Nowcasting US GDP: The role of ISM Business Surveys

Transcription

1 Nowcasting US GDP: The role of ISM Business Surveys Kajal Lahiri George Monokroussos Department of Economics - University at Albany, SUNY y March 11 Abstract We study the role of the well-known monthly di usion indices produced by the Institute for Supply Management in nowcasting current quarter US GDP growth. We investigate their marginal impact on these nowcasts when large unbalanced (jagged edge) macroeconomic data sets are used in real time to generate them. We nd some evidence that these ISM indices can be helpful in improving the nowcasts in the beginning of the month when new ISM information becomes available ahead of other monthly indicators. JEL Classi cation: C33, C53. KEYWORDS: Nowcasting, Forecasting, ISM, PMI, factor models, Kalman lter. We thank Domenico Giannone and David Small for making available the data and code that were used in Giannone, Reichlin, and Small (8) and Giannone, Modugno, Reichlin, and Small (1). We also thank Lucrezia Reichlin, our discussant Urs Mueller, and other participants at the 7th International Workshop on Flash indicators in Verbier for useful comments, and Cagri Baydil and Herbert Zhao for research assistance. George Monokroussos thanks the Research Department of the Federal Reserve Bank of Boston for its hospitality. y Lahiri: Department of Economics, University at Albany, State University of New York, 14 Washington Avenue, Albany, NY 1. klahiri@albany.edu. Monokroussos: Department of Economics, University at Albany, State University of New York, 14 Washington Avenue, Albany, NY 1. gmonokroussos@albany.edu 1

2 1 Introduction: ISM variables and their role in nowcasting The Institute for Supply Management (ISM) produces a well known monthly report on business conditions based on rms responses to a questionnaire it sends out every month. Business executives are asked about their rms s production, employment, inventory levels, etc., during the preceding month and their responses are then used to construct di usion, or summary indices of business activity. These di usion indices can be useful tools in assessing the current state of various sectors and of the economy in general. Most prominently, their Purchasing Managers Index (PMI) for manufacturing sectors is a very well known economic indicator that always receives publicity and is closely watched by those aiming to get a head start in forecasting the economy s movements in real time. As Koenig () discusses, the PMI, and the other ISM di usion indices in general, have two main advantages: First, and most importantly, it s timeliness: New ISM manufacturing information comes out on the rst business day of every month (and the ISM non-manufacturing variables shortly afterwards) and it contains the reports based on the previous month s questionnaires. No other economic variable of such importance becomes available rst thing every month, on a consistent basis 1. Second, ISM data are typically subject to at most small revisions, presumably because of their nature as survey responses. As such, they preserve most of the real-time nature that is crucial in many estimation and forecasting exercises (see, inter alia, Orphanides (), Koenig et al (3), etc.) Furthermore, ISM variables and the PMI in particular have been shown to have forecasting power for GDP and the business cycle. For instance, Dasgupta and Lahiri (1993) have shown that the PMI can be useful in forecasting GDP changes; similar results on the performance of the PMI as a leading indicator include Klein and Moore (1991), Dasgupta and Lahiri (199), Kau man (1999), Koenig () and Lindsey and Pavur (5). While this literature has had a long and interesting history, there has been intense interest over the last few years and an increasing number of important contributions in nowcasting, which is the task of predicting the present, the very recent past, or the very near future of GDP, and some other macro variables as well. Such important contributions are, inter alia, Evans (5), 1 Payroll information is released by the Labor Department on the rst Friday of every month.

