Prepared by Ryan T. Kennelly, Economic Analyst Center for Business and Economic Research Lee Business School University of Nevada, Las Vegas October 2012 The Center for Business and Economic Research University of Nevada, Las Vegas 4505 S. Maryland Parkway Las Vegas, Nevada 89154-6002 (702) 895-3191 [email protected] http://cber.unlv.edu Copyright© 2012, CBER ABSTRACT: In this paper, we create new economic indexes for the metropolitan area of Las Vegas, Nevada. We first construct a coincident index, using two employment series of Las Vegas as proxies for the current state of the economy. We create the coincident index using the method outlined by The Conference Board (2001). After producing the coincident index, we construct a leading index. First, a rigorous set of criteria ranging from traditional, proven methods to modern day econometric techniques determine which data series lead the Las Vegas economy. The resulting series contain elements from the national, local, and neighboring states’ economies. Once again, we employ The Conference Board method intending to yield a suitable leading index, but aren’t satisfied with the results. Therefore, we develop a new process, using regression analysis to weight each variable. Furthermore, we restrict our data to different time periods producing three different sets of weights. Our most successful leading index comes from limiting the data to the most recent business cycle. This captures currently intact and relevant economic relationships, resulting in an accurate, reliable leading index. Table of Contents 1. Introduction ...................................................................................................................................1 2. Previous Work on Economic Indicators ...........................................................................................2 3. Constructing a Coincident Index for Las Vegas .................................................................................6 3.1 Conference Board Method ........................................................................................................7 4. Constructing a Leading Index for Las Vegas ................................................................................... 10 4.1 Criteria for Leading Economic Indicators .................................................................................. 11 4.2 Causality Testing ..................................................................................................................... 13 4.3 Constructing a Leading Index .................................................................................................. 17 4.3.1 Conference Board Method ............................................................................................... 17 4.3.2 Leading Index – A New Method ........................................................................................ 19 4.4 Leading Index Evaluation ........................................................................................................ 33 5. Summary of Findings .................................................................................................................... 35 6. Concluding Remarks ..................................................................................................................... 36 Bibliography .................................................................................................................................... 38 List of Figures Graph 1: Las Vegas Coincident Index ...............................................................................................9 Graph 2: Las Vegas Coincident Index vs. Las Vegas Real GMP ...........................................................9 Graph 3: Leading Index – Conference Board Method ..................................................................... 18 Graph 4: Leading Index – Conference Board Method ..................................................................... 18 Graph 5: Leading Index – All Data ................................................................................................. 24 Graph 6: Leading and Coincident Indexes – All Data ...................................................................... 25 Graph 7: Leading Index, Financial Crisis – All Data ......................................................................... 25 Graph 8: Leading Index – 1992 Data .............................................................................................. 28 Graph 9: Leading and Coincident Indexes – 1992 Data ................................................................... 29 Graph 10: Leading Index, Financial Crisis – 1992 Data .................................................................... 29 Graph 11: Leading Index – 2002 Data ............................................................................................ 31 Graph 12: Leading and Coincident Indexes – 2002 Data ................................................................. 32 Graph 13: Leading Index, Financial Crisis – 2002 Data .................................................................... 32 List of Tables Table 1: List of Economic Indexes ....................................................................................................5 Table 2: Augmented Dickey-Fuller Results ..................................................................................... 14 Table 3: Granger Causality Testing ................................................................................................ 16 Table 4: Regression Results – Intermediate Table .......................................................................... 21 Table 5: Regression Results – All Data ........................................................................................... 22 Table 6: Leading Index Regression Results – 1992 Data .................................................................. 27 Table 7: Leading Index Regression Results – 2002 Data .................................................................. 30 Table 8: Leading Index Weights – 2002 Data.................................................................................. 31 Table 9: Correlation Testing .......................................................................................................... 35 New Economic Indexes for Las Vegas, Nevada Ryan T. Kennelly* 1. Introduction Nowadays, businesses and policymakers use economic indexes to track and predict the economy. They do this to make informed decisions, ones that will best serve themselves or their constituents. At the national level, gross domestic product (GDP) represents the current state of the economy. However, GDP is only available quarterly, has reporting lags, and is revised frequently. Because of these problems, The Conference Board publishes two monthly indexes, one that coincides with and one that leads the U.S. economy. These are useful in assessing the current state of the economy and predicting where it is heading, an invaluable tool to policymakers and businesses. Lately, criticism has plagued these indexes, primarily the leading index for its performance after the end of Great Recession in June 2009.1 Nevertheless, many regional economists have attempted to replicate indexes similar to The Conference Board’s. The Conference Board employs the method developed by the Department of Commerce (1977, 1984). As with any method, there are pros and cons. We will later discuss these in detail. The method used by The Conference Board has been around a long time, and although successful, it doesn’t have much of an econometric appeal. As econometric theory advanced, other methods emerged that made use of the new techniques. Most notably, Stock and Watson (1989) created an index model using vector autoregressive (VAR) techniques. Their main assumption is that the current state of the economy was unobserved. Although advanced econometrically, their leading index * The author would like to thank Stephen P.A. Brown, Nasser Daneshvary, Rennae Daneshvary, Stephen M. Miller, and Reza Torkzadeh for helpful comments and insights. 