve through historical data in a
way that minimizes error between the curve and the data. The plot can be a straight line or a

2
Theodore J. Gordon, "The Methods of Futures Research," The Annals of the American Academy, July
1992.
curved line. Statistical processes are used to plot the line through the data points. If the fit is
good, the plot can be extended into the future to produce a forecast.

The value of some variables may depend on factors other than time. Population size, for
example, may be dependent on numerous variables, such as the number of young women in the
population a year ago, their education, or personal income. Models that relate a factor, such as
population, to other input variables, such as education or personal income, can be constructed
using a method known as regression analysis. In regression analysis, the subject under study
population is the "dependent variable," and the factors linked within the equations are
"independent variables." Regression equations can be linear and involve a few independent
variables or be nonlinear or polynomial and involve many variables. Regression equations can be
written for time series, or they may be "cross-sectional." That is, written to represent, the
relationship among independent variables at points in time.

Sometimes the dependent variable of one equation is used as an independent variable in another
equation. In this way, "simultaneous" equations are built to describe the operation of complex
systems (such as national economies) in econometrics. Jay Forrester pioneered this field. An
example of his work is described in World Dynamics.

In time-series analysis, regression analysis, and simultaneous-equation modeling, the equations
are determined by statistical relationships that existed in past times. By contrast, in simulation
modeling, the equations are constructed to duplicate, to a greater or lesser degree, the actual
functioning of the system under study. For example, a simulation model that attempts to
duplicate the historical size of population might involve the following logic: population today is
simply the number of people who existed last year, plus the number of people born and minus
the number of people who died during the year. Such an equation can be used as a forecasting
model.

Statistical modeling (that is, regression analysis) is usually a cross-sectional view of the
relationship among variables at a single point in time. The number of cases included in a sample
provide the variation to construct the parameter estimates. In contrast, simulation modeling is a
longitudinal analysis of the variables and their relationship over time. Thus, time is a primary
variable as is the effect that variables have on each other. The difference is largely between a
cross-sectional or a longitudinal analysis. The methods overlap when time is used as a variable
in the regression equation that leads to time-series analysis. The purpose of the two methods is
still different, however. The emphasis in statistical analysis is to estimate the regression
coefficients as indicators of the structure of a system; the emphasis in simulation analysis is to
use those coefficients in extrapolating the value of variables over time. Statistical techniques can
be used to arrive at the coefficients, but they are only a means to than end.

From a historical perspective, these modeling methods flow from the evolution of statistics.
Mathematics and statistics found increasing applications in the 1800s. During this time, several
statistical ideas and methods of analysis were developed by Francis Ysidro Edgeworth, Francis
Galton, Karl Pearson, George Udny Yule, and other British scientists and mathematicians. Sir
Francis Galton invented the correlation coefficient about 1875. Galton argued that plants and
animals vary according to patterns and developed a number of statistical methods to study these
patterns. Galton further applied these techniques to the study of heredity, which he was the first
to call eugenics.

The importance of cycles in trend data was noted by W. M. Persons in 1916. A major
development around 1925 involved techniques to identify cycle lengths rather than merely
assuming them. Statistics mostly remained unrefined until the 1920s, when it emerged as a
branch of science by a small group of English statisticians. An early use of multiple linear
regression occurred in psychology, where it was used to evaluate test results. During World War
II, this field of analysis was developed further as part of the war effort in Great Britain and the
United States.

As early as the 1600s, Sir William Petty, an English economist, recognized the value of applying
mathematics and statistics to economics. Not until the late 1800s did Leon Petty lay the foundation
for econometrics by writing a mathematical description of a market economy. Econometricians
Ragnar Frisch of Norway and Jan Tinbergen of the Netherlands were the first to be awarded a Nobel
Prize for their work in economics.

