Financial markets and retail construction cycles.

Moss, Steven ; Parker, Darrell ; Laposa, Steve 等

INTRODUCTION AND LITERATURE

It is not universally accepted that specific activities such as retail construction have a cycle of their own. It can be argued that time series data with apparent cycles may be nothing more than chance happenings (Markridakis, Wheelwright & Hyndman, 1998). At the macro-economic level it is generally believed that the US economy has a business cycle (Reilly, 1985; Ritter, 1995). A business cycle is a wavelike or oscillating pattern about the secular trend (Mendenhall & Sinich, 1993). Granger and Newbold (1977) argue that when a series is plotted through time it may appear smoother than white noise. The autocorrelation or smoothness can be the effect of sums of cosine waves in the data that form a linear cycle model. Cycles can be two to ten years or longer in length and are not always a function of the overall business cycle (Bowerman & O'Connell, 1987).

Supply modeling for real estate markets has continued to progress over the last 14 years, and there are standard products available from information specialists. Early supply models built upon the foundation laid by Rosen (1984), who presented a model for office stock integration with new construction, vacancy rate, and rent. Wheaton (1987) extended Rosen's work adapting econometric models to national office building time series data. Born (1988), Grenadier (1995), Clapp, Pollakowski, and Lynford (1992) have developed models for the office market and real estate cycles. Real estate cycle estimation is important for investors and real estate professionals. Tsolacos (1999) states that "Quantitative studies of retail property developments are very useful to real estate professionals because the output of this work can be incorporated into the information set that analysts use to form expectations about cyclical trends of new retail development".

Capital markets, interest rates and the real estate construction cycle may be interrelated. During periods when real estate financing is readily available at low interest rates overbuilding may occur reducing rents, increasing vacancy rates and increasing the risk of defaults on mortgages. The property manager should consider the relationship between lease duration and timing within the real estate cycle. During the bottom of the cycle when rents are at a low point it would be inadvisable to sign long-term leases at the current rent level.

Benjamin, Jud and Winkler (1998a) propose that modeling retail supply is important for academics and professionals. They conclude that retail sales are related the demand for retail space but that demand and supply are inelastic. Benjamin et. al. conclude that interest rates do not have a significant impact on retail supply. Benjamin, Jud and Winkler (1998b) model individual MSA's retail supply adjustment as a function of retail sales and interest rates. Benjamin et. al conclude that supply adjustments are inelastic in the short run in relationship to retail sales and elastic in the long run. Scott and Judge (2000) argue that property investment values are influenced by the development cycle.

There is evidence that bank lending patterns change over time and may follow cycles. The Federal Reserve Bank 1997, Nugent (2001), and the Kulish (2001) report that banks change lending standards over time. The changing lending standards make it easier or more difficult for developers to obtain loans. Asea and Blomberg (1998) concluded that banks lending standards systematically vary over a cycle. They also use national unemployment rates as a proxy for national lending standards. Within a regional analysis of lending, unemployment becomes a local market measure. The empirical question remains as to the robustness of this role of unemployment to lending at the local level.

Henneberry's 1999 study showed that there are cycles in office building construction in England. Additionally, Henneberry argues that the cycles are convergent due to mature capital markets and at the same time divergent based on regional factors. Tsolacos (1999) used a single series from 1985 to 1996 and modeled retail construction starts as a function of consumer spending, rents, and yield levels using a VAR model. Tsolacos (1999) concluded that there are retail construction cycles in England, but that the cycles are not convergent in relationship to national investment markets. Scott and Judge (2000) analyzed a series from 1956 to 1996 and concluded that there is a retail property value cycle in Great Britain with a cycle length of 7.8 years Hudson-Wilson and Pappadopoulos (1999) concluded that there are real estate cycles and they are linked to capital markets. Hudson-Wilson and Pappadopoulos also concluded that real estate cycles are divergent across markets and that traditional modeling techniques fail to recognize these cycles. Hudson-Wilson and Pappadopoulos contend that real estate cycles are a key factor missing in the estimation of default risk for commercial mortgage-backed securities (CMBS). Lawson (1998) proposes that there are retail real estate cycles and that different cities and different types of retail have different cycles.