1 The U.S. Leading Index published by the Conference Board has indicated a quick recovery since the end of the Great Recession; whereas, in reality, the economy is recovering at a slow pace. 1 for the United States performed poorly during the early 1990s. An overview of their method is in section 2. Although neither The Conference Board nor Stock and Watson method is perfect, many economists created variants of them, realizing the potential rewards of having information about the economy. Some created indexes for the nation, others for state and regional areas. In this paper, we do two things. First, we create a coincident index for the Las Vegas metro area using The Conference Board method. Secondly, we develop a new method for creating a leading index, combining the best qualities of the above methods into one. We first use economic theory to determine a list of data series that may lead our coincident index. Then we employ Granger causality testing to actually show which series lead the “current state.” Furthermore, we run a series of regressions that finalize the selections for our leading index. In our method, we are able to keep the traditional, proven methods of The Conference Board and infuse them with the most modern econometric techniques. By merging the old with the new, we create a leading index that will satisfy both the statisticians and the econometricians2, all while hopefully predicting the Las Vegas economy in an accurate manner. 2. Previous Work on Economic Indicators The roots of creating economic indexes can be traced back to Mitchell and Burns (1938). They were among the first to construct series that lead, lag, and coincide with the national economy. Their philosophy was to let the data speak for themselves, and not make underlying assumptions about which data series should lead the economy. Observation and careful statistical work were at the core of their methods. Not surprisingly, being the forerunners of this practice, their efforts were met with substantial criticism. Koopmans (1947) claimed it was “measurement without theory,” and reprimanded the 2 Statisticians are more in favor of letting data speak for themselves, while econometricians prefer to have a solid base of economic theory and then use statistics to validate. 2 authors for “observing and summarizing the cyclical characteristics of a large number of economic series” without a formal theoretical framework. At that time, Koopmans spoke on behalf of the Cowles Commission, located at the University of Chicago. The heart of the Cowles Commission was that economic theory suggests testable hypotheses. Their method was to use theory to build a model of the economy, test the suggested hypotheses, and subsequently reject or fail to reject the underlying theory. Koopmans’ criticism of Mitchell and Burns represented the Commission’s disdain for work that wasn’t solid in statistical and economic theory. After Mitchell’s death in 1948, Vining (1949) – a fellow economist at the National Bureau of Economic Research (NBER) – issued a rebuttal on behalf of his colleague. He claimed that Koopmans’ and the Cowles Commission’s methods put a “straightjacket on economic research.” He also argued that the current macroeconomic theory wasn’t strong enough to construct systems of equations that accurately represent the economy. In addition, he asserted that Koopmans’ approach overlooked the benefits of simple observation. These remarks started a feud between the “statistical economists” represented by NBER and the “econometricians” represented by the Cowles Commission. However, after the Great Depression and subsequent advances in statistical theory and econometric modeling, the empirical first methods of NBER were pushed aside for models based on Keynesian economics. It wasn’t until the 1970s that the debate between empirical-inductive and theoretical-deductive methods resurfaced. The models developed in the 1950s and 1960s couldn’t predict or explain the high inflation and high employment witnessed in the 1970s. The “statistical economists” argued that this unpredicted result was a symptom of the fundamental problems underlying the Keynesian model. In conjunction with the mounting criticism of the theoretical-deductive methods, Sims (1980a, 1982) created a process of tracking the economy with roots in Mitchell’s empirical-inductive methods. It consisted of a dynamic time series model, containing properties of an unrestricted system of equations 3 to summarize business cycle facts. This is more commonly known now as vector autoregressions. VAR models are used to capture linear dependence between a set of variables. By not focusing on individual coefficients within the model, the data were able to “speak for themself”. While Sims’ method had far more formal theoretical work than Mitchell’s, the underlying message remained the same – use careful statistical work with minimal theory to provide insights into economic behavior. Once again, Sims’ work was met with the same criticism as Mitchell’s. Sims’ work with VAR models leads us to the last major contribution in measuring business cycles, Stock and Watson (1989). In summary, the methodology is as follows. First, they construct a coincident index. Their main assumption is that “current state of the economy” is unobserved and reflected in several indicators. Each indicator is influenced by past values of the unobservable current state along with other forces. So for an indicator I, we have: (1) where S is the unobservable state, u the error term, and a a constant. In addition, it is assumed that S and u follow autoregressive processes: (2) (3) where et and zt are error terms. Combining (1), (2), and (3) for many monthly indicators produces a system of equations that can estimate the change in the unobservable state, . They estimate this model using maximum likelihood in standardized log differences, set the index equal to 100 at some point in time, and construct the coincident index with the estimated changes. To construct a leading index, indicators determined to lead the economy serve as the ’s, and the equation is changed to: and the system is used to predict future changes in the unobservable state. 