Equations were subsequently organized into models which analyzed the effects of monetary and fiscal policy on inflation and unemployment. Agricultural models were developed to study not a single crop but rather many crops together so that one could understand the interacting relationships among the prices for a whole group of commodities.
Work then spread beyond microeconomics and macroeconomics across a broad range of economic specialities. Growth models were developed to analyze the role of savings and investment on the process of capital accumulation. In finance, portfolio models were constructed to consider the trade-off between yield and risks in collections of stocks and bonds. Econometricians exploited the higher speeds and greater memory of computers by estimating entire models simultaneously rather than a single equation at a time. Energy economists estimated and simulated models which measured the dynamic response of the public to higher oil, gas and electricity prices.
Today, computational methods range across almost the entire
scope of economics. Recent work on solution methods for
dynamic stochastic problems has permitted theorists and
econometricians to analyze intertemporal economic behavior under
uncertainty. Discrete choice models have been applied in the
transportation sector to analyze the choice between riding the bus
or train. General equilibrium models have been developed to
study the effects of tariff reductions on prices as well as on the
income of wage earners. Global warming issues are analyzed
with models which study the response of electricity producers to
carbon emission taxes. Expert system methods have been applied
to the construction of models which analyze market behavior in
double auctions. With these past contributions as background, the
research opportunities now available in computational economics
are considerable.
B. Opportunities
Many economic models are naturally formulated as dynamic
programming problems since this approach includes optimization
over time in an uncertain environment. However, it is only
recently that advances in computer workstation and mainframe
speeds have reached a level where we are making progress on the
estimation of these models. The models are estimated by
algorithms which employ an "outer" and an "inner" loop. The
outer loop contains a search for the values of the parameters of the
model which best fit the data. This search is in a space which is
the same dimension as the number of parameters. The inner loop
of the algorithm solves a dynamic programming problem
corresponding to the parameter values supplied by the outer
algorithm. The inner loop involves a search over an essentially
infinite-dimensional space of alternative strategies or decision-
rules. These types of algorithms are numerically well-behaved
but pose large computational demands, requiring extremely fast
processors, and huge memory and disk resources. Recently this
methodology has been applied to a wide range of topics including
women's choice of birth control, aircraft engine replacement,
patent renewal, regulation of nuclear power plants,
portfolio/savings decisions schooling choice, and retirement
behavior.
In related work, econometricians are pushing forward on "estimation by simulation" methods. Recently, research has found ways to estimate models which were previously thought to be computationally intractable due to the fact that they required evaluation of multiple integrals with no closed-form solution. Examples occur in models of discrete choice such as the choice of mode of transportation for urban commuters. The new simulation approaches allow one to avoid explicit evaluation of the multiple integrals and instead take Monte Carlo draws to yield simulated choices which are unbiased. This approach provides great potential speed-ups; however, the statistical objective function in such problems can sometimes be a discontinuous or "non-smooth" function of the underlying parameters, causing difficulties for traditional nonlinear "hill-climbing" algorithms that have been employed to optimize smooth objective functions. Even in cases where the objective function is smooth, there is a problem that traditional optimization methods are "myopic" and typically only locate a local optimum, whereas what is really needed is a globally optimal solution. Recent developments in "simulated annealing" and random search algorithms are guaranteed to eventually find a global optimum, but these methods are very computer-intensive.
In recent years there has been increased use by economists of large-scale data sets to model microeconomic behavior. In order to comprehend such massive collections of data, we will need powerful workstations that will allow us to interactively analyze the data using multi-media audio-visualization technologies and other graphical tools. These methods combine the strength of computer numeric processing capabilities with human's superior ability to interpret and understand information in audio-visual format. These methods can also be very effective for identifying and isolating measurement errors that are inevitably present in large survey datasets.
There has been much discussion in the scientific literature about the potential for global warming if we continue to emit large amounts of carbon dioxide and other greenhouse gases into the atmosphere. As yet, there has been much less attention to economic aspects of the problem. This situation is beginning to change with the construction of economic models which enable us to analyze the effects of "carbon emission taxes." If imposed, these taxes would cause shifts away from coal, which has high carbon dioxide emission, and towards oil and natural gas, which have lower emissions. Econometric models are a natural means of analyzing this kind of response.
Economic analysis sometimes posits such elaborate and complex optimizing behavior that it seems unlikely that consumers and producers are sophisticated enough to derive such rules. In the language of computer science, many economic problems are "NP- Complete" and so even slight increases in the "size" or "realism" of certain models can quickly overwhelm the capacity of the largest and fastest supercomputers. A new approach uses artificial intelligence methods to create economic agents with rather simple optimizing behaviors but with the ability to learn. These agents are then placed in a simulated environment to explore whether their learning leads them to become as sophisticated in their behaviors as the complex optimizing behaviors which were posited. One example of this kind of work is the efforts to analyze the double auction institutions used by the Chicago Board of Trade, the New York Stock Exchange and many other financial institutions around the world. This has been done with a double auction tournament with artificially intelligent traders, which has been sponsored by the Santa Fe Institute.