The purpose of this study is to: 1. Show there is evidence of convergent cycles in retail construction, and 2. Model the retail construction cycle as a function of capital markets. Panel data will be used to model 58 MSA's time series for a period of 27 years each. A series length of 27 years would normally be far too short to estimate autoregressive functions and evaluate potential cycles. A cross-sectional/time-series regression model (panel data) using all 1566 observations preserves the integrity of the time series structure. This will allow for time series analysis and cycle estimation in a relatively short time series.

DATA

The dependent variable in this study is retail supply (RS). RS is gross retail construction, in square feet, for 58 metropolitan statistical areas (MSAs) for the periods 1970 to 1996. The data is comprised of 58 time series of 27 periods each, for a total number of 1,566 observations. The retail supply data was obtained from FW Dodge.

There are two kinds of independent variables used to model retail construction, local MSA level and capital market variables. The local variables vary by both MSA and by year. The capital market variables vary by year, but are the same for all MSAs in a given year. The capital market variables are non-farm, non-residential loan balances outstanding for the United States and Other Areas (CB) as reported for commercial banks by the FDIC.

CB has been adjusted to 1996 dollars (ACB). The second capital market variable is the prime interest rate (Prime) as reported by the Federal Reserve Board. Both ACB and Prime are measured as first differences

The first local independent variable is households (HH) by year, by MSA. Households are the number of families in a MSA. Households are used in this model instead of population. Retail expenditures such as televisions, furniture, drapes, etc. maybe more closely related to the household units, instead of the individuals within the households. Household data was obtained from Woods and Poole.

The second local independent variable is retail sales (Sales) by MSA, by year, adjusted to 1996 dollars. Retail sales were obtained from the US Department of Commerce, Bureau of Economic Analysis. Descriptive statistics for all variables are shown in Table 1.

METHODOLOGY

A cross-sectional/time-series regression (panel data), shown in equation 1, is utilized in this research (Hsiao, 1986). Ambrose and Nourse (1993) used this methodology to analyze capitalization rates.

[y.sub.it] = [summation] [Bx.sub.it] + [E.sub.i] + [M.sub.t] + [N.sub.it] (1)

for,

i = 1 to 58 t = 1 to 27

x is the vector of independent variables

B is the vector of regression coefficients

[E.sub.i] is the individual effect

[M.sub.t] is the time effect

[N.sub.it] is the random error term

One advantage of the cross-section/time-series regression model for panel data is that the regression parameters are estimated with all 1566 observations, whereas a yearly index of the 58 MSAs would have only 27 observations. Indexing would also eliminate the ability to test for differences in response by MSA. Pooling all 1566 observations without regard to time period would remove the possibility of incorporating lagged relationships in the model.

The fixed effects model also provides tests for response differences relative to time (Mt) or individual MSA ([E.sub.i]). These tests involve partitioning the residuals of the linear regression equation by year and by MSA and performing an ANOVA on the partitioned residuals to test for main effects relative to time period or MSA (Hsiao, 1986). A finding that time period is significant in the residuals would indicate that the mean response across MSAs differs by year. On average for all 58 MSAs, the model would systematically over-predict in some years and under-predict in others. A pattern in the residuals based on time period may be evidence of a cycle.

A finding that individual or MSA effects are significant would indicate that some MSAs have higher or lower retail construction starts (after regressing out the effects of the independent variables) relative to other MSAs regardless of the time period. This could be an indication that the model is divergent across MSAs.

Each MSA's retail supply exhibits a positive trend over time. The time series methodology used in this paper, (see methodology), requires that all variables be transformed to stationary series. A stationary series is defined as a time series with a constant mean and variance (Vandaele, 1983; Fuller, 1976). Using a stationary series avoids problems with spurious correlations. To obtain a stationary series, first differences of retail supply (DRS) were obtained. Since the data set has 58 time series the individual and average observations were analyzed for stationarity.

The mean and variance of the series now appear constant, satisfying the conditions of a stationary series. The auto-correlation in the series will be discussed in the results sections to follow. The first differences of RS represent the new retail space added each year.