Basic input-output concepts were developed almost 200 years ago by Francois Quesnay. Leon
Walras refined the mathematics involved in this technique in the early 1900s. However, Wassily
Leontief published the first national input-output table in 1936 and inaugurated the modern use of
input-output analysis. He received a Nobel Prize in 1973 for his work in this field.

Large-scale simulation modeling could not be developed until computer technology progressed
so that large data sets could be easily handled. Some of the first applications of simulation
modeling occurred at the Rand Corporation in the late 1960s. In July 1970, members of the Club
of Rome attended a seminar at the Massachusetts Institute of Technology (MIT). During this
seminar, members of the Systems Dynamics Group expressed a belief that the system analysis
techniques developed by MIT's Jay W. Forrester and his associates could provide a new
perspective on the interlocking complex of costs and benefits as a result of population growth on
a finite planet. Research began under the direction of Dennis L. Meadows and was presented in
his book, Limits to Growth, followed by Toward Global Equilibrium: Collected Papers and The
Dynamics of Growth in a Finite World. More recent treatments include Donella Meadows'
Beyond the Limits: Confronting Global Collapse, Envisioning a Sustainable Future, the sequel
to Limits to Growth, and Groping in the Dark: The First Decade of Global Modeling, which
provides an excellent overview of global modeling. Some of the first applications of simulation
occurred at Rand in the mid-1970s. Simulation modeling is used extensively today in fields
ranging from astrophysics to social interaction.


II. DESCRIPTION OF THE METHOD AND HOW TO DO IT

Time-Series Analysis

Curve Fitting

In curve fitting, a computer software program examines historic time-series data for how well a
series of equations can fit or duplicate those data. The equations can be either linear or
nonlinear. The latter are represented by a Gompertz curve or a logistic function and are often
referred to as "S"-shaped curves. Sinusoidal variations of parameters also fall into the nonlinear
category.

When curve fitting is employed, the equation chosen for making the forecast or for projecting the
historic data is generally the one that displays the best fit or highest correlation coefficient
(usually using a "least-squares" criterion which refers to any methodology which evaluates the
relationship between a set of values and alternative estimators of those values by choosing that
estimator for which the sum of the squared differences between the actual values and the
estimated values is lowest) with the past data points. In certain cases, the practitioner may know
beforehand that the system or parameter with which s/he is dealing cannot exceed 100 percent of
the market. In such cases, only certain equations are selected a priori for use. Figure 1 presents
a set of curves typically used in curve fitting.

COMMON CURVE-FITTING ALGORITHMS

V = e
-(A+BX)
(exponential)
V = M Y + B (linear)
log V = M Y + B (log-log)
log V = M log Y + B
V = M log Y + B
1/V = M Y + B
1/V = M / Y + B
V = M /Y + B
log log V = M Y + B
log log V = M log Y + B
log V = M / Y + B
1/log V = M / Y + B
1/V = M log Y + B
1/log V = M log Y + B

Note: M = Slope
B = Additive Constant
Y = Year 1900
V = Calculated value

Most curve-fitting software programs allow the user to select a curve, but selection of the proper
general curve shape can be difficult. Two different curve shapes can, for example, each fit the
historical data well and yet produce markedly different extrapolations. In effect, selecting the
curve shape may predetermine the forecast. Analyses can also include identification of cyclic
components in the historical data (for example, "economic recessions") and, as shown in Figure
2, superimposed positioning of the cyclic components on the arrayed forecast.

Figure 1. History and Baseline Projection

Averaging Methods

Data on historical trends can be smoothed using several methods. The arithmetic mean can be
calculated and used as a forecast. However, this assumes that the time series is based on a fairly
constant underlying process. Simple moving averages, which are computed dropping the oldest
observation and including the newest, provide a method of dampening the influence of previous data
sets. Linear moving averages provide a more sophisticated averaging method better suited to
addressing trends that involve volatile change. In effect, linear moving averages are a double
moving average, that is, a moving average of a moving average.

