Printer Friendly

Practical economic forecasting for small regions.


Forecasts of regional economic variables, such as income and employment, represent valuable information to businesses that serve regional markets and to government agencies that must implement programs at the county or regional level. Their value is greatly enhanced to the extent that they are derived in a manner consistent with economic theory, subject to objective review and capable of being reproduced over time. This paper reports on an econometric modeling approach that has been used successfully in a practical setting.

DETAILED PROJECTIONS of regional economic activity, produced by region-specific forecasting models, can be of great benefit to businesses such as utilities and commercial banks that serve regional markets, and to state and local governments that must plan or implement programs for multicounty areas.

In practice, however, the costs of building and maintaining regional economic models and the constraints on data availability often discourage the modeling efforts necessary to develop regional economic forecasts. Typically, businesses and government agencies interested in regional economic projections either make do with rough impressions not supported by much analysis or simply "allocate" to the regional level from a more aggregated geographic entity for which projections are available.

The regional economic literature provides numerous examples of regional models (see Bolton for a recent survey). But these models, for the most part, are not readily applicable to forecasting in a practical setting, be it industry or government. The typical model reported in the literature is designed to test hypotheses about regional economic structure, and therefore emphasizes theoretical completeness and structural detail. As a consequence, it is likely to possess a degree of sophistication, in its specification or in its estimation, that quickly confronts the resource limits of all but the largest practitioners. The typical forecaster in a practical setting will value a model primarily for its ability to produce reasonable and tolerably accurate forecasts in a regular and timely manner.

In the context of limited resources available, ease of data collection and maintenance, and also simplicity of model specification and estimation, are likely to be more highly valued than structural or econometric sophistication. Another limitation of reported models from the viewpoint of practical forecasting is that they are typically built for predefined geographic entities, such as states or Metropolitan Statistical Areas (MSA), and rely heavily on data that are available for those entities but not available in general. Yet it is not often that the region of interest to a business coincides with a state or an MSA; for the typical business, greater flexibility of regional definition is needed. Even state government agencies, whose ultimate responsibility is a state jurisdiction, may find that the aggregated data mask important intrajurisdiction differences.

This paper reports on a regional economic modeling approach used by East Kentucky Power Cooperative, Inc. (EKPC), a rural electric cooperative that serves 280,000 residential customers and 15,000 commercial customers in east-central Kentucky. These models use quarterly, county-level data to produce regional forecasts of income, employment, wages, population, labor force and the unemployment rate. They are comparatively easy to estimate, utilizing ordinary least squares within a linear framework, and are specified with a view toward obtaining plausible and realistic simulation properties. The endogenous data set is confined to a limited number of concepts that are available, in principle, for all U.S. counties, and are relatively easy to update from U.S. government-supplied tapes. The use of county level data facilitates a bottom-up approach to the construction of economic regions, introducing a high degree of flexibility into the process of regional definition. The approach presented here thus reflects the concerns of regional economic modeling in practice.

The purpose behind EKPC's regional model is to provide data for use in forecasting electricity sales. The importance to EKPC's load forecast of regional economic projections is heightened by two factors: (1) the geographic extent of its service area; and (2) the composition of its customer base. The roughly eighty counties served by EKPC reflect a wide variety of economic activity. Coal mining dominates the economy of the southeast portion of the service territory; heavy industry, such as steel and chemicals, is very important to the economy of the northeast portion; and service-based employment, along with state government employment, provides the economic base for much of the western part of the service area. This diversity suggests the value of regional economic detail in forecasting the demand for electricity. In addition, the growth prospects of EKPC's customer base, which consists mostly of residential and small commercial accounts, is highly dependent on regional employment and income. This dependence is in contrast to the electric utility whose customer base is dominated by large industrial customers, whose prospects are likely to be more directly influenced by national or even international economic events.


The definition of economic regions is a major consideration in practical applications of regional economic modeling. In this regard, a distinction needs to be made between the service or market area of a company and the regional economy or economies in which it operates. The service or market area is not likely to be a valid unit for economic analysis. its boundaries may be determined in a variety of ways, among them the location and strength of competitors (as in the case of a commercial bank), and the geographic extent of powerlines (as in the case of an electric utility). The regional economy, on the other hand, exists independently and represents, in theory, a cohesive economic entity sharing a common trade area and a common labor market. In most cases, the forecaster is interested in identifying and modeling the regional economy or economies that are relevant to his service area. The company's sales growth is dependent on the growth of the region, not just the growth of the service area.(1) In practice, the scarcity and expense of regional data present numerous obstacles to the definition of the regional economies. No data are readily available that allow forecasters to determine, analytically and systematically, the correct regional boundaries. These decisions must often be made, as in the case of EKPC, on the basis of judgment.(2)

Judgment is guided by common sense. If the regions are too broadly defined, the economic diversity that was originally sought will be lost. If the regions are too narrowly defined, they are not likely to have any viability as economic entities, and this circumstance will increase the problem of developing good regional economic data pertinent to the individual regions.