The household series exhibit varying degrees of positive trend over time. First differences (DHH) were obtained to transform HH to a stationary series. The DHH represents the net change in households each year within an MSA. Retail sales also show trends within each MSA. First differences of sales, (Dsales) are used to transform sales to a stationary series

RESULTS

Total retail construction, shows upward trends with no distinct cyclical pattern. DRS, shows autocorrelation but no distinct repeating cycle. To determine if DRS is cyclical the effects household formations (DHH), retail sales (Dsales) and short-term autocorrelation will be regressed out of DRS. Then the residuals of the DRS model will be analyzed to determine if there is a cycle present (Bails & Peppers, 1993; Markridakis et.al, 1998).

The first step in developing the regression model is to observe the autocorrelation and partial auto-correlation function to determine the autoregressive structure for DRS.

The autocorrelation function and partial autocorrelation function show that lags of DRS up to 6 periods may need to be included in the model. The Ljung-Box Q-statistic, significant at the 1% level, confirms the DRS series is not white noise.

After the autoregressive structure of the dependent variable is determined, the independent variables, DHH and Dsales, are added to the model. During this process the objective is to achieve white noise in the residuals. The regression model shown in equation 2 is the result.

DRSt =216.24+.082D[Sale.sub.t-1]+12.09[DHH.sub.t]+13.16 [DHH.sub.t-1]+.57[DRS.sub.t-1]+.07[DRS.sub.t-2]+.16[DRS.sub.t-6] (2)

(4.3) * (5.7) * (2.6) ** (12.4) * (1.7) *** (6.3) *

[R.sup.2] = .677

where,

t statistics are shown in parenthesis

* significant at the 1% level.

** significant at the 5% level

*** significant at the 10% level

The Ljung-Box Q test, significant at the 10% level, indicates that the residuals may not be white noise. The autocorrelation function and partial autocorrelation function for equation 2 indicate that the residuals for equation 2 are white noise.

Equation 2 confirms Benjamin, Judd and Winkler (1998) findings that there is a relationship between retail sales changes and retail construction. Benjamin, Judd and Winkler findings suggested that this relationship lagged on average 8.1 years, varying from 2 to 20 years by MSA. Our results using the cross-sectional/time-series regression model with first differenced variables show intermediate lags of retail sales are not significant. Longer lags, six to 9 years, are statistically significant, but have a negative coefficient and almost no observable impact on the [R.sup.2]. The model also indicates that household formations play a significant role for predicting retail construction starts.

Dokko, Edlestein, Lacayo and Lee (1999) found that when developing a model for income and value cycles that each MSA had its own cycle and required its own estimation (the cycle was divergent across MSAs). The significance of the regression coefficients in equation 2 using all 58 MSAs simultaneously indicates that the relationships are to some degree stable across MSAs for retail construction. If the relationship between DRS and the independents changed by MSA the coefficients would not be statistically significant in the cross-sectional/time-series regression.

To determine if there is a long-term cycle in retail construction, the residuals from equation 2 were averaged by year. The yearly average is based on a yearly sample size of 58, one data point for each MSA in the study.

This analysis also indicates that a cycle does exist for the residuals and is stable (convergent) across MSAs. If each MSA were on an independent cycle (divergent) the residuals by year would have been random and cancelled each other out resulting in a mean residual of zero for each year. This finding is confirmed by the ANOVA model for the residuals of the cross-sectional/time-series regression (equation 2) shown in Table 2.

The ANOVA indicates the residuals are not significantly different by MSA but are significantly different by year. The convergent cycle confirms the findings of Henneberry (1999) and Hudson-Wilson and Pappadopoulis (1999). The cycle length is approximately 8-10 years the same as found by Scott and Judge (2000). Low points of the construction cycle are in the early 1980's and 1990's. High points of the cycle are in the late 1970's, mid 1980's and mid 1990's.