4 While advanced econometrically speaking, it is not without faults. The Stock and Watson model for the United States completely missed the early 1990’s recession (the first recession after they constructed it), by not predicting or representing it. Furthermore, Phillips (1999) compared and contrasted the Stock and Watson model with Conference Board’s method, and concluded that The Conference Board’s method was superior in the regard of predicting turning points of the economy. While all these methods were developed to construct indexes for the nation, indexes can be useful at the local level as well. Many regional economies mirror the national economy to an extent, but fluctuate on their own. For example, the economy of Las Vegas is heavily influenced by tourism. While tourism plays a role nationally, it is much less significant than in Las Vegas. Following that thought, we would want to use variables and data that capture tourism in indexes for Las Vegas. The same process applies to other regional economies where manufacturing or oil may play a large role. Many economists have realized the value of indexes and created them for their own area. Table 1 is a list of some regional indexes. Under the “type of index” column, C stands for coincident and L for leading. In addition, the last column denotes the method used, either as Conference Board (CB), Stock and Watson (SW) or a variant (V). Table 1: List of Economic Indexes Author(s) Dua and Miller Phillips Gazel and Potts Balcilar et al. Crone Kurre and Riethmiller Crane Slaper Region Connecticut Texas Southern Nevada Nevada All 50 States Erie, PA Milwaukee, WI Indiana Year 1996 2005 1995 2010 2006 2005 1993 2009 Type of Index C,L C,L C,L C,L C,L L L L Method CB SW-V CB-V CB SW CB-V CB-V CB-V 5 Since Mitchell and Burns begin to track the economy with indexes in 1938, economists have developed many methods for “best” measuring or predicting changes in the economy. The most prevalent of these techniques either stem from The Conference Board or Stock and Watson. While The Conference Board method has a better track record and is more storied, the Stock and Watson method has great econometric appeal. As it stands, when constructing a coincident or leading index for a region, one must carefully weigh the pros and cons of each before settling on a particular process. 3. Constructing a Coincident Index for Las Vegas Before creating a leading index, we start by creating a coincident index to capture the current state of the economy. For the national and state level, GDP can serve this purpose. Even some metros – Las Vegas included – have gross metropolitan product (GMP) compiled by the Bureau of Economic Analysis (BEA), but it is only released on an annual basis. Since GMP is released infrequently, we need a new measure. Now the question becomes, how do we measure the economy? In essence, we are looking for series that consistently move with the local economy. For Las Vegas, the economy is best modeled using two proxy series: Las Vegas establishment employment and Las Vegas household employment.3 Both are reported monthly on a timely basis and collected from the Department of Employment, Training and Rehabilitation (DETR). Current Employment Statistics (CES) reports total nonfarm employment, whereas Local Area Unemployment Statistics (LAUS) reports household employment. For a metro area, data series that represent the current state of the economy are rare. Employment is generally considered a good proxy for how well the economy is performing even if it may lag a little at the national level. We also tried including the unemployment rate, but it made the index too volatile. 3 Employment is commonly used in coincident indexes at the state or regional level. Nationally, however, employment may be a lagging indicator. 6 We utilize the procedure outlined by The Conference Board (2001) to combine the two series into an index. We chose The Conference Board method over Stock and Watson for its proven record of accuracy not only at the national level, but also at the state and regional levels. This is built upon the method outline by the U.S. Department of Commerce (1977, 1984). 3.1 Conference Board Method First we compute the symmetric monthly changes for each variable. This ensures that positive and negative changes in a series receive the same weight. For a series in levels, this amounts to: ( ) ⁄ ( ) (1) where Xit is the data for month t of component i. The second step takes each component’s symmetric changes, multiplies them by their standardization factor4 (wi), and sums them together. ∑ Now we can compute the raw index. The series represents the symmetric changes in the index ( ), so we can use equation (1), replacing X with I and c with r. Then we have: ( ) 4 Standardization factors determine how monthly changes in each component contribute to the index. The standardization of each component is the inverse of the standard deviation of the symmetric changes. This allows each component to contribute to the monthly change in the index. Also, the standardization factors are normalized to sum to one. For us, the standardization factor on Las Vegas MSA Employment is 0.5162 while the factor on Las Vegas Household Employment is 0.4838. 7 Solving for yields: [ ] Step four is to rebase the index to the desired year. For us, this is 1982, for reasons we will explain later when developing our leading index. To do this, the index levels obtained in step 4 are multiplied by 100 then divided by the preliminary index level of the desired base year. Using this method, we create a coincident index for the Las Vegas economy. Two employment series (Las Vegas MSA and household employment) serve as proxies for the current state of the economy. The coincident index is illustrated in Graph 1 below. Our data constrain us to the dates from 1980 onward, and Las Vegas recessions5 as indicated by the index are shaded gray. The recessions in the early 1980s and 2000s are represented, along with the financial crisis of 2008. However, what about the recession in the early 1990s? During the 1990s, the Las Vegas economy was booming due to increased tourism. Consequently, the effect of the 1991-1992 national recession was hardly noticed locally, as represented by the index. For comparison, we can use Real GMP data from the BEA for Las Vegas. These are shown in Graph 2. Unfortunately, the BEA only has GMP for the Las Vegas metro area back to 2001, but even looking at this small time period, we can be comfortable with our choices to accurately represent the current state of the economy. One potential problem with this index could be volatility, but throughout the approximately 30 years, the recessions in the early 1980s and 2000s along with the financial crisis of 2008 are clearly defined. 5 Defining recessions we will be similar to the method used by NBER. While there is no fixed rule, generally five to six months of growth or decline will signify a trough or peak. However, the decision has to be a judgment call in the end. 8 Graph 1: Las Vegas Coincident Index 380 330 280 230 180 130 80 Graph 2: Las Vegas Coincident Index vs. Las Vegas Real GMP 400 93000 390 88000 380 370 83000 360 350 340 78000 Coincident Real GMP 73000 330 68000 320 310 63000 9 4. Constructing a Leading Index for Las Vegas The intuition of leading economic indicators is pretty straightforward. The idea being that although we know the current state of the economy through our coincident index, there must be measurable changes that occur in the economy before peaks and troughs. The hard part is identifying and representing those changes. But why not create a model of the economy, and then use it to predict the future? There are a variety of reasons, as outlined by Kurre and Riethmiller (2005). For starters, a full-blown econometric model takes significantly more time and resources than a leading index. Also, a large amount of data is needed to create a model. Much less data are needed to create an index, something especially important for a metro area, as well-maintained local data series are rare. Lastly, since a model is built upon past patterns, it can easily miss turning points. So not only is creating an index easier to do, it is also more reliable. In this section, we first address how to choose leading economic indicators with older methods, ones not based in modern econometrics. After obtaining that list of candidate variables, we use Granger causality testing to further narrow our choices. Following that, we make an attempt at combining the series using The Conference Board method. Unsatisfied with the results, we add another step, creating a new process. The methodology consists of running a series of regressions to finalize the variables to be used in our leading index and summing the coefficients on individual data series to weight each variable in the final leading index. 