Computational economics has given impetus to the growth of sectoral economics. These models stand halfway between the more familiar models of firms on one hand, and economy-wide models, on the other hand. Examples are models of the agricultural sector in Mexico and Egypt which cover many crops and many inputs such as land, water and fertilizer. These nonlinear programming models enabled agricultural officials to study the response of farmers to policy measures such as the removal of price controls and planting restrictions.
About twenty years ago there was considerable research on economic growth models. However, most of this research was with mathematical rather than computational methods, and the analysis was for the most part restricted to models with one or two sectors so that analytical solutions could be obtained. Also, it was not possible to consider the effects of economies of scale in these models because of the resulting non-convexities. Some computational work on growth models occurred during this period but it was restricted by the relatively slow computers of the time. Today, computer workstations and supercomputers have the capability to solve nonlinear growth models with tens to hundreds of sectors and also the ability to include economies of scale. Thus, in a time when there is rising concern about economic growth, researchers can now return to this area with much more powerful tools of analysis than were available in the earlier era.
One of the puzzles of macroeconomics is why producers respond to downturns in the economy by rapid adjustment of production and employment rather than by adjustment of prices. One of the methods for analyzing this kind of behavior is the use of state- dependent models with stochastic coefficients. Though algorithms for the estimation of macroeconometric models with stochastic coefficients are well established, empirical applications have until now been confined largely to single equation models because of the high dimensionality of these systems. In the near future, models of this type with many equations can be estimated.
In order to improve our understanding of the response of individuals and firms to taxes, wages and prices, economists are increasingly turning to the use of large microdata panels. These panels are possible to create but they require substantial matching across existing databases, extensive editing and long-term commitments of resources. Thus, despite great efforts, the present generation of economic databases may fail in the next decade because they are not designed to deal with the volume, complexity, and diversity of the data which will need to be accessible for economic research. It is imperative that we bring together establishment and firm information on outputs (products and services), inputs (capital, labor, research and development, materials, purchase services), financial and management characteristics, and demographic information on workers (work histories, education, training, etc.) in an efficient computational environment. Also, we must enhance the infrastructure for maintaining and communicating data and the programs that manipulate data.
Economic theory has long been the province of analytical methods which shunned computational techniques. However, that situation is changing rapidly as theorists are discovering that computers can be invaluable aids in their work. We see three efforts which need to be made in the application of computational techniques to economic theory. First, there must be greater diffusion among economists of standard numerical methods and techniques. Many of the published uses of numerical analysis in economics are inefficient and do not use standard techniques. While the simple and ad hoc methods used are sufficient for a particular problem, they leave little room for extensions and generalizations. The utilization of standard numerical techniques has often resulted in highly efficient solution methods, as was the case with application of linear integral equation methods to the Lucas (1978) asset pricing model. However, it is surely the case that once we bring to the analysis the idiosyncratic structure of an economic problem, there will be ways in which specially designed algorithms will be able to use that structure to speed computations. Therefore, a second effort should be development of new numerical methods which are designed to do well on economic problems. The use of standard methods and the development of these methods will be of little interest if they do not advance our knowledge of economic problems. Therefore, a third priority should be special support of researchers who exploit the state of the art in numerical techniques to make their economic analyses more realistic, more believable and better adapted for data analysis.
One of the applications of computational methods to economic theory is in financial markets. An important puzzle of financial theory is why corporations use such a diversity of assets. Various explanations have been developed - tax clienteles and signalling being the most common examples. However, these arguments are made largely in simple static models, often without any general equilibrium analysis. Numerical methods will make it possible to test these alternatives in more realistic dynamic contexts and with a full equilibrium analysis, resulting in models which will be more appropriate for data analysis.
In recent years there has been a major trade agreement with
Canada and now there is discussion of a similar agreement with
Mexico. Some of the important economic analyses for the
Canadian agreement was done on both sides of the border with
computable general equilibrium models. These models enable one
to study the effects of changes in tariffs and other trade barriers on
prices and wages as well as on production and employment. The
models were sufficiently useful in the Canadian case that they are
now being extended to cover trade with Mexico.
C. Infrastructure
A strong research program in computational economics requires
the proper infrastructure in the sense of hardware, software and
networks as well as programming assistants and graduate student
support.
Economic research groups have joined the personal computer revolution, but for the most part they have not upgraded their equipment to the workstations and networks which are powerful tools for modern computational research. Partially this is due to the fact that an individual personal computer can be acquired and operated quite well as a stand-alone device. While a single workstation can be acquired and maintained by a lone investigator, the overhead time required to maintain the software and hardware on such a machine is substantial. Workstations, on the other hand, are more efficiently used in groups centered around servers with an individual employed to maintain the hardware and software on a group of five to twenty machines. Therefore, we recommend that the economics program at NSF be aware of these economies of scale and help research groups get over substantial start-up costs. Once the initial investment is made in a set of workstations, a server and the supporting network, then additional machines can be added to the system at a relatively modest cost.