Prior studies such as Tsolacos (1999), Benjamin, Jud and Winkler (1998), Henneberry (1999), and Hudson-Wilson and Pappadopoulis (1999) have suggested that retail construction cycles may be influenced by capital market variables. The second objective of this research is to determine if the retail construction cycle, observed in the residuals of equation 2, is influenced by capital market variables. To model the retail construction cycle the average residual from equation 2 serves as the dependent variable. The independent variables are U.S. commercial banks non-farm, non-residential loan balances adjusted to 1996 dollars (DACB) and the prime rate (DPrime), both in first difference form. Both variables are the same by year for each MSAs. The resulting regression is shown in equation 3.

Avg[Res.sub.t] = -23.27 + 9.44[DACB.sub.t] - 7.57[DACB.sub.t-2] - 37.05D[Prime.sub.t-2] (3)

(2.32) * (2.14) * (1.79) **

* significant at the 5% level

** significant at the 10% level

[R.sup.2] = .498

Where,

DACB = First differences of commercial bank portfolios

DPrime = First difference of the prime interest rate

The [R.sup.2] for equation 3 indicates that 50% of the retail construction cycle can be explained by two capital market variables. The convergent retail construction cycle is inversely related to changes interest rates on a lagged basis. The model also indicates that when banks increase their outstanding loan balances there is an increase in retail construction followed by smaller lagged reduction in retail construction two years later. The conclusion can be drawn that to the degree there are national lending and interest rate cycles they are influencing the retail construction cycle.

Having established that there is a convergent retail construction cycles influenced by capital markets, the next step is determine the relative importance of the retail construction cycle driven by capital market variables versus the local MSA level variables in equation 2. This is accomplished by adding the U.S. commercial bank portfolio and prime rate variables to the cross-sectional/time-series regression. The results are shown in equation 4 and reported fully in Table 3.

[DRS.sub.t] =146.87+.03D[Sale.sub.t-1]+11.55[DHH.sub.t]+14.62 [DHH.sub.t-1]+.54[DRS.sub.t-1]+.09[DRS.sub.t-2]+.14[DRS.sub.t-6] (2.0) ** (2.5) ** (3.1) * (12.0) * (2.4) ** (5.7) *

12.03[DACB.sub.t] - 7.89[DACB.sub.t-2] - 56.85D[Prime.sub.t-2] (4) (4.9) * (4.0) * (4.3) *

[R.sup.2] = .694

where,

t statistics are shown in parenthesis

* significant at the 1% level.

** significant at the 5% level

The results show that the relationship with the national financial variables remains the same when the model is simultaneously estimated across all 58 MSAs. The results also indicate that the impact of the capital market variables is incremental. All the local MSA level variables from equation 2 remain statistically significant with only small changes in the coefficients.

The Ljung-Box Q test, autocorrelations and partial autocorrelations for the residuals of equation 4 now indicate that the residuals are white noise. In other words, the cycle in the residuals has been removed by adding the capital market variables to the model.

The additional predictive power of the retail construction cycle can be determined by comparing the [R.sup.2] of equation 2 and 4. This provides a measure of the importance of the capital market variables and thus the retail construction cycle. By adding the capital market variables driving the retail construction cycle to the model the [R.sup.2] only increases from .677 to .694. This would indicate that although there is retail construction cycle influenced by national bank lending and interest the additional explanatory power of the capital markets is minimal compared to local variables such as local retail sales, house hold formations, and short term construction trends. The cross-sectional/time-series model was also estimated with only the capital market variables. The variables were all significant in the resulting equation. The R2 for the equation was only .059 confirming the minimal impact of the capital markets on MSA level retail construction starts.

Another check for robustness can be found by examining the impact of recessionary periods upon the estimation. Asea and Bloomberg (1998) use monthly data and find that changes in lending standards have a more profound effect on the economy during an expansion than a contraction. For annual data at the MSA level the use of recessionary dummy variables while significant do not change the nature of the results and add little to improve the explanatory power of the model.

To further verify that the results are robust each of the capital market variables was regressed against the local variables including all possible time lags of the local variables implied by equation 4. Each of the local variables was then regressed against the capital market variables and all possible time lags of the capital market variables implied by equation 4. Using the resulting [R.sup.2] from the equations a variance inflation factor was then calculated for each independent variable. The results are shown in Table 4.

The VIF's are all close to one indicating that the model does not have a significant amount of multi-colinearity between the local and capital market variables.