10 4.1 Criteria for Leading Economic Indicators If we can find series that lead the economy, The Conference Board method detailed above could give us a way to create an index. However, how do we choose the series?6 Kurre and Weller (1999) along with Kurre and Riethmiller (2005) outline the following criteria for choosing leading indicators. 1) Ocular Regression – This method relies on visually inspecting the data, to see if they lead the coincident index. Before modern day econometric techniques, this was the most prominent way to choose series. 2) Using Previous Economic Indicator Work – Sometimes the best place to start is by looking at the work of others. We have mentioned a few of the many indexes that have been created, and building upon another’s success can be very effective. 3) Economic Theory – Does this variable make sense in terms of the economy? In Las Vegas, tourism is a strong part of our local economy. Would it make sense that an index of Southern California’s economy (where a large part of the tourism comes from) may lead our economy? That is, if Southern California’s economy is doing well, should we expect to reap the benefits in Las Vegas? While helpful, the above criteria are very broad, and do little to narrow down what could be an extremely long list of candidates. Crane (1993) along with Kurre and Weller (1999) also suggest the following criteria: 1) Data Availability and Lags – We don’t want to use data that are not available frequently, or has just begun to be collected. Ideally, we want at least one full business cycle of data to 6 It is desirable to use more than one series to either predict or represent the economy. The more series that we use, the less likely that exogenous factors unrelated to the economy’s performance will distort our index (this is truer for leading indexes where we might have series exclusive to one sector of the economy. We want an index that represents the economy as a whole, not just one particular piece). 11 determine relationships within the economy. Additionally, we don’t want the data to be revised repeatedly or have large publication lags. Having to wait six months before publishing an index that leads the economy by six months is counterproductive. 2) Substitutability – Does the data series potentially substitute for some national series used in the U.S. Leading Economic Indicator (LEI – published by The Conference Board)? For many regional indexes, the U.S. LEI serves as a base model for their own index. The LEI contains average weekly initial claims for unemployment insurance. State-level initial claims for unemployment insurance would be a substitute for that variable. 3) Missed Turning Points – An ideal series will never miss a turning point, but such series are rare. 4) False Turning Points – A series with a lot of false signals may never miss a turning point, satisfying criterion number 3, but is of no use – having a similar effect to the “Boy Who Cried Wolf”. How would we know if we were actually headed into a recession if it were predicted every few months? 5) Volatility – This is an extension of criterion number 4, an extremely volatile (bumpy) series will make it harder to determine whether or not we are at a true turning point. 6) Length of the Lead – A four to six month lead is ideal, anything less doesn’t serve a purpose and anything more becomes vague as the consistency of the lead will become variable. 7) Consistency of the Lead – If the length of the lead varies, it is difficult to predict when the economy will reach its peak or trough. However, if the indicator always leads by the same amount, we are able to make a good prediction. 12 4.2 Causality Testing All the criteria that we discussed thus far can be classified as traditional methods (none rely on econometric tests). These were used before Stock and Watson paved the way on using econometrics to create indexes. For us, the list of possible leading economic indicators was still large after eliminating many possibilities using traditional criteria. It was clear that we needed some way to further evaluate indicators, and that we should look to econometrics for help. Granger causality testing became our solution. Granger causality testing determines whether one variable is useful in predicting future values of another. The first step is to check whether the series is integrated or stationary. If a series is integrated, a shock to it will be permanent, whereas a stationary series will eventually return to its trend. As per common econometric practice, an Augmented Dickey-Fuller (ADF) test was used. The format of an ADF test is as follows: for a data series y, where is a constant, is the error term, and p is the lag order of the autoregressive process. The lag order can be found using an Akaike Information Criterion (AIC).7 The null hypothesis of an ADF is while the alternative is . If we fail to reject, the data series is said to have a unit root (is integrated). We first run the ADF test with the series in levels, and if we fail to reject the null, the test is done after first differencing the series. We continue differencing until the series is stationary. In Table 2 are results of selected variables.8 7 An AIC test measures the relative goodness of fit of a model by determining the tradeoff between bias and variance in model construction. The general form of the test is: AIC = 2k – 2ln(L), where k is the number of parameters in the model, and L is the maximized value of the likelihood function. So for each n = 1…p, we have an AICn. The min{AICn} denotes the preferred model where n gives the lag order of the autoregressive process. 8 A large number of national and local data series were tested after clearing the criteria in section 4.1. However, to avoid long lists of results, only selected variables are shown. 13 Table 2: Augmented Dickey-Fuller Results Variable Levels 1st Difference AZ -2.79* - CA - -2.65* Coincident - -5.044*** M2 - -4.20*** McCarran - -28.21*** Occu - -18.51*** S&P 500 - -15.98*** Taxable Sales - -10.05*** Visitor - -19.48*** *, **, and *** denote significance at the 10%, 5%, and 1% level, respectively All data are reported monthly and are seasonally adjusted. AZ and CA are the Arizona and California Leading Indexes, respectively, as compiled by the Philadelphia Fed9. Coincident is the coincident index created previously in this paper. M2 is collected by the Board of Governors of the Federal Reserve System. McCarran is the total passengers enplaned and deplaned from McCarran airport as reported by itself. Occu is the hotel/motel occupancy rate in Las Vegas provided by Las Vegas Convention and Visitors Authority. SP500 is the S&P 500 Index, which is provided by Standard and Poors. Tax is Clark County taxable sales as reported by the Nevada Department of Taxation, and Visitor is the Las Vegas visitor volume collected by Las Vegas Convention and Visitors Authority. All data predate 1982 except for the Arizona and California Leading Indexes, making analysis from that year forward intuitive. As seen above, AZ is stationary in levels, whereas the remaining variables are stationary in first differences. 