The service area of EKPC is sufficiently diverse and extensive that several regional economies are assumed to exist within its boundaries. It also lies adjacent to several large metropolitan areas that EKPC does not serve but that influence the service area. The boundaries of regional economies were established by grouping together counties that were perceived to share economic similarities or affinities. As an example, the Bluegrass region was defined to include those counties in the EKPC service area with an affinity to the Lexington metropolitan area. The decision on which counties to include was made largely on the basis of travel distance for shopping and commuting. Given this criterion, the Bluegrass region, for example, was defined as encompassing a fifteen-county area lying roughly in the center of EKPC's service area that has the Lexington, Kentucky metropolitan area at its core.,, While no clear test of the viability of regional boundaries exists, the residence adjustment to income (collected at the county level) indicates, in net, how much income earned by residents of a region derives from employment in that region.

Consistent with the idea of modeling the regional economy and not just the service area, the regions defined in the EKPC data include substantial areas not served by EKPC member cooperatives, chiefly towns and cities, where electric loads are served by investor-owned utilities or by municipals. A prime example is the city of Lexington. Although this city is not itself served by EKPC, it is included in the defined region because the surrounding, served territories are strongly dependent on the Lexington economy. Rural areas typically depend substantially on nearby cities and towns for jobs and income. Income, employment and population levels may differ in the rural and urban areas of a single region, but the growth of the region still depends on the economic performance of the region as a whole, and especially the towns and cities.

On the other hand, cities peripheral to EKPC's service area and also not served, such as Louisville, Cincinnati and the Huntington-Ashland metropolitan area, were excluded from the definition of nearby regions. Economic linkages are presumed to exist between served areas and these cities, but it was considered important to distinguish the geographic limits of EKPC's service region. Furthermore, it was judged that the additional cost of collecting data for these areas (which spread over four different states) was not warranted by the additional benefit.


Data collection and development are costly and time-consuming processes and key issues in practical regional model building, especially when the model is expected to be reestimated regularly. The basic regional economic data are those readily obtainable in magnetic form, typically from U. S. government sources. The decision to include other kinds of regional information requires a careful consideration of benefits and costs. Locally collected data on housing starts, for example, may improve the forecasts of the construction employment and income, but the improvement must warrant the expense of collecting this information. Incorporating additional data would probably be highly labor intensive. More sources would have to be contacted, more publications would have to be monitored, and more data would have to be regularly entered into databases by hand. In addition, more variables would have to be forecast, either outside the main regional model or as part of that model, which would increase the cost and complexity of the modeling process.

Regional models often incorporate constructed data, which is allocated to the region from the state or national level on the basis of employment, income or some other variable actually measured at the regional level. While such data may serve the needs of particular model specifications and produce forecasts of variables of particular interest, it has several important drawbacks (see Freedman). For one thing, because constructed data are a function of ratios of existing data that itself can be used directly in the estimation, it provides additional detail without necessarily providing additional information. For another, the process of constructing data through allocation fails to confront true regional differences that may exist in the concepts being constructed.

The EKPC regional models utilize a variety of national and regional data, all of which are readily available in magnetic form. The major groups of variables and the supplying agencies are summarized in Table 1. Each variable in the regional data set corresponds to a single government source variable; there is no derivation of output or other concepts not actually measured locally.

One problem with the source data is discontinuities that arise primarily because of changes in the coverage provisions of the unemployment compensation law (under which the data are collected). Another problem is missing values that occur because of the nondisclosure rule.(4) These cases require some kind of data derivation to produce consistent and meaningful time series. In EKPC's case, the affected variables were regressed against closely related regional variables for which consistent data were available and the predicted values

were employed as if actual in the affected intervals.


Like most regional modeling efforts, the approach described in this paper departs from the aggregate income-expenditure approach that underlies the standard macroeconomic model, primarily because no data on final demand components exist at the regional level. The region is treated as a satellite of the national economy.

The model structure reflects the need for straightforward econometrics and meaningful measures of regional economic performance within the limitations imposed by considerations of cost. Figure I outlines the structure of a "typical" regional model used in the case of EKPC. The actual model structure is more complicated, consisting of thirty-eight stochastic equations and eight identities. The model structure is basically recursive. Changes in the national economy, translated to the regional economy through their impacts on employment and income in "export" industries such as mining and manufacturing, are the prime cause of change in the regional economy. There is no feedback, however, from the regional economy to the national economy. The change in "local" industries, such as trade and services, is a function primarily of total regional income. In practice, regional industries may defy easy categorization as "export" or "local". There are likely to be instances where service industries, normally considered to be local, actually serve national markets, and instances where manufacturing industries, usually considered export industries, serve regional customers exclusively.