Exponential Smoothing

Exponential smoothing refers to a class of methods in which the value of a time series at some
point in time is determined by past values of the time series. The importance of the past values
declines exponentially as they age. This method is similar to moving averages except that, with
exponential smoothing, past values have different weights and all past values contribute in some
way to the forecast.

Exponential smoothing methods are useful in short-term forecasting. They can often produce
good forecasts for one or two periods into the future. The advantage of exponential smoothing is
in its relatively simple application for quickly producing forecasts of a large number of variables.
For this reason, this method has found wide application in inventory forecasting. Exponential
smoothing should not be used for medium-or long-term forecasts. Such forecasts depend heavily
on the most recent data points and thus tend to perform well in the very short term and very
poorly in the long term.

Explanatory or Causal Analysis

Multiple-Regression Analysis

Multiple-regression analysis is used to determine the relationship between the dependent variable
and a number of independent variables. Frequently, this analysis is used to create an equation
that explains the behavior of the dependent variable. If the value of a variable-for example,
sales-were thought to be determined by three other variablesprice, advertising, and quality
multiple-regression analysis might be used to determine the nature of the relationship. It would
yield an equation like:

Sales = 100 0.2 x price + 0.1 x advertising + 0.6 x quality

Forecasting with multiple-regression analysis requires:

Sets of historical values for each of the predictor variables, and the dependent
variable for the same time period. Independently derived extrapolated or
forecasted values of the predictor variables for each time or occasion when a
value of the dependent variable is desired.

Many high-quality software statistical packages facilitate the regression process. Not only do
they produce high-quality graphs of actual and calculated data, but also they produce a wealth of
information regarding how well the regression forecast matches actual data.

Figure 2 illustrates an application of regression analysis. Here the problem was to derive an
equation for the number of employed persons in the United States (the dependent variable). The
independent variables were:

year (yr);
gross national product (GNP);
number of unemployed persons (UNEMP);
population (POP); and
GNP deflator (GNPDEF).

The equation that resulted from the analysis was:

EMPLOYED = .6202(yr) + .0403(GNP) .0056(UNEMP)
.2609(POP) .1288(GNPDEF) 1116.9

Figure 2. The Match Between a Regression Equation and Actual Data

Figure 2 graphs an example of multiple regression. In this example, the dependent variable is
"Employed," and the independent variables include year, GNP unemployed, and population. By
visually inspecting the two lines above, we can see that the defined relationship has yielded
results very close to the actual data. Another indication is that the R
2
is .9838, which means that
this equation explains 98 percent of the variance.

Often, in multiple regressions, the correlations among variables is useful to examine. The
correlation coefficient describes the degree to which the variation in one variable is related to
that in another. The correlation coefficient varies from +1 for a positive correlation to -1 for a
negative correlation. A value of 0 means no correlation exists. This method is often useful in
testing hypotheses about the relationships among variables. For example, if you thought that the
sales of a product were affected by the amount of advertising, you could calculate the correlation
coefficient. The result would tell you the degree to which a change in advertising is associated
with a change in sales. Figure 3 illustrates a typical correlation coefficient printout.


Variable

GNPDEF

GNP

ARFC

Pop.

GNPDEF
Probability =
Number =
GNP
Probability =
Number =
ARFC
Probability =
Number =
Pop
Probability =
Number =

0.992
1.000
16

0.465
0.930
16
0.446
0.917
16

0.979
1.000
16
0.991
1.000
16
0.364
0.835
16

0.971
1.000
16
0.984
1.000
16
0.457
0.925
16
0.960
1.000
16
Figure 3. Correlation Coefficient

Forward-Stepwise Regression Analysis

Forward-stepwise regression analysis uses all the same calculations as in multiple regression
except that it "steps" through the regression procedure by entering only the independent variables
that make a statistically significant contribution to the amount of variance explained. In the first
step, each independent variable is examined in a separate regression in order to choose the one
that accounts for the most variance. In each additional step, the remaining variables are
examined. The variable that contributes the most to explaining the remaining variance is
entered into the equation. The stepwise-regression method of analysis only selects variables that
meet the F-value or probability criteria entered by the user. The stepwise-regression method
continues until all selected variables are entered or until no remaining variables meet the
entrance criteria.