9 The Arizona and California leading indexes are comprised of state-level housing permits, state initial unemployment insurance claims, the interest rate spread between the 10-year Treasury bond and the 3-month Treasury bill, and delivery times from the Institute for Supply Management (ISM) manufacturing survey. 14 The second step of Granger causality testing is to run the following regression: ∑ where is the coincident index and ∑ is the economic indicator we are testing for month t in stationary form. For our purposes, n was chosen to be six, as a six-month lead is most desirable.10 After obtaining the regression results, an F-test is run on all the chosen indicators lags to determine significance. If an indicator’s lags are statistically significant, it is said to Granger cause the coincident index. This tells us that the indicator is useful in predicting future movement of the coincident index – confirming that it is a leading economic indicator. Table 3 shows the data series that were significant in predicting future values of the coincident index. 10 A six-month lead provides time for businesses and policymakers to make adjustments but lacks the problems of longer leads. Generally, as you increase the predicted lead time, the variability of that lead increases as well. A sixmonth time period will consistently lead around six months, whereas a twelve month lead might result in an actual lead of between seven months and fifteen months. In addition, tests were run for n=3 all the way up to n=12 with similar results in terms of relative significance. 15 Table 3: Granger Causality Testing Variable F-Statistic Probability AZ 5.01433 0.00006*** CA 5.18454 0.00004*** M2 1.79475 0.0991* McCarran 5.52141 0.00002*** Occupancy 4.36805 0.0003*** S&P 500 5.46449 0.00002*** Taxable Sales 8.96316 0.00*** Visitor Volume 4.098 0.0005*** 4.98336 0.00005*** 1.71645 0.1161 1.287981 0.2619 Gaming 11 Housing Permits 12 Convention Attendance 13 *, **, and *** denote significance at the 10%, 5%, and 1% level, respectively 11 The numbers shown for total monthly gaming revenue in Clark County are actually for reverse causality. So growth in the coincident index causes growth in gaming revenue. Causality from gaming revenue to the coincident index was insignificant. 12 Clark County housing permits, monthly. 13 Las Vegas convention attendance, monthly. 16 4.3 Constructing a Leading Index After the methods described in 4.1 and 4.2, we have a list of variables that lead the “current state” of the economy. In this section, we first make an attempt to create a leading index using The Conference Board method, but fail. We then develop a new method, consisting of multiple “hollowedout” regressions in three different time periods, with the final result being our leading index for Las Vegas, Nevada. 4.3.1 Conference Board Method Our first thought after identifying leading economic indicators was to use The Conference Board method once again. As previously mentioned, this method has a history of accuracy, both at the national and regional level. The variables used for this leading index are: Arizona and California Leading Indexes M2 money supply Total McCarran enplaned and deplaned passengers Las Vegas hotel and motel occupancy rate S&P 500 Index Clark County taxable sales Las Vegas visitor volume The resulting index is portrayed in Graph 3 (shaded areas are recessions indicated by the coincident index). Also, Graph 4 shows the leading index along with the coincident index. 17 Graph 3: Leading Index – Conference Board Method 435 385 335 285 235 185 135 85 Graph 4: Leading Index – Conference Board Method 385 435 335 385 285 335 285 235 235 185 185 Leading Coincident 135 85 135 85 18 There are a number of flaws with this leading index, the most prominent being that there is no lead before recessions. The index drops after the recession has begun for both the early 2000’s recession and the 2008 financial crisis – a quality of a coincident index. Having no early warning about upcoming recessions defeats the purpose of a leading index. 4.3.2 Leading Index – A New Method The results from The Conference Board method suggest that we need a new way to create the leading index. Traditional techniques and Granger causality testing have provided a list of variables, but we must combine them in a different fashion. The method proposed here uses regression analysis, along with subsequent F-tests to determine the significance of each variable. Before we run the regression, each variable is indexed with December 1981 = 100.14 Now we run the following regression: ∑ ∑∑ where Coint is the coincident index for month t, X is a vector of all the significant economic indicators in Table 3, and c a constant. In addition, j and i represent the indicator and month, respectively. For this particular index, n = 4 and m = 6. The reasoning behind this is that we are looking for a four to six-month lead on the coincident index.15 Note that this regression is very similar to a Granger Causality Test. We deem this a “hollowedout” regression - referring to the first few lags, which are included in the Granger test that we omit from our equation. This “hollowing out” technique tells us which variables are able to lead our coincident index in the desired time period of four to six months. Next a series of F-tests were run on each triplet of variables. Of the groups that are not significant, the least significant was dropped from the equation, and the regression run again. Table 4 is 14 15 The Arizona and California Leading Indexes begin in 1982, making this date an intuitive choice. Other values of n and m were tried, ranging from 2 to 8, but using these values provided the best fit. 19 an intermediate table; in this case, we would eliminate Occu from the equation and run the regression again. The same process was repeated until only significant variables remained - the final results are shown in Table 5. 20 Table 4: Regression Results – Intermediate Table Variable Coefficient t-Statistic Joint Significance COINCIDENT(-4) COINCIDENT(-5) COINCIDENT(-6) 0.930026 0.190887 -0.179986 7.194023 1.140515 -1.414815 0.0000 AZ(-4) AZ(-5) AZ(-6) 0.389606 -0.232628 -0.189127 1.442082 -0.445364 -0.711194 0.0000 CA(-4) CA(-5) CA(-6) 1.345443 -1.707921 0.441561 3.55211 -2.365805 1.159865 0.0000 M2(-4) M2(-5) M2(-6) -0.109994 0.162127 -0.091268 -0.879361 0.833961 -0.729266 0.0000 MCCARRAN(-4) MCCARRAN(-5) MCCARRAN(-6) 0.021539 -0.000153 -0.038171 1.402653 -0.009087 -2.586364 0.0327 OCCU(-4) OCCU(-5) OCCU(-6) 0.049161 -0.026387 0.004026 0.770945 -0.371538 0.065449 0.879 SP500(-4) SP500(-5) SP500(-6) 0.013199 -0.000602 -0.016321 2.708033 -0.081608 -3.31846 0.0000 TAX(-4) TAX(-5) TAX(-6) 0.02475 -0.006748 -0.001783 2.912818 -0.750408 -0.201218 0.0021 VISITOR(-4) VISITOR(-5) VISITOR(-6) 0.000118 0.015212 0.036079 0.004423 0.481762 1.357567 0.035 C -4.947006 -0.778915 R-squared Adjusted R-squared 0.999623 0.999592 21 Table 5: Regression Results – All Data Variable Coefficient t-Statistic COINCIDENT(-4) 0.932064 7.267581 COINCIDENT(-5) 0.178844 1.076606 COINCIDENT(-6) -0.182983 -1.482249 AZ(-4) 0.400223 1.490975 AZ(-5) -0.246003 -0.473379 AZ(-6) -0.183583 -0.693205 CA(-4) 1.317544 3.513204 CA(-5) -1.689709 -2.350372 CA(-6) 0.451552 1.19594 M2(-4) -0.115677 -0.932516 M2(-5) 