Employment is estimated for ten sectors: agriculture; mining; construction; manufacturing; regulated industries; wholesale trade; retail trade; finance, insurance and real estate; services and government. The employment equations are in the form of labor demand relationships, where labor demand is a function of output. In the case of manufacturing employment, regional manufacturing output is proxied by the Federal Reserve Board production index for manufacturing. For most of the other employment sectors, regional income and regional population serve as proxies for regional output. The real mortgage rite is included in the contract construction equation to reflect the demand for housing. In many regions, including this one, commercial employment has maintained a strong growth rate even during cyclical downturns in regional income and employment. A reasonable hypothesis is that this growth is due to structural changes, including changing demographic characteristics and changes in the composition of output, which are fundamentally exogenous to the structure of flows between sectors specified here. Consequently, either a time trend or national employment levels in commercial categories are often included to reflect these forces.


Regional wages, by sector, are calculated as total sectorial labor earnings divided by average sectorial employment, and are estimated for each of the ten employment sectors. These measures of wages will be influenced by the age and experience of the regional workforce, and thus are not true market-equilibrium wages, but are only representative of those wages. The changes in average regional wages, however, will be subject to most of the same determinants as changes in market wages. Therefore, the basic specification for average wages reflects demand conditions in regional labor markets through the inclusion of the regional unemployment rate and changes in labor productivity through the inclusion of U.S. output per manhour. Lagged wages are often included to capture the adjustment process as regional wages respond to changing labor market conditions.

The manufacturing wage is included in some of the other sectorial wage equations on the theory that manufacturing often serves as an employment option of first choice and that conditions in the market for manufacturing labor will influence conditions in the labor markets for other sectors.


Separate income equations are estimated for each of the ten employment and wage sectors. Regional employment and regional wages are the usual explanatory variables. Other income concepts estimated are farm gross earnings, contributions to social insurance, transfer payments, proprietor's income and the residence adjustment. Farm gross earnings are usually trended. The next three concepts are related to regional population and income and to appropriate national variables. The residence adjustment transforms income from place-of-work to resident.(5) The residence adjustment is trended into the future.

Population and Labor Market

Three concepts are estimated within the labor market block: the regional unemployment rate, the regional labor force and regional population. The unemployment rate is a function of total regional employment and the regional labor force. Labor force, in turn, is determined by regional population. Population is determined by average regional wages and the regional unemployment rate. This specification for population focuses on its migratory component, which is expected to be influenced by changing economic opportunities in the region.


R-squared and Durbin-Watson statistics were considered in determining the final specifications of the equations. Efforts were made to improve the fit of the equations, usually through respecification. Low Durbin-Watson statistics, calculated for some of the equations, were believed to be due to inherent specification problems, which, while not desirable, are nevertheless unavoidable within the limitations of the data, and are not considered serious enough to bias the forecast results.(6) The main focus of the modeling exercise, however, was to produce reasonable forecasts of the regional economic variables. Therefore, structural variables were judged primarily on the basis of whether their signs were plausible in light of economic theory, and whether the magnitude of their coefficients were such as to make a material contribution to the forecast. Both sign and size of coefficients are crucial to simulation and forecasting. In general, "low" t-statisties were not, by themselves, sufficient reason to reject a structural explanatory variable or a specification.

Binary variables were added to the equations to reflect structural change not captured by the available structural variables. In some cases, the binaries correspond to a known structural change (that is, a plant closing or a coal strike). In other cases, they correspond to a suspected change. Binaries were retained to the extent that they were statistically significant and improved the forecast.

The results achieved with the model constructed for the Bluegrass region will be used to illustrate the simulation and forecast accuracy to the EKPC modeling approach. The model was estimated for the period 1975-85, two years short of the period for which actual data were available. With the coefficients obtained from this estimation, the model was simulated for the period 1975-85, using actual exogenous and lagged endogenous variables, and forecast for the years 1986 and 1987, using actual exogenous but predicted endogenous variables. One way to assess the solution values of the model is to calculate the mean absolute percentage error (MAPE) of each of the endogenous variables.(7)

Overall, the performance of the Bluegrass model compares well with that of other models reported in the literature. While no widely agreed-upon standards exist, a MAPE of less than 3 percent is considered to be tolerable (Glickman). For the Bluegrass region, 80 percent of the endogenous variables in the simulation period and 74 percent of the endogenous variables in the forecast period had MAPEs of less than 3 percent. These compare to simulation performances of 31 percent, 77 percent, and 79 percent, respectively, for the Delaware (Latham, Lewis and Landon), Los Angeles (Hall and Licari) and Chicago Duobinis) models.