CONCLUSION

This paper has presented evidence that there is a convergent cycle in retail construction. The retail construction cycle may be obscured by the effects of differing levels of growth in factors such as retail sales and household formations within each MSA. After the effects of varying growth rates and short term autocorrelation in the individual MSA have been regressed out a distinct convergent cycle appears. The cycle length estimated by this research is 8 years confirming the cycle length observed by Scott and Judge (2000).

If the cyclical fluctuations observed are random it is unlikely that the cycle would have remained consistent across 58 time series. If the cycle observed were simply random behavior in the time series it is more likely that with a sample size of 58 the cycle should average out to a mean residual of zero by year. It has been shown that the residuals are significantly different from zero by year.

Having determined that there is an underlying convergent cycle in retail construction the cycle timing can be analyzed. The bottom of the cycle is in the early 1980's and 1990's. This coincides with the last two recessions in the US economy. The retail construction cycle can be explained by capital markets, bank lending and interest rates. The impact of the capital markets is incremental but has little additional explanatory power for predicting retail construction at the MSA level.

When predicting MSA level retail construction the local market conditions are key to understanding retail construction. MSA level retail sales changes, MSA level household formations, and MSA level short-term trends in retail construction provide 98% of the explanatory power in this analysis.

DATA APPENDIX

RS - Retail Supply - Gross Retail Construction. DRS - First differences of retail supply. [DRS.sub.t] = [RS.sub.t] - [RS.sub.t-1]

CB - Commercial Balances - non-residential loan balances outstanding for the United States and Other Areas as reported for commercial banks by the FDIC.

ACB - Commercial Balances adjusted to 1996 dollars. [DACB.sub.t] = [ACB.sub.t] - [DACB.sub.t-1]

PRIME - the prime interest rate as reported by the Federal Reserve Board. [DPrime.sub.t] = [Prime.sub.t] - [Prime.sub.t-1]

HH - Households are the number of families in a MSA. DHH - Net change in households each year within an MSA. [DHH.sub.t] = [HH.sub.t] - [HH.sub.t-1]

Sales - Retail sales by MSA, by year, adjusted to 1996 dollars. Dsales - First differences of sales. [Dsales.sub.t] = [Sales.sub.t] - [Sales.sub.t-1]

REFERENCES

Ambrose, B. W. & Nourse, H. O. (1993). Factors influencing capitalization rates, Journal of Real Estate Research, 8(2), 221-238.

Asea, P. K.& Blomberg, B. (1998). Lending cycles, Journal of Econometrics, 83, 89-128

Bails, D. & Peppers, L. (1993). Business Fluctuations Forecasting Techniques and Applications, (2nd Ed.). Englewood Cliffs, NJ: Prentice Hall.

Benjamin, J. D., Jud G. D. & Winkler, D. T. (1998a). A simultaneous model and empirical test of the demand and supply of retail space, The Journal of Real Estate Research, 16(1), 1-13.

Benjamin, J. D., Jud G. D. & Winkler, D. T. (1998b). The supply adjustment process in retail space markets, The Journal of Real Estate Research, 15(3), 297-307.

Born, W. L., (1988). A real estate market research method to screen areas for new construction potential, The Journal of Real Estate Research, 3(3), 51-62.

Bowerman, B. & O'Connel,l R. (1987). Time Series Forecasting Unified Concepts and Computer Implementation, Boston: Duxbury Press.

Clapp, J. M., Pollakowski, H. O. & Lynford, L. (1992). Intrametropolitan location and office market dynamics, Journal of the American Real Estate and Urban Economics Association, 20(2), 229-258.

Dokko, Y., Edelstein, R. H., Lacayo, A.J. & Lee, D. C. (1999). Real estate income and value cycles: A model of market dynamics, Journal of Real Estate Research, 18(1), 69-95. Federal Reserve Board, (1997), Bank Lending in the U.S., Economic Trends, Dec., 16.

Fuller, W. (1976). Introduction to Statistical Time Series, New York: John Wiley & Sons.

Granger, C.W.J. & Newbold, P. (1977). Forecasting Economic Time Series, New York: Academic Press Inc.