3 Banbura et al (1), Giannone et al (8, 1), Camacho and Perez-Quiros (1), Kuzin et al (forthcoming), Barhoumi et al (1). Much of this nowcasting literature takes advantage of recent advances in factor models and related techniques that allow researchers to extract useful information from large data sets with many variables and thus deliver forecasting gains. While we do not want to discard any variables that could be potentially useful in forecasting GDP, directly employing a large model with many variables would require estimating a large number of parameters, which of course would compromise the estimated model s forecasting performance. Factor models, that the nowcasting literature employ, manage to deal successfully with that issue. An additional challenge that comes with this task of extracting the maximum amount of useful information from these large data sets in real time is that as new data releases arrive throughout the quarter they are incorporated at various times into these panels, which are thus unbalanced panels, or have jagged edges. The nowcasting literature typically employs standard Kalman ltering techniques to deal with this issue of di erent variables having di erent endpoints at any given point in time. This paper s primary goal is to bring together the extant literature on ISM and PMI with the new literature on nowcasting. In particular, and while the PMI has been studied as a potential nowcasting tool largely in isolation in the past (see, inter alia, Koenig (), Pelaez (3a, 3b), Cho and Ogwang (6, 7) ) this paper revisits the PMI and other ISM variables within the context of the large jagged data sets employed in nowcasting. For that purpose, we adopt the approach of the seminal paper of Giannone, Reichlin, and Small (8) - henceforth GRS. GRS employ a dynamic factor model and the Kalman smoother to nowcast US GDP. We revisit some of their work, while paying closer attention to individual ISM variables and their role in nowcasting GDP. The paper is organized as follows: Section provides some historical background and more details on the ISM and the di usion indices it produces. Section 3 investigates, in a more traditional way, the bivariate relationship between ISM indicators and GDP growth. The rest of the paper then studies the role of ISM indicators in nowcasting GDP growth from large panels. In particular, section 4 summarizes the econometric approach of GRS that we employ and section 5 presents results from our nowcasting exercises with various ISM variables. Finally, Section 6 o ers some 3

4 concluding remarks. ISM data: A closer look Except for a four-year interruption during World War II, the ISM (formerly known as NAPM, the National Association of Purchasing Management) has been sending every month since 1933 a national survey to a sample of purchasing and supply executives of more than 4 companies in manufacturing industries across the U.S. The resulting report containing the data compiled from the survey responses is the Manufacturing ISM Report on Business (ROB). These survey responses re ect the change in the current month over the previous one for 1 indicators which are new orders, production, employment, supplier deliveries, inventories, customers inventories, prices, backlog of orders, exports, and imports. Di usion indexes are then created based on the responses to these survey questions. For instance, for production, the possible responses to the question What is the trend for production? are positive, neutral or negative (compared to the preceding month). The resulting di usion index is created by adding the percentage of positive responses to half the percentage of the neutral responses. This number varies between and 1 and it represents the percent of companies that increased their production during the month. Basically, a level above 5 indicates that more executives are reporting increase for that variable than are reporting decrease. Di usion indices are then seasonally adjusted. The Purchasing Managers Index (PMI) is the equally weighted (. each) 3 composite index of ve of these seasonally adjusted di usion indexes: New Orders, Production, Employment, Sup- The details of this seasonal adjustment are available in: 3 When the PMI was rst introduced in 198 by Thedore Torda, a senior economist of the US Department of Commerce, it was constructed as an equally weighted composite index of ve of the ten seasonally adjusted di usion indexes which are the results of the ISM manufacturing survey: new orders, production, employment, supplier deliveries and inventories. The PMI was back-calculated prior to 198, and is available starting in In 198, the US Department of Commerce changed the weights of these ve di usion indexes in order to maximize the relationship between the PMI and the GDP. The new weights were.3 for New Orders,.5 for production,. for Employment,.15 for Supplier deliveries and.1 for Inventories (Torda 1985). Several studies have discussed the plausibility of using fewer components and di erent weights which can improve the PMI. Pelaez (3a) proposed an alternative to PMI, which is based on regressions of the growth rate of GDP and industrial production index on current and lagged values of PMI components. The weights were allowed to evolve over time in one version and remained xed in another. Pelaez (3b) used an index composed of three of the PMI components (new orders, employment and supplier deliveries). Cho and Ogwang (6) used only the employment component of PMI. Cho and Ogwang (7) applied principal component analysis using six of the ten non-manufacturing di usion indexes to compose a non-manufacturing PMI. In 8, ISM eventually returned to equal weights. 4