0.168258 0.872001 M2(-6) -0.092467 -0.742446 MCCARRAN(-4) 0.024172 1.641814 MCCARRAN(-5) 0.000163 0.009824 MCCARRAN(-6) -0.037636 -2.608692 SP500(-4) 0.013256 2.732441 SP500(-5) -0.000815 -0.111104 SP500(-6) -0.016195 -3.306797 TAX(-4) 0.025071 2.978394 TAX(-5) -0.006247 -0.699496 TAX(-6) -0.001385 -0.158401 VISITOR(-4) 0.0119 0.567229 VISITOR(-5) 0.004524 0.198381 VISITOR(-6) 0.0359 1.751307 C -1.870116 -1.532513 R-squared 0.999622 Adjusted R-squared 0.999594 Joint Significance 0.0000 0.0000 0.0000 0.0000 0.0253 0.0000 0.0005 0.0285 22 In this regression, all the available data were used (December 1981 – May 2011). Now we must determine the weight of each variable for the index. This was done by taking the sum of the absolute values of the coefficients of an individual triplet and dividing by the sum of the absolute values of all the coefficients. ∑ | | ⁄ ∑ ∑ | | For month t, each series was multiplied by its weight and then summed to yield the final leading index. ∑( ) Before continuing, an explanation of the weights is warranted. We were faced with a few choices for constructing the weights. For example, we could have not taken absolute values and allowed for some components to have negative weights as opposed to the above. The intuition behind our method is that we want to know the magnitude of the effect for each component. Although each component is indexed to 100 in December 1981, they grow at different rates. Weighting by the magnitudes automatically adjusts for the different growth rates when constructing the leading index. 23 4.3.2.1 Leading Index: Full Sample The final result is displayed in Graph 5, with recessions indicated by the coincident index shaded in grey. In Graph 6, we see both the coincident and leading index together to get a better picture. Upon first glance, it seems that the leading index serves all of its purposes well. There is a significant lead on all peaks and troughs in all three recessions; it’s not volatile and is consistent. However, looking at the very end of the index, we can see that the leading index begins to rise quickly, whereas the coincident is struggling. This is pictured in Graph 7. Although this leading index is a good fit for the previous recessions, it doesn’t seem to represent the last couple of years well. This problem leaves us unsatisfied and looking for a new solution. Graph 5: Leading Index – All Data 585 485 385 285 185 85 24 Graph 6: Leading and Coincident Indexes – All Data 585 Leading Index 385 Coincident Index 335 485 285 385 235 285 185 185 135 85 85 Graph 7: Leading Index, Financial Crisis – All Data 612 390 607 602 380 597 Leading Index 370 592 587 360 582 350 577 Coincident Index 572 340 25 4.3.2.2 Leading Index: 1992 Data Because the problem occurs within the most recent data, perhaps we are putting too much weight on older data. After every business cycle, the economy restructures itself. So it’s possible by constructing the leading index with data back to 1982, we are capturing relationships that no longer exist. To combat this, we use the same method two more times. The first time we use data from April 1992 (the end of the early 90’s recession as reported by NBER) until the most recent. The second time we use data from December 2001 (the end of the early 2000’s recession as reported by NBER) onward. This is an attempt to capture the most recent and intact economic relationships among the variables. For simplicity, we will call these variations the 1992 and 2002 leading indexes, respectively. Note that in both cases, we have data that span at least one business cycle, the minimum for evaluating interactions among economic variables. First, we look at the 1992 leading index in the same format as we looked at the all data leading index. Regression results are reported in Table 6, and Graphs 8, 9, and 10 display the results. 26 Table 6: Leading Index Regression Results – 1992 Data Variable Coefficient t-Statistic COINCIDENT(-4) 0.679876 4.753338 COINCIDENT(-5) 0.107933 0.585553 COINCIDENT(-6) -0.005852 -0.04081 AZ(-4) 0.942234 2.32879 AZ(-5) -0.408533 -0.52998 AZ(-6) -0.521127 -1.32916 CA(-4) 1.096243 2.355805 CA(-5) -1.252997 -1.44989 CA(-6) 0.367067 0.786028 M2(-4) -0.246002 -1.92106 M2(-5) 0.292813 1.509301 M2(-6) -0.097663 -0.75651 MCCARRAN(-4) 0.04364 2.778842 MCCARRAN(-5) 0.022084 1.226023 MCCARRAN(-6) 0.000929 0.057325 SP500(-4) 0.006776 1.364842 SP500(-5) -0.001418 -0.19346 SP500(-6) -0.01501 -2.93151 VISITOR(-4) -0.011234 -0.44404 VISITOR(-5) -0.04979 -1.81817 VISITOR(-6) -0.030394 -1.17592 C -17.75056 -6.81302 R-squared 0.999205 Adjusted R-squared 0.999127 Joint Significance 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0012 27 Note that in this version of leading index, Tax is no longer included. We have no economic intuition or theory as to why this is, other than that the economy of Las Vegas was restructured in the early 1990s. The 1992 leading index still has all the good qualities of a reliable leading index, such as smoothness and consistence. Although the end of the financial crisis is still not modeled perfectly, it is definitely better than its predecessor as the increase at the end is less steep. Graph 8: Leading Index – 1992 Data 585 485 385 285 185 85 28 Graph 9: Leading and Coincident Indexes – 1992 Data Leading Index 385 585 Coincident Index 335 485 285 385 235 285 185 185 135 85 85 Graph 10: Leading Index, Financial Crisis – 1992 Data 638 390 633 628 380 623 370 618 Leading Index 613 360 608 350 603 Coincident Index 598 340 29 4.3.2.3 Leading Index: 2002 Data Although the 1992 leading index models the end of the financial crisis better, there is still room for improvement. Constructing the 2002 index allows us to capture the most recent economic relationships. Table 7 contains the regression results; Table 8 shows the weights, while Graphs 11, 12 and 13 give us closer looks.16 Table 7: Leading Index Regression Results – 2002 Data Variable Coefficient t-Statistic Joint Significance COINCIDENT(-4) COINCIDENT(-5) COINCIDENT(-6) 0.435439 0.055073 0.146041 2.405474 0.2479 0.767815 0.0000 AZ(-4) AZ(-5) AZ(-6) 1.191803 -0.939418 -0.008065 1.986601 -0.823958 -0.013734 0.0000 CA(-4) CA(-5) CA(-6) 0.425271 -0.280396 -0.291863 0.746419 -0.257141 -0.460029 0.0000 MCCARRAN(-4) MCCARRAN(-5) MCCARRAN(-6) 0.04229 0.00289 0.009558 3.012942 0.200433 0.697116 0.008 SP500(-4) SP500(-5) SP500(-6) 0.009723 -0.002149 -0.009086 1.57964 -0.264387 -1.614556 0.0961 C 7.96521 0.67018 R-squared Adjusted R-squared 0.995629 0.994992 16 In addition to the “hollowed-out” approach, we also ran the model including lags one through three of the coincident index. This allowed the coincident index to fully explain itself within the time period, and subsequently eliminated the S&P 500 Index from our final model. This is not surprising, considering the S&P 500 Index was only marginally significant in the “hollowed-out” model. However, we feel the “hollowed-out” approach is superior, since when given the data at time period t (looking to forecast four to six months into the future), lags one through three of the coincident index aren’t available. 