The simulation performance of the population equation was especially good, having an historical MAPE of 0.1 and a forecast MAPE of O.O. This result, typical of all the EKPC models, suggests the adequacy of simple models for predicting aggregate population growth.

The MAPEs were greater for variables that tend to be highly cyclical, such as the unemployment rate and manufacturing employment, than for variables that tend to be more stable, such as population and total regional income. Often the peak-to-trough variations in the cyclical variables were more extreme than predicted by the model. Also, the MAPEs were generally greater for the more disaggregated variables. For example, total employment had a smaller MAPE than all but one employment sector in both the simulation and forecast periods.


The purpose of this paper has been to describe an approach to regional economic modeling that has been successfully employed by East Kentucky Power Cooperative Inc. to produce regional economic forecasts for the purposes of electric load forecasting and general regional economic analysis. The approach has several distinct advantages as a tool for practical regional economic forecasting. Foremost among them are its flexibility in defining regions and its ability to produce reasonable forecasts with comparatively simple model specifications and estimation techniques. It also has the advantage of utilizing widely available data that can be collected from a relative handful of sources, thus minimizing the time and expense of maintenance.

The limitations of the approach reported here derive mostly from the scarcity of regional data both for defining regions to model and for specifying equations. As a consequence, it is less useful as a basis for structural economic analysis. Furthermore, the effective use of the modeling framework described here requires sound judgment on the part of forecasters as to what constitutes a reasonable forecast. With these limitations in mind, however, this approach can provide a more accurate and systematic approach to regional economic forecasting than the strictly intuitive approach followed by many businesses and government agencies.


1 In the case where electric utilities are using manufacturing employment as a proxy for manufacturing output as an input to an industrial load forecast, a narrower regional definition may be employed. This would arise from the need to capture closely the economic variations associated with the utility's customers.

2 EKPC produces separate regional economic forecasts for each of its twelve member cooperatives, thus introducing an additional constraint into the process of regional definition, i.e., that each member cooperative fall within a single region.

3 The counties comprising the Bluegrass region are Anderson, Bourbon, Bracken, Clark, Fayette, Franklin, Harrison, Jessamine, Madison, Menifee, Montgomery, Nicholas, Powell, Scott and Woodford.

4 Under the nondisclosure rule, the government suppresses economic data in order to maintain the privacy of firms when it represents an aggregate of fewer than three reporting units in a county.

5 If the residence adjustment is positive, then the region in aggregate is exporting labor (i.e., commuters) to other regions. The income earned by these workers must be added to place-of-work income for the region to sum to total resident income. If the residence adjustment is negative, then the region in aggregate is importing workers and their income must be subtracted from place-of-work income to derive total resident income.

6 An Appendix, showing the estimated equations for the Bluegrass region, is available from the authors.

7 The MAPE statistic is calculated by summing the absolute values of the percentage difference between the predicted value and actual value of the endogenous variable over the number of observations. The MAPE statistics for all the endogenous variables in the Bluegrass region are shown in an Appendix available from the authors.


Bolton, R. (1985). "Regional Econometric Model." Journal of Regional Science. 25(4): pp. 495-520.

Duobinis, S.F. (1981). "An Econometric Model of the Chicago Standard Metropolitan Statistical Area." Journal of Regional Science. 21(3): pp. 293-319.

Freedman, D. (1981). "Some Pitfalls in Large Econometric Models: A Case Study." Journal of Business. 54(3): pp. 479-500.

Glickman, N.J. (1974). "Son of The Specification of Regional Econometric Models.' " Papers of the Regional Science Association. 32: pp. 155-177.

Hall, O.P. and J.A. Licari (1974). "Building Small Region Econometric Models: Extension of Glickman's Structure to Los Angeles." Journal of Regional Science. 14(3): pp. 3:37-353.

Latham, W.R., K.A. Lewis, and J.H. Landon (1979). Regional Econometric Models: Specification and Simulation of a Quarterly Alternative For Small Regions." Journal of Regional Science. 19(1): pp. 1-13.
COPYRIGHT 1991 The National Association for Business Economists
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1991 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Fiske, John R.; Lamb, James C.; Morss, Mark F.
Publication:Business Economics
Date:Jul 1, 1991
Previous Article:The business economist at work: the research department of the Federal Reserve Bank of Cleveland.
Next Article:Statistics corner: issues for official statistical systems in democratic market societies.

Terms of use | Copyright © 2017 Farlex, Inc. | Feedback | For webmasters