5 plier Deliveries and Inventories. These indexes are given in Figure 1 4. The composite index PMI again ranges from to 1, with 5 again being the critical reference. ISM speci es a reading above (below) 5 as indicating that the manufacturing sector is in expansion (contraction). Similarly, and as Koening () discusses, a PMI above 41. indicates an expansion of the overall economy. Therefore, the PMI below 41. indicates contraction in both the overall economy and the manufacturing sector. Some studies even de ne the critical reference as 47 for expansion of the manufacturing sector and 4 for expansion of the economy. The index is released at 1: a.m. EST on the rst business day of each month, requires little revision, and is widely recognized by many economists and business practitioners as a reliable short-term barometer of economic activity ( Dasgupta and Lahiri 1993). While, and as discussed earlier, these are clearly desirable features, there are disadvantages to the PMI as well: One such drawback is simply a result of the PMI being a di usion index: Its increases or decreases do not capture the intensity with which business conditions are changing. Furthermore, the PMI doesn t account for size di erences across rms: As such, it may miss important shifts in business conditions if, for instance, such shifts are primarily concentrated in a few large rms. For our purposes the ultimate criterion is to what extent the PMI helps nowcast GDP, especially in the context of the many other variables that can be used for that purpose. The last panel of Figure 1 plots the PMI and we can clearly see there that there is a strong, yet far from perfect relationship between the PMI and the business cycle. In addition to these manufacturing indices, the ISM has begun composing non-manufacturing indices as well. While services isn t as cyclical as the manufacturing sector is, it is well known that the share of manufacturing in the economy has dropped dramatically over the last half century. In that context, non-manufacturing di usion indices seem central to achieving better coverage of the economy. The non-manufacturing Report on Business becomes available on the third business day of each month soon after the manufacturing ROB, and it is again based on survey questions (asked of 375 executives in 16 non-manufacturing industries across the country). There are 1 non-manufacturing seasonally adjusted ISM di usion indices: business activity, new orders, em- 4 The shaded areas in the gure are the National Bureau of Economic Research (NBER) de ned economic recessions. 5

6 ployment, supplier deliveries, backlog of orders, new export orders, inventory change, inventory sentiment, imports and prices. There is also a composite Non-Manufacturing Index (NMI) which is available only since 8, however. This NMI is based on business activity, new orders, employment, and supplies deliveries. Figure illustrates the four indicators that are used in the construction of the NMI, as well as (bottom panel) the NMI that can be constructed using the entire time series sample on these four indicators. Again, this gure suggests that there is some relationship between these indices and the business cycle. Our task in what follows is to investigate more explicitly the importance of all these manufacturing and non-manufacturing ISM indices in nowcasting GDP growth. 3 ISM indicators and GDP growth: A bivariate investigation In order to check how PMI nowcasts current quarter GDP growth (Y t ), we transformed the monthly PMI index to a quarterly series using the formula: P MI q;t (1=9)P MI(t 1; ) + (=9)P MI(t 1; 3) + (3=9)P MI(t; 1) + (=9)P MI(t; ) + (1=9)P MI(t; 3), where P MI(t; i) is the level of the PMI in the ith month of quarter t. This weighted-average de nition is appropriate given that GDP growth measures the percent change in the quarterly average level of economic activity, see Koenig (). We regressed Y t on this quarterly index using the equation that follows. A di erence term (P MI q;t ) is added to allow for lagged adjustment and serial correlation. Y t = c 1 (P MI q;t c ) + c 3 P MI q;t + u t Using observations over 1948:1-1:1, the estimated equation was Y t = :8((P MI q;t 4:7) + :P MI q;t, adj:r = :5; DW = 1:83 (1.5) (4.7) (4.67) t-values are reported in parentheses. The advantage of estimating in this non-linear form is that it will yield an estimate of the threshold parameter c : a PMI value above it suggests the economy is expanding. u t is the usual regression error term. The estimates are remarkably similar to those in Koenig () who used data till rst quarter of. The explanatory power is.5 with a threshold value of 4.7. The in-sample root mean square error was.18, and Theil s U statistic compared to no change as the naïve forecast was.69. These values and estimates over di erent sub-periods suggest that the composite quarterly PMI has reasonably good and stable 6