There are substantial externalities to equipping economic researchers with workstations. One is that researchers, who become accustomed to using networks, find that they can easily collaborate with colleagues located at other universities within the U.S. or abroad. The sharing of software and data and the joint authorship of papers by investigators at various locations is a real spur to creativity which economists should fully exploit.
Networks have another use which provides potential efficiency gains. Late-night users of computer networks observe that there is substantial underutilization of processor capacity on other network machines. One way to balance the load curve is to share computer resources over a network, with software that allows large tasks to be spawned on other machines in periods when there are idle cycles. Thus, network-distributed processing, a special case of parallel processing involving ten or twenty networked processors, offers a powerful way to solve large computational problems. Some experiments along these lines have already been started on the network of Sun workstations at the Federal Reserve Board and we think that NSF should encourage more research of this type.
Software availability is a problem for some research groups in economics. The economics program at NSF has regarded software acquisition as the proper role of the universities and yet many university budget officers are unprepared to cope with this category of expenditures. Software expenses can be substantial - not only in the initial outlay but in the maintenance fees and the cost of buying upgrades. We recommend that when NSF approves the acquisition of hardware in a research budget, the principal investigator be encouraged to request the needed software and funds for the maintenance of that software.
One of the ways to increase efficiency in computational research is to use the "blueprint" approach. Architects do not design each building completely de novo, rather they frequently use some elements from previous blueprints and integrate those with new ideas to complete the design. Similarly, economists need not design a completely new model each time they begin a project. Rather, they can make good use of the previous models as a starting point. However, computer listings of models are not usually published and they may be difficult to obtain. As networks come into increasing widespread use, it will be easier for investigators to request and receive not only the data but also the models used by previous investigators. There remains the difficulty that most investigators become expert with one application and would have high startup costs to master others. For example, two previous models may have been done in the statistical packages TSP and TROLL and the investigator is an expert in another statistical package, viz. SAS. We need software that will translate a model among some of the primary statistical and optimization packages which are widely used by economists. Commercial spreadsheets have made progress on this front and we believe that some investment by NSF in model translators for statistical and optimization packages would yield high dividends.
One of the essential elements of computational economics research is ready access to data. Much data access in economics is still done in the "old-fashioned way" by writing to ICPSR (Inter- University Consortium for Political and Social Research), or Bureau of Census or Labor Statistics to obtain tapes which must then be decoded and the data manipulated to get it into a useable form. We believe that this system could be made much more efficient by the development of a "National Economics Server." This server would maintain a menu-driven interface with which a user could request the desired data. The server would not contain data but would rather have links to institutions which maintain economic databases. The server would fetch the requested data from various institutions, organize it into a common format, download the data and then bill the user for services rendered. Such a server would mean that data dictionaries and other views of the organization of economic data could be provided which are not concerned with the organizations that collect and maintain the data but, rather, with the conceptual view which economists hold. For example, data listings are now organized primarily by the institution that maintains the data and secondarily alphabetically. One alternative form of the data organization would be along the lines of the national income accounts. Also, the individual would not have to learn the format and procedures of several different data organizations but only that of the single server. Finally, the National Economics Server could become a repository of models as well as data which could greatly facilitate the "blueprint" approach to economic research discussed above.
Much of recent research in economics has made use of microdata panels. The creation of these data panels requires substantial matching across existing databases, extensive editing and long- term commitments of resources. A comprehensive initiative is needed in this area to bring together establishment and firm information on product and services outputs, capital and labor inputs, financial and management characteristics, and demographic information on workers.
Some of the most useful data in economics is now in a form called
"complex data." The designs for the collection of such data are
not done by a single person but rather by groups of people. The
data is used not by a single individual but by many persons who
access it in parallel. The data is collected over extended periods of
time and there are alterations of the data design during the period
of collection. The collection and management of this data requires
high capacity communications networks, relational database
management systems and teaching resources to train individuals in
the use of the system. NSF needs to view these complex data
information facilities as national laboratories and provide funds for
their ongoing maintenance in supportive institutions.
D. Budget, Action Items and Conclusions
About twenty years ago NSF provided substantial funding to a
large project in computational economics. These funds were for
the development of an "economists' computing environment"
called the TROLL system which ran on large mainframe
computers. Some of us have considered the possibility that it is
now time for another project of this nature in which NSF would
be asked to fund a single large project to create a new computing
environment for economists which operates on workstations and
networks. Though in our view it may be easier to obtain funding
for a single large project than for many small ones, we do not
think that the most efficient way to move economic research
forward now is a single large project. Rather, we have tried to
demonstrate in this report that computational economics is now
very broad - it extends almost across the entire scope of
economics. There are many rich research opportunities in diverse
parts of the field and there are many young economists now with
the proper training to move ahead sharply in computational
economics.
Therefore, we have presented a budget for roughly 18 million dollars a year, the heart of which is funding of roughly 125 projects per year including funding for postdoctorates, graduate students, programming assistants, equipment and supplies. The budget also includes funding for the National Economics Server.
We have also identified a number of action items. The most important of these is that NSF should facilitate the shift in economic research to a workstation and network based computational infrastructure. Also, we have suggested the creation of the National Economics Server to provide data and models in the manner described above.