Polynomial-Regression Analysis

Polynomial-regression analysis is similar to multiple regression except that only one independent
variable exists. The independent variable, however, is used in the equation several times. If it is
used twice, the second term will be squared. If it is used three times, the third term will be
cubed. This type of analysis is most useful when a complicated nonlinear relationship between
two variables is suspected.

Autocorrelation Analysis

Autocorrelation accommodates searches for the existence of repeatable cycles. If data are
suspected of containing a repeating cycle, autocorrelation analysis can help to determine if such
a cycle is really there. Suppose, for example, data on monthly sales of some product are
available, and a seasonal influence is suspected. Autocorrelation analysis can help determine
this.

Autocorrelation analysis consists of correlating the data at a particular time with data at an earlier
time. A phase 1 autocorrelation correlates each data point with the data point one time period
earlier. Thus, the second point is correlated with the first point, the third with the second, the
fourth with the third, and so on. An autocorrelation of phase 2 will correlate a series with itself
with a lag of two periods. Thus, the third point is correlated with the first, the fourth with the
second, etc. An autocorrelation of phase 12 will correlate a series with itself lagged 12 periods.
Thus, when analyzing monthly data, a phase 12 autocorrelation would correlate all the January
points with each other, all the February points with each other, etc. If sales data really reflect a
seasonal influence in the sales data, then sales will tend to be high in the same months each year
and low in the same months. This pattern will show a high autocorrelation with a phase of 12. A
typical autocorrelation presentation is shown in Figure 4.

Figure 4. Autocorrelation Coefficients

Simultaneous Equations

Simultaneous equations are often used when a high degree of interdependence exists between the
time series to be forecast and other variables. Economics and other fields recognized that variables
that might otherwise be used in regression analysis were not independent. This interdependency
precluded the use of regression analysis.

Econometrics sometimes relies on simultaneous equations to forecast key economic variables, such
as GNP or gross domestic product (GDP) and consumer spending. An econometric model can be
composed of a number of simultaneous equations of different types and functional forms. Factors
that are most strongly influenced by one another are classified in an econometric model as
endogenous, while those that can be determined outside the system of simultaneous equations are
called exogenous. One equation must be specified for each endogenous variable.

Input-Output

An input-output table is a technique for determining transactions within and among different
sectors. Such a table could be constructed, for example, that summarizes the subtractions and
additions within all industries and between them and the consumer. If a table for the system
being studied is not available, a considerable amount of time and money is usually required to
collect the necessary data and construct the table. Once in hand, the table can be used to trace
changes that occur in all sectors as a result of changes in one sector. A computer is usually used
to trace how a change in the output of one sector affects the inputs required by other sectors,
affecting the outputs of still other sectors, and so forth.

Simulation Modeling

In simulation modeling, an attempt is made to duplicate the system being modeled in the form of
equations, not solely by drawing on statistical relationships among variables, but rather by logic
and inference about how the system works. For example, using regression to forecast the number
of people who live in Accra, one might find a very precise, but spurious, correlation with the
number of apples sold in New York. The statistics might appear excellent, but the logic flawed.
Furthermore, the coefficients produced in a regression analysis have no actual meaning. The
perspective changes in simulation modeling. This method starts with an idea about how the
system functions and gives meaning to the coefficients. The validity of the model is tested by
comparing its output with actual historical data.

For example, suppose we want to construct a model for use in forecasting employment. A time-
series analysis might take historical employment data, fit a curve, and extend the data into the
future. A regression analysis would relate employment to factors considered important, such as
overall population, unemployment, wages, exports, imports, etc. The difference between these
methods is valued for different purposes. The purpose of regression analysis is the structure of
variables based on their covariation. Future values can be estimated but rarely are (except
perhaps in screening candidates for school or employment). In simulation, that same structure is
an input to the real purpose, which is the estimation of future values and the behavior of those
values over time. Therefore, the distinction should be drawn at the purpose and not at the
mathematical level.