7 explanatory power in nowcasting GDP growth. The above link function does not identify the separate contribution of rst, second and third monthly PMI values in the nowcast. This is achieved using an alternative link function by regressing Y t on three independent variables rst monthly PMI, second monthly PMI and the third monthly PMI of a quarter. These regressions are reported in Table 1 where we nd that the adj:r values of the regressions are.54,.59 and.6 respectively, suggesting that last two monthly values have very modest marginal explanatory power over the rst monthly value. This later formulation gives a slightly better explanatory power than the link function in the equation above. We also examined the temporal stability of the relationship between GDP growth and PMI in real time using GDP growth gures based on rst announcements. These correlations over and over di erent sub-periods were calculated to be: :.67; :.73; :.59; :.68; 1984-:.44; and 1-1:.7. Interestingly, during the relatively stable period 1984-, the correlation was the lowest, and PMI seems to have higher correlation with GDP growth in periods of higher volatility. The fact that during 1-1 the correlation was very high, is not consistent with the hypothesis that correlation of PMI with GDP has diminished after 1984 due to better inventory control practices (McConnell and Perez-Quiroz ). One explanation of our nding is that during crisis periods, the factors that PMI represents (e.g., supplier deliveries, Inventories, etc.) tend to have closer relationship with the economy. During prolong periods of stable growth PMI may possibly lose its relative advantage. We now examine the nowcasting capability of non-manufacturing (NMI) ISM index that is available since 1997:7. Using the composite quarterly NMI values, the following non-linear regression was estimated: Y t = :8(NMI q;t 45:7) + :NMI q;t ; adj:r = :569; DW = 1:85 (5.17) (5.) (3.75) The explanatory power is similar to that with PMI and suggests a higher value (45.7) of the threshold while using NMI compared to PMI to signal GDP growth. In order to examine the separate contributions of the three monthly values of NMI, we also regressed Y t on the rst, second, and third monthly NMIs, The results are reported in Table. The adj:r for the three regressions were.37,.44 and.5 respectively, and were signi cantly 7

8 lower than those due to PMI. In Figure 3 we have plotted quarterly GDP growth together with monthly PMI and NMI over 1997:7-1:1 that covers last two recessions. The three monthly series look similar to each other, with the NMI slightly muted than the corresponding PMI series. Both PMI and NMI moved in remarkable tandem with GDP growth and the recession cycles. In both recessions, they started rising quite ahead of the business cycle troughs with considerable lead time. Also, the cyclical variation in the monthly indices is a lot more prominent than the composite quarterly index suggested by Koenig (199), particularly at the beginning of the recessions. Over all, consistent with numerous previous studies, we nd that both PMI and NMI by themselves have very signi cant explanatory power in nowcasting current quarter GDP growth, and the relationship seems to be stable. Given that ISM indices are announced at the beginning of each month ahead of other monthly indicators, and have strong cyclical movements ahead or coincident with cyclical turning points, the media and the policy makers interest in ISM is understandable. However, ISM indicators exist together with many others in the economy with which the ISM indices are expected to be highly correlated. So the question is: how signi cant is the marginal contribution of ISM indicators given other indicators, and how important is the fact that PMI and NMI are announced ahead of other indicators at the beginning of each month. 4 Nowcasting GDP growth using the framework of Giannone, Reichlin, and Small The approach of GRS that we employ combines a dynamic factor framework with the Kalman smoother. It thus deals with both central challenges associated with nowcasting, (as discussed in the introduction, namely the large number of variables with a potential proliferation of parameters, and the jagged edges of the data set). Furthermore, it has the potential to capture essential dynamics in the various time series of the panel. The rest of this section summarizes the GRS approach, keeping their notation intact. The reader is referred to their paper for more detailed information. It is assumed that the information included in the large number of explanatory variables is captured by a few common factors: 8