30 Table 8: Leading Index Weights – 2002 Data Variable AZ CA McCarran SP500 Total Weight 0.66 0.31 0.02 0.01 1 In this version of the leading index, even more variables are eliminated. One interesting note is that M2 was dropped from the index. Although M2 has been a strong leading economic indicator in the past, we can probably attribute its lack of appearance to the Federal Reserve beginning to pay interest on banks’ reserves, causing a huge spike in the money supply. This spike causes M2 to not be a good indicator, and therefore was eliminated from the regression due to insignificance. We can infer that the strong past relationship of M2 and economy holds it in the previous two leading indexes. Graph 11: Leading Index – 2002 Data 685 585 485 385 285 185 85 31 Graph 12: Leading and Coincident Indexes – 2002 Data Leading Index 685 385 585 335 Coincident Index 485 285 385 235 285 185 185 135 85 85 Graph 13: Leading Index, Financial Crisis – 2002 Data 390 725 715 380 705 370 Coincident Index 695 360 685 Leading Index 350 340 675 665 32 This final version of the leading index provides everything we are looking for. Just like the previous two it is smooth and predicts all turning points with a consistent lead of 4-6 months. On top of that, it models 2010-2011 well, not quickly increasing but showing a long, slow recovery, as seen in Graph 13. By limiting our data to the most recent business cycle, we are able to capture the relevant economic relationships needed to produce a respectable leading index. 4.4 Leading Index Evaluation Now that we have found a useful leading index, let’s revisit the criteria laid out earlier. The following criteria are considered traditional methods, as they don’t rely on modern econometric techniques. 1) Data Availability and Lags – Because we choose series for the leading index with these criteria in mind, they are satisfied. 2) Substitutability – We have a few components in our leading index that are potential substitutes for components in the U.S. LEI. Both the Arizona and California leading indexes have state-level initial unemployment insurance claims and new building permits. The U.S. LEI includes these variables at the national level. Also, our index has the S&P 500 Index, common to the U.S. LEI as well. 3) Missed Turning Points – Our index doesn’t miss any turning points by leading all peaks and troughs for recessions. 4) False Turning Points – Defining a turning point by three to four consecutive months of increase or decline in the index, the index doesn’t give false turning points. 5) Volatility – Looking at Graph 11 we can see that the leading index is smooth, not jagged. 33 6) Length of the Lead – The index leads peaks by an average of 5 months and troughs by an average of 6.5 months. Earlier we commented that a 4-6 month lead is most desirable. 7) Consistency of the Lead – The length of the lead for peaks is very consistent, with each one being exactly five months. However, the lead on troughs is a bit more variable. For the early 80’s recession, we have a lead of 4 months. The lead for the trough of the early 2000 recession is 3 months. For the ‘08 financial crisis, we have almost a double-dip like structure in the coincident index. An argument can be made for the true trough being either in Nov 2010 or June 2011. Either way, the leading index models this period well, with a trough in Dec 2009 (lead of 11 months) followed by a period of stagnate growth resulting in another trough in Oct 2010 (lead of 8 months). Based on traditional methods, our leading index satisfies nearly all of the criteria. The only blemish occurs in the consistency of leading troughs, specifically during the financial crisis. However, looking at the period from 2007 until now, in Graph 13, we still conclude that the leading index provides an accurate forecast of the coincident index. Continuing the trend of combining traditional methods with econometrics, we evaluate the leading index based on correlation testing as well. Taking the first differences of the coincident and leading indexes, we run a correlation test between lags of the first differenced leading index and the first differenced coincident index. We do this for three different time periods. The first time period consists of all data, from 1982 to 2011. Our highest correlation coefficient is 0.49, indicating a relevant but only moderate correlation between the indexes. We then proceed in a manner similar to creating our leading index, restricting the data to the two most recent business cycles, and then only the most recent. Our correlation coefficients increase in both cases, reaching a high of 0.74 for our 3rd and 4th lag in the most recent data. 34 Is it surprising that the correlation coefficients increased drastically? Recall that when we created our leading index, we used data restricted to the most recent business cycle. Our intuition was that this captured the most recent and relevant economic relationships and will therefore be a better predictor in the future. A critique of our leading index could be the low correlation coefficient attributed to a more loose connection in the past. However, a leading index is not created to model the past; it is created to help predict the future, and that is where ours excels. Table 9: Correlation Testing ∆CI ∆CI ∆CI ∆LI(-3) 0.51 0.67 0.74 ∆LI(-4) 0.52 0.68 0.74 ∆LI(-5) 0.49 0.65 0.69 ∆LI(-6) 0.48 0.64 0.67 ∆LI(-7) 0.48 0.63 0.65 Jan '82 - Aug '11 Apr '92 - Aug '11 Dec '01 - Aug '11 5. Summary of Findings Our first task was to create a coincident index for Las Vegas to represent the “current state of the economy.” Since employment is generally used in regional indexes as a measure of the well-being of the economy, we combined two employment series of Las Vegas and created an index using The Conference Board method. We chose this method over Stock and Watson’s after examining their results and the literature comparing the two methods (which was in favor of The Conference Board). After constructing our coincident index, we were faced with many choices for constructing our leading index. Instead of pursuing a purely empirical-inductive or theoretical-deductive approach, we chose a path that allowed us to infuse the traditional, proven methods with modern day econometric techniques, such as Granger causality testing, to discover leading economic indicators. Following our approach from the coincident index, we employ The Conference Board method again, but are left with unsatisfactory results. 35 Because our previous results suggested we needed a new method, we ran a series of “hollowedout” regressions, enabling us to truly see what combination of leading indicators best predicts the coincident index. After calculating weights based upon our last regression, we have a vastly improved leading index, but are not convinced by its performance, particularly in the last few years. By constricting the time periods of data used in the regressions, we are able to capture the structural changes in the economy that happen over time, and redefine what indicators do and don’t belong in the index. Our final restriction to only the most recent business cycle provides an index unmatched by any other period. This is our final leading index. 