Simulation modeling is complicated but has the advantage of forcing attention on how things really
work. No prepackaged programs for simulation modeling are available for statistical analysis; each
model entails a custom job. However, some specialized modeling languages, such as Dynamo, were
developed for these applications.


III. STRENGTHS AND WEAKNESSES OF THE METHOD

Although time-series analysis is quick and easy, it provides little fundamental understanding of
future behavior. Since the future is predicated solely on the past without an underlying feel for
causal factors, time series is a naive forecasting method.

While various forms of explanatory or causal forecasting strive to explain fundamental causal
relationship, it too is predicated on past behavior and therefore also presents a naive forecast.

Major Strengths of Regression

The strength of regression as a forecasting method is that it capitalizes on historical relations
between the predicted (dependent) and predictor (independent) variable. It uses all the
information in the historical data pairs to determine the future values of the predicted variables.

The "goodness of fit" of the Y
c
to the historical Y values can be used to compute a measure of
the strength of the linear relationship between the historical X, Y pairs which then can be used to
calculate "confidence limits" or probable upper and lower bounds for the predicted values of Y.
In general, the closer the Y
c
to the historical Y values, the narrower the confidence limits; that is,
the actual values of Y are less likely to depart from the predicted values. The correlation
coefficient is an index that can be used to calculate a figure of merit from the accuracy with
which the calculated values of Y, Y
c
match the actual past-history data. The square of the
correlation coefficient ranges from 0 through 1. A value of 0 means total failure of the Y
c
values
to correspond with the corresponding Y values. A value of 1 for the square of the correlation
coefficient means that the Y and Y
c
values correspond exactly. Values between 0 and 1 may be
interpreted as expressing the proportion of variability in the historical Y values that could be
accounted for by the calculated linear relationship between X and Y.

Major Limitations of Regression

The method of least squares, as commonly used, implies that the predicted values of the
independent variable (X) are devoid of error or uncertainty; that is, the only possible error or
uncertainty is in values of the dependent variable (Y). Often this assumption is questionable.
For example, the predictor variable can be forecast incorrectly. Take a specific example:
suppose we want to forecast the prime interest rate and develop a good statistical equation
relating the Consumer Price Index (CPI) and the prime interest rate. A forecast of the future CPI
trend may then be used to generate a forecast of the prime interest rate, using bivariate linear
regression. Accuracy of this forecast depends on how strongly the past CPI values and prime
interest rate are related and on how accurately the future CPI trend is predicted. The latter source
of inaccuracy is not normally taken into account in calculating upper and lower bounds for the
forecasted values of the dependent variable or, more generally, in evaluating the accuracy of this
forecasting method.

When the past-history data are subject to error, the effect of the error makes the predicted values
of Y vary less than they should. Values of Y that should fall below the mean will generally be
forecast as such, but less so than they should be, similarly for values that should be above the
mean. The greater the "noise" in the past history, the greater this effect; and no way exists using
this method to distinguish a weak relationship between X and Y from a strong relationship that is
obscured by noise or error of measurement.

The methods. as commonly applied, assumes that all past-history data pairs are equally
important. While "weighted" data pairs can be used to generalize, the method is not common.

The method fundamentally generates a "regression forecast." The forecast of Y depends on a
prior forecast of X. Similarly, the forecast of X might depend upon a prior forecast of W. But
somewhere in this series a forecast must exist that does not depend on another forecast. One way
to break the chain is to have time itself as the predictor or independent variable. This option,
however, necessarily results in the predicted (dependent) variable, which depends only on time,
either increasing or decreasing without limit over time.