Grenadier, S. R. (1995). The persistence of real estate cycles, Journal of Real Estate Finance and Economics, 10(1), 95-119.

Henneberry, J. (1999). Convergence and difference in regional office development cycles, Urban Studies, 36(9), 1439-1466.

Hudson-Wilson, S. & Pappadopoulos, G. (1999). CMBS and the real estate cycle, Journal of Portfolio Management, 25(2), 105-112.

Hsiao, C. (1986). Analysis of Panel Data, Cambridge, MA: Cambridge University Press.

Kulish, N. (2001). Economy: Banks are toughening lending rules, Wall Street Journal, May 18, Sec A., 2.

Lawson, R. (1998). Area office, retail markets nearing top of cycle, Nashville Business Journal, 14(15), 39-42.

Makridakis, S., Wheelwright, S. & Hyndman, R. (1998). Forecasting Methods and Applications, (3rd Ed.), New York: John Wiley & Sons.

Mendenhall, W. & Sincich, T. (1993). A Second Course in Business Statistics: Regression Analysis, (4th Ed.), New York: Macmillan.

Nugent, J. (2001). No surprise: Banks are tightening syndicated loan standards, Bank Loan Report, Federal Reserve Board, March 5, 16(9), 4.

Reilly, F. (1985). Investment Analysis and Portfolio Management, (2nd Ed.), Chicago: The Dryden Press, 336-337.

Ritter, J. (1995). An outsiders guide to real business cycle modeling, Review (Federal Reserve Bank of St. Louis), 77(2), 49-61.

Rosen, K. T. (1984). Toward a model of the office building sector, Journal of the American Real Estate and Urban Economics Association, 12(3), 261-270.

Scott, P. & Judge, G. (2000). Cycles and steps in British commercial property values, Applied Economics, 32(10), 1287-1298.

Tsolacos, S. (1999). Retail building cycles: Evidence from Great Britain, Journal of Real Estate Research, 18(1), 197-218.

Vandaele, W. (1983). Applied Time Series and Box-Jenkins Models, New York: Academic Press Inc.

Wheaton, W. C. (1987). The cyclic behavior of the national office market, The Journal of the American Real Estate and Urban Economics Association, 15(4), 281-299.

Steven Moss, Georgia Southern University

Darrell Parker, Georgia Southern University

Steve Laposa, PricewaterhouseCoopers LLP

Table 1: Descriptive Statistics

Variable Mean S.D. Min Max

RS 1,864.82 1,506.23 98.90 11,105.50
ACB 186.93 78.98 93.97 315.99
PRIME 9.42 3.16 5.25 18.87
HH 724.23 626.74 99.13 3,331.78
SALES 29,803.93 28,348.95 6,377.87 212,629.70

Table 2: Analysis of variance for the residuals from equation 2.

Source SOS DOF

INDIV 48658409.29 57
TIME 71027746.38 20
JOINT 119686155.68 77
ERROR 780955504.23 1140
TOTAL 900641659.92 1217

Source Mean Square F Sign. Level

INDIV 853656.3035 1.246 0.107
TIME 3551387.3194 5.184 0.000
JOINT 1554365.6583 2.269 0.000
ERROR 685048.6879
TOTAL

Table 3: Regression Results

 Variable Coeff Std Error T-Stat Signif

Constant 146.867 53.068 2.767 0.005
DASALES{1} 0.032 0.016 2.021 0.043
DHHOLDS 11.549 4.585 2.519 0.012
DHHOLDS{1} 14.623 4.696 3.114 0.002
SUPPLY{1} 0.544 0.046 11.961 0.000
SUPPLY{2} 0.091 0.038 2.376 0.017
SUPPLY{6} 0.138 0.024 5.688 0.000
DACB 12.033 2.459 4.893 0.000
DACB{2} -7.889 1.980 -3.982 0.000
DPRIME{2} -56.847 13.175 -4.315 0.000
[R.sup.2] = 0.694

Table 4

Variable [R.sup.2] VIF

DACB .17 1.21
DPRIME .03 1.03
DHHOLDS .01 1.01
DASALES .24 1.31