6. Concluding Remarks Since Burns and Mitchell’s work in 1938, there have been countless debates on how to properly construct indexes that represent and predict the economy. First, the empirical-inductive approaches of NBER prevailed, followed by their disappearance due to advanced models based on Keynesian theory. But when the unpredicted high inflation and high unemployment arose during the 1970s, some claimed the full-blown econometric models were dead and turned away from those theoretical-deductive methods. Then Stock and Watson developed their own method, using modern-day econometric techniques never before seen in this field, fueling the debate between the two approaches. In this paper, we attempt to find a harmonious middle ground in creating new economic indexes for the metropolitan area of Las Vegas, Nevada. We first construct a coincident index, using two employment series of Las Vegas as representatives for the current state of the economy. We combine these series using the method outlined by The Conference Board, developed from the statistical approaches at NBER and the Department of Commerce. After producing the coincident index, we construct a leading index. A rigorous set of criteria ranging from traditional, proven methods to modern day econometric techniques determine which 36 series lead the Las Vegas economy. After finding our candidate variables, we employ The Conference Board method with little success. Therefore, we develop a new method of creating a leading index, using regression analysis to weight each data series. Restricting our data to specific time periods produces three different sets of indicators and weights. Limiting data to the most recent business cycle captures the currently intact and relevant economic relationships, resulting in an accurate, reliable leading index. As with all economic indexes, we can’t be entirely sure that it will perform as well as desired. The fate could be much the same as Stock and Watson’s index, failing to predict the first recession after its construction. If that were the case, each part of the method would have to be revisited and reevaluated. That being said, we feel that given the economic theory and econometric practices available now, that this method represents the best possible way to construct an economic index. The stance taken here to use both empirical-inductive and theoretical-deductive methods is much like that of Hoover’s (1994). He drew an analogy between economics and astronomy. Astronomers don’t point their telescopes in random directions hoping to make a discovery. Although a telescope is an extremely valuable tool, without knowing where to look, it becomes useless. It is through observations and careful calculations that an astronomer knows where to look. Economics is much the same. Observation and careful statistical work, such as that done by Mitchell and Burns, can tell us where to look. Only then can we use our telescope, econometrics, to discover and create something new. 37 Bibliography Auerbach, A.J. (1982), The Index of Leading Economic Indicators: “Measurement Without Theory,” Thirty Five years Later, The Review of Economics and Statistics, 64, 589-595. Balcilar, M., Gupta, R., Majumdar, A. and Miller, S.M. (2010), Forecasting Nevada Gross Gaming Revenue and Taxable Sales Using Coincident and Leading Employment Indexes, University of Connecticut Working Paper. Burns, A.F. and Mitchell, W.C. (1938), Statistical Indicators of Cyclical Revivals, National Bureau of Economic Research Bulletin 69, New York. Reprinted as Chapter 6 of G.H. Moore, ed. Business Cycle Indicators. Princeton: Princeton University Press. 1961. Crane, S.E. (1993), Developing a Leading Indicator Series for a Metropolitan Area, Marquette University, Economic Development Quarterly, 7, 267-281. Crone, T.M. (1988), Using state indexes to define economic regions in the U.S., Journal of Social and Economic Measurement, Special Issue on Regional economic Models, 25. Crone, T.M. (2000), A New Look at the Economic Indexes for the States in the Third District, Federal Reserve Bank of Philadelphia. Crone, T.M. (2006), What a New Set of Indexes Tells Us About State and National Business Cycles, Federal Reserve Bank of Philadelphia. Dua, P. and Miler, S.M. (1996). Forecasting and Analyzing Economic Activity with Coincident and Leading Indexes: The Case of Connecticut. Journal of Forecasting, 15, 509-526. Gazel, R.C. and Potts, R.D. (1995), Southern Nevada Index of Leading Economic Indicators, Center of Business and Economic Research at University of Nevada, Las Vegas. Gazel, R.C. (1995), The Southern Nevada Index of Leading Indicators: A Good Performance in 1995, Center of Business and Economic Research at University of Nevada, Las Vegas. Gazel, R.C., Potts, R.D. and Schwer, R.K. (1997), Using a Regional Index of Leading Economic Indicators to Revise Employment Data, Center of Business and Economic Research at University of Nevada, Las Vegas. 38 Koopmans, T.C. (1947), Measurement Without Theory, The Review of Economics and Statistics, 29, 161-172. Kurre, J.A. and Weller, B.R. (1999), Is Your Economy Ready to Turn on You? Constructing a Leading Index for a Small Metro Area, Penn State Erie. Kurre, J.A. and Riethmiller, J.J. (2005), Creating an Index of Leading Indicators for a Metro Area, Economic Research Institute of Erie. Lucas, Robert E., Jr (1976), Econometric Policy Evaluation: A Critique, The Phillips Curve and Labor Markets, Carnegie-Rochester Conference Series on Public Policy 1, eds. Brunner, Karl and Allan H. Meltzer, Amsterdam: North-Holland, 19-46. Niemira, M.P. and Klein, P.A. (1994), Forecasting Financial and Economic Cycles, New York: John Wiley. Phillips, K.R. (1999), The Composite Index of Leading Economic Indicators: A Comparison of Approaches, Journal of Economic and Social Measurement, 25. Phillips, K.R. (2005), A New Monthly Index of the Texas Business Cycle, Federal Reserve Bank of Dallas – San Antonio Branch, Journal of Economic and Social Measurement, 30, 317-333. Shiskin, J. (1961), Signals of Recession and Recovery, National Bureau of Economic Research. Simkins, S. (1999), Measurement and Theory in Macroeconomics, North Carolina A&T State University. Sims, Christopher A. (1980a), Comparison of Interwar and Postwar Business Cycles: Monetarism Reconsidered, American Economic Review 70:2, 250-257. Sims, Christopher A. (1980b), Macroeconomics and Reality, Econometrica 48:1, 1-48. Sims, Christopher A. (1980c), Scientific Standards in Econometric Modeling, Discussion Paper No. 82-160, Center for Economic Research, Department of Economics, University of Minnesota. Sims, Christopher A. (1982), Policy Analysis with Econometric Models, Brookings Papers on Economic Activity 1, 107-152. 39 Slaper, T.F. and Cohen, A.W. (2009), The Indiana Leading Economic Index: Indicators of a Changing Economy, Indiana Business Research Center, Indiana Business Review. Stock, J.H. and Watson, M.W. (1989), New Indexes of Coincident and Leading Economic Indicators, National Bureau of Economic Research Macroeconomics Annual, 351-394. The Conference Board (2001), Components and Construction of Composite Indexes, Business Cycle Indicators Handbook, 47-63 United States Department of Commerce (1977), Composite Indexes of Leading, Coincident, and Lagging Indicators: A Brief Explanation of Their Construction, Handbook of Cyclical Indicators, A Supplement to the Business Conditions Digest, Bureau of Economic Analysis, 73-76. United States Department of Commerce (1984), Composite Indexes of Leading, Coincident, and Lagging Indicators: A Brief Explanation of Their Construction, Handbook of Cyclical Indicators, A Supplement to the Business Conditions Digest, Bureau of Economic Analysis, 65-70. Vining, Rutledge (1949), Koopmans on the Choice of Variables to Be Studied and the Methods of Measurement, The Review of Economics and Statistics 31, No. 2, 77-86. Vining, Rutledge (1951), Economic Theory and Quantitative Research: A Broad Interpretation of the Mitchell Position, The American Economic Review 41, No. 2, Papers and Proceedings of the Sixty-third Annual Meeting of the American Economic Association, 106-118. 40 An affirmative action/equal opportunity institution.
© Copyright 2026 Paperzz