IV. FRONTIERS OF THE METHOD

Statistical methods described here rely heavily on the assumption that forces shaping history will
continue to do so. The frontiers of statistical methods must surely include techniques to test that
assumption and, where found wanting, permit the introduction of perceptions about change.
Trend Impact Analysis, Probabilistic System Dynamics, Cross Impact Analysis, Interax: all of
these methods attempt to meld judgment with statistics.

Staying strictly within the bounds of statistics, some research directions that could benefit the
field include:

Developing methods that test time series for chaos (see section Frontiers of Futures
Research). Unless special cases are involved, fitting chaotic time series using any of the
techniques suggested here will probably be unproductive.
Exploring new forms of regression equations that match the time series under study. For
example, using an S-shaped function when regressing variables that relate to
technological maturation or substitution.
Simulating policies in regression models. To date, "dummy variables" whose value is
either zero or one have been used in regression studies to indicate the presence or absence
of a policy. This method is not refined.
Using improved clustering or multidimension scaling techniques to improve the
efficiency of searching for variables that can fit in a group or relate to one another.
Including nonlinear relations in simulation models, which can sometimes result in
apparent chaotic behavior.


Several key periodicals track developments in this field:

Applied Economics
Futures
International Economic Review
International Journal of Forecasting
Journal of Applied Business Research
Journal of Business and Economic Statistics
Journal of Developing Arts
Journal of Econometrics
Journal of Forecasting
Journal of Macroeconomics
Journal of Policy Modeling
Oxford Bulletin of Economics and Statistics
Quarterly Journal of Business and Economics
Technological Forecasting and Social Change

V. WHO IS DOING IT

Fortunately quantitative analysis is facilitated by a large number of software programs that are
growing in sophistication. These programs have made more sophisticated quantitative analysis
widely available and relatively inexpensive to forecasters and planners. The most sophisticated
methods tend to be used in a university setting by consultants, mathematicians, and
econometricians who have devoted considerable time to the study of the subject.

Software packages are available from numerous sources, including:

SYSTAT Inc.
1800 Sherman Avenue
Evanston, IL 60201
(708) 864-5670, Fax (708) 492-3567

Abacus Concepts, Inc.
1984 Bonita Ave.
Berkeley, CA 94704
(510) 540-1949

Manuguistics
2115 East Jefferson Street
Rockville, MD 20852
(301) 984-5000, Fax (301) 984-5094

Smart Software, Inc.
(Charles Smart, President)
4 Hill Road
Belmont, MA 02178
(617) 489-2743 Fax (617) 489-2748

SPSS, Inc.
444 North Michigan Ave., Suite 3000
Chicago, IL 60611
(312) 329-2400, Fax (312) 329-3668

SAS Institute Inc.
SAS Campus Drive
Cary, NC 27513
(919)677-8000, Fax (919)677-8123
TSP International
P.O. Box 61015
Palo Alto, CA 94306
(415) 326-1927

The Futures Group, Inc.
80 Glastonbury Boulevard
Glastonbury, CT 06033-4409
(203) 633-3501, Fax (203) 657-9701

BIBLIOGRAPHY

Ayres, Robert U. "Envelope Curve Forecasting." In Technological Forecasting for Industry and
Government. Edited by James R. Bright. New Jersey: Prentice-Hall, Inc., 1968, pp. 77-94.

Ayres, Robert U. "Extrapolation of Trends." Technological Forecasting and Long-Range Planning.
New York: McGraw-Hill, 1969, pp. 94-117.

Benton, William K. "Time Series Forecasting." Forecasting for Management. Massachusetts:
Addison-Wesley, 1972, pp. 67-139.

Bright, James R. "Trend Extrapolation." A Brief Introduction to Technology Forecasting:
Concepts and Exercises. Austin, Texas: Pemaquid Press, 1972, pp. 6-1 to 6-70.

Congressional Research Service. "Trend Extrapolation." Long-Range Planning. Prepared for the
Subcommittee on the Environment and the Atmosphere. Washington, D.C.: U.S. Government
Printing Office, May 1976, pp. 434-443.

