ALBUQUERQUE, N.M. – A prototype economic model developed on one of the world’s fastest supercomputers is including data about human beings to an extent never achieved before.
The idea is to better predict the effect of new economic programs and policies.
The computer program Aspen, devised at Sandia National Laboratories, is designed to combine staples of macroeconomics — interest rates, trade policies, legal restrictions, and so on — with data on the savings and spendings of individual moms and pops, wherever in the United States they may be.
The latter data, part of microeconomics, for technical reasons has never been combined with macroeconomics for the U.S. economy as a whole, though economists have speculated the combination would result in more accurate forecasts.
Said economics Nobel Prize winner Lawrence Klein, “It’s a promising way of modeling the economy. No one has had the commitment and hardware to do it properly.” The idea of using real-world data about households and other small economic agents “was a technique proposed in the l960s when computers were in their infancy and data availability was much less,” he said. “Now, computational problems should not be an obstacle, and if you specify the area, you could make a test.” Klein, an economics professor at the University of Pennsylvania, is a consultant on the project.
The program in its final form should be useful for businesses, venture capitalists, company financial officers, bankers, and others interested in analyzing business cycles, studying financial implications of technology shifts, and predicting the effects of legal or policy changes of the Federal Reserve or the Department of the Treasury.
Sandia physicist Rich Pryor presented the program’s possibilities to technical personnel of the U.S. Federal Reserve Boards in St. Louis and in Chicago in December, accompanied by Sandia economists Tom Quint and Nipa Basu. Pryor, lead investigator on the project, has been invited to make another presentation with Basu and Quint at the Federal Reserve Board’s headquarters in Washington, D.C., in mid- March.
Use of the current prototype program is available at a nominal fee to individuals, businesses and government agencies interested in playing out their own economic scenarios with video game-like screen interactions, using a protected computer link into Sandia. A telnet connection and X Window System software are necessary to log in.
A more detailed economic simulation is expected to be on-line by May.
According to Richard Anderson, a research officer at the Federal Reserve Bank in St. Louis, “The program shows promise as a tool to analyze certain problems, such as interactions between households, that economists haven’t been able to attack. The method handles a lot of households.”
Computer sub-codes called genetic algorithms allow economic actors to “learn” from their mistakes, rewarding correct decisions by creating “descendants” of those people or institutions who make them, and eliminating those who chose incorrectly.
Because the program’s macroeconomic component relies mainly upon mathematical probability, the model should be better able to compute the effect of new economic programs like the proposed flat tax, NAFTA (North American Free Trade Agreement) and GATT (General Agreement on Tariffs and Trade), for which there are no historical experiences required by standard macroeconomic models, said Basu. “Anyone qualified who logs on even to the prototype model will understand its potential as a tool, and, by modifying the parameters, see some new ways of solving economic problems.”
Said Stephen Gibson, executive director of the Bionomics Institute in San Rafael, Calif., “The approach Pryor’s taking is one we’d like to encourage. We believe the economy is more like an eco-system — more of a complex adaptive system — than a giant machine.” Gibson believes Pryor’s program reflects this understanding.
In Pryor’s programs, households deposit money in banks, people are hired and fired, banks lend to companies, corporations buy corporations, and all “learn” from the experience. “If a corporation in the model raises the price on goods and the price rise is right, then a genetic algorithm reinforces that decision — its correctness allows it to survive, and other corporations may be deleted. Things successful breed with things successful and make offspring,” Pryor said.
The prototype is a rudimentary model of a simple market economy consisting of government, industry, and households interacting with each other. Household demand for products is dependent on family size and income, the government collects taxes and pays benefits to the unemployed, and firms compete against each other by varying the price of their products and “learning” from the result. The model was developed primarily for testing message-passing algorithms and the logical ordering of the decisions of agents. When fed real-world data, the model produced results close to reality, including a business cycle, though it was not yet accurately represented.
A more complex model currently under development includes a banking system, the Federal Reserve, a bond market, a real estate market, and more types of industry. The model emphasizes interest-sensitive sectors of the economy, should be useful in analyzing monetary issues, and is expected to be available for use by May. Further inclusions will include government functions of regulation, organization of bond markets, and financing of the public sector.
“The simulation represents the behavior of basic decision-making agents within the economy,” Pryor has written. “By explicitly setting out how an individual or group tries to maximize usage of something they’ve bought, or make a profit on something they’ve sold, the program realistically models the behavior of agents in a complex environment.”
Barbara Bergman, an economics professor at American University in Washington, D.C. and consultant on the project, said, “This worthwhile research can be useful in policy planning. Conventional macroeconomics has not advanced very far in modeling the economy since 1950.” Macroeconomists use “macroequations that involve total consumption of the whole economy, total income, total assets, total rate of growth. You put together a hundred equations and solve them simultaneously without dealing with consumption of individual families and that family’s assets from data base provided by the census bureau and other information-gathering groups.”
Said Jim Shaw, president of BergenShaw International, a consulting firm that specializes in advanced pattern recognition systems in San Jose, Calif., “Economics is sort of like religion. It has unconfirmed results. Emergent properties often escape the attention of macroeconomists, while what Rich is doing is based on probabilities — a 30 percent chance, say, of someone doing this or that. Randomness enters his program.”
According to Axel Leijonhufvud, director of the Center for Computable Economics at the University of California at Los Angeles, “The model of the Pryor group has adaptive features. Other models depict taxpayers as though the way they save and pay taxes are precalculated, like the flight of an ICBM, from childhood to retirement. One assumes one’s agent has all the information needed for optimal decisions, that he has solved his problems at a very tender age, there’s no feedback or alteration in the behavior pattern.”
The decisions of householders in their budgeting, shopping, educational needs and so on are generally considered only in microeconomic theories that — because of the huge amount of computation necessary and scarcity of data available — could compute only sections of the economy.
The actions of large forces such as trade policies and legal restrictions are generally considered macroeconomic — easily enterable in computers, but providing a picture painted by a very broad brush, with the corresponding probability of omitting factors of importance.
While macromodels can provide accurate forecasts, their general reliance on past experience limits their applicability when new economic policies are introduced. The Sandia mathematical model doesn’t need historical experience on the large scale to build its model. “If you want to assess the impact of GATT or NAFTA, econometric macromodels can’t do it because there’s no past experience. But if you work from microeconomics, you can build up,” Pryor said.
The program — “a big job,” says Pryor — was created at a national laboratory because its resources and stability of staff made it possible.
The program also has potential military uses, in which the agents are tanks and planes being built or destroyed rather than commercial companies and households prospering or declining.