De Gooijer, Jan G., and Kuldeep, Jumar. "Some Recent Developments in Non-Linear Time Series
Modeling, Testing, and Forecasting." International Journal of Forecasting 8, October 1992,
pp. 135-156.

Engle, Robert F., and Hendry, David F. "Testing Superexogeneity and Invariance in Regression
Models." Journal of Econometrics 56, March 1993, pp. 119-139.

Fiebig, Denzil G.; McAleer, Michael; and Bartels, Robert. "Properties of Ordinary Least Squares
Estimators in Regression Models with Nonspherical Disturbances." Journal of Econometrics
54, October-December 1992, pp. 321-334.

Forrester, Jay W. World Dynamics. Cambridge, Massachusetts: Wright-Allen Press, 1971.

Lanford, H. W. "Analytical Techniques." Technological Forecasting Methodologies: A Synthesis.
New York: American Management Association, 1972, pp. 32-113.

Lenz, Ralph C., Jr. "Forecasts of Exploding Technologies by Trend Extrapolation." In
Technological Forecasting for Industry and Government. Edited by James R. Bright. New
Jersey: Prentice-Hall, Inc., 1968, pp. 57-76.

Leone, Robert P.; Oberhelman, H. Dennis; and Mulhern, Francis J. "Estimating Individual Cross-
Section Coefficients from the Random Coefficient Regression Model." Journal of the Academy
of Marketing Science 21, Winter 1993, pp. 45-51.

Leybourne, S. J. "Estimation and Testing of Time-Varying Coefficient Regression Models in the
Presence of Linear Restrictions." Journal of Forecasting 12, January 1993, pp. 49-62.

Makridakis, Spyros; Wheelright, Steven C.; McGee, Victor. Forecasting: Methods and
Applications. John Wiley & Sons, 1983.

Martino, Joseph P. "Trend Extrapolation." In A Guide to Practical Technological Forecasting.
Edited by James R. Bright and Milton E. F. Schoeman. New Jersey: Prentice-Hall, Inc., 1973,
pp. 106-125.

Meadows, Dennis. The Limits to Growth. New York: Universe Books, 1972.

Meadows, Dennis L., and Meadows, Donella, eds. Toward Global Equilibrium: Collected Papers,
Cambridge, Massachusetts: Wright-Allen Press, 1973.

Meadows, Donella, H.; Meadows, Dennis L.; Randers, Jorgen. Beyond the Limits: Confronting
Global Collapse, Envisioning a Sustainable Future, Post Mills, Vermont: Chelsea Green
Publishing Co., 1992.

Meadows, Donella, et al. Groping in the Dark: The First Decade of Global Modeling. New York:
John Wiley & Sons, 1982.

Meadows, Dennis L., et al. Dynamics of Growth in a Finite World. Cambridge, Massachusetts:
Wright-Allen Press, Inc., 1974.

Nazmi, Nader. "Forecasting Cyclical Turning Points with an Index of Leading Indicators: A
Probabilistic Approach." Journal of Forecasting 12, April 1993, pp. 215-225.

Otter, Pieter W. "Canonical Correlation in Multivariate Time Series Analysis with an Application to
One-Year-Ahead and Multiyear-Ahead Macroeconomic Forecasting." Journal of Business &
Economic Statistics 8, October 1990, pp. 453-457.

Stanford Research Institute. "Forecasting Techniques Using Time Series and Projections."
Handbook of Forecasting Techniques. Springfield, Virginia: National Technical Information
Service, 1975, pp. 39-100.

Wooldridge, Jeffrey M. "Some Alternatives to the Box-Cox Regression Model." International
Economic Review 33, November 1992, pp. 935-955.

Statistical Modeling

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistical Modeling

Uploaded by

Copyright:

Available Formats

AC/UNU Millennium Project Futures Research Methodology V2.

You might also like