9 x tjj = + F t + tjj (1) where x tjj is an n 1 vector of observed explanatory variables 5 available in vintage j (j = 1; :::; J), in month. In practice GRS create 15 vintages per month (so 45 in a quarter) based on their analysis of the pattern of data releases every month (this pattern is about the same across months, which justi es this approach). F t is an r 1 vector of the common factors and is an n r matrix of factor loadings. The dynamics of the common factors are modeled as follows: F t = AF t 1 + Bu t () where u t is a q 1 white noise vector of shocks to the common factors and B is an r q matrix of rank q. A is an r r matrix with all roots of det(i r Az) lying outside the unit circle. GRS parameterize (and make a robustness case for) their benchmark speci cation with r = q = and we keep the same parameterization in what follows. Note that the idiosyncratic error terms tjj are assumed to be cross-sectionally orthogonal white noises and also orthogonal to shocks u t. Note that with equations (1) and () we have a state space framework and we can thus apply standard Kalman ltering techniques to estimate the common factors, given parameter estimates. GRS proceed with the estimation as follows: First, they apply principal components to a balancedpanel subset of the original jagged-edge data set 6 and estimate the parameters above by OLS regressions on these principal components. Then the common factors are estimated by running the Kalman smoother using the entire (thus unbalanced) data set, where parameter estimates replace true parameter values in the state space speci cation above. Given such estimates of the common factors, GDP nowcasts emerge simply as the tted values from OLS regressions of the quarterly GDP series on these quarterly estimated factors 7. The jagged edge data set that GRS have put together consists of close to macro variables 5 These variables have been transformed to induce stationarity. We keep GRS s transformations intact, and we thus refer the reader to (the appendix to) GRS for the details. 6 This balanced panel is created by discarding observations that are not available for all the variables. 7 Unlike the GDP series, most of the observables in the GRS data set come at the monthly frequency; GRS thus transform them by applying a lter (Mariano and Murasawa 3) that converts the monthly series to an (approximately) quarterly quantity when observed at the end of a quarter. Thus the OLS regressions are run at the quarterly frequency for both the dependent variable and the regressors. This is GRS s approach to the more simplistic bridge regressions of standard nowcasting practice. 9

10 for the US economy starting in January of 198. These variables, most of which are at the monthly frequency, and which include real and monetary quantities, prices, surveys, are grouped into blocks, or vintages, 15 per month, on the basis of a stylized calendar of monthly data releases that remains (mostly) unchanged across months. This data set (of the original GRS paper) was collected at a single point in time, and thus cannot be used to take into account data revisions and their possible e ect on the nowcasts. We employ this "pseudo real-time" data set in parts of the analysis that follows. However, we also put to use in the next section a true real-time version of the GRS data set, which enables us to account for data revisions as well. This data set 8 is updated every Friday with any and all releases that are available at that point. Therefore we essentially have a series of overlapping, real-time sets with any later data set di ering from previous ones for one or both of two possible reasons: updated gures for a given observation(s) of one or more variables in the data set, and/or more recent observations for one or more variables in the data set. The GRS data that we use in what follows does include the ve ISM manufacturing variables discussed earlier that are used to construct the PMI, but not the ISM non-manufacturing series. We thus augment the GRS data set with non-manufacturing information and with any additional ISM manufacturing indices that could be potentially bene cial, but we have to exclude some variables on the basis of low response rates or short time series. We thus end up considering materials buying prices (in addition to PMI and its ve indices) on the manufacturing side, and several non-manufacturing indices: inventories, new orders, deliveries, employment, business activity. 5 Assessing the Nowcasting Performance of ISM variables Our goal in this section is to evaluate the ISM variables (both when considered in isolation and when bundled together) performance in nowcasting GDP, always in the context of an environment where many macro variables are part of our information set at any point in time within any quarter. Other recent work (for instance, and in addition to GRS, see Banbura et al (1) and Camacho and Perez-Quiros (1) ) has stressed the importance of the timeliness issue when it comes to survey data s impact on nowcasts. Timeliness must certainly be a central factor when it comes to the ISM data in particular; as discussed earlier, the ISM manufacturing indices become available 8 The data, which were kindly provided to us by David Small, are the series used in Giannone et al (1), minus a few proprietary series. 1

11 rst thing every month and their non-manufacturing counterparts come out soon after that. We thus dig deeper into both of these issues, namely the marginal impact of the ISM variables on the nowcasts and the timeliness factor, with a view to gaining some insight on the following 3 questions: Suppose you are located at the end of month 1 or the beginning of month, and your task is to nowcast (or forecast, if month 1 is the last month of the previous quarter) GDP for the quarter in which month belongs. First, do ISM data help in this task? Second, which (any or all) ISM variables do help and what s the role of PMI in this task? Third, what s the added value, if any, of non-manufacturing ISM data? Our rst take in attempting to answer these questions is by looking at the time series of the nowcasts, and by assessing how their in-sample t is a ected by the various ISM indices (together or in isolation). However, this simple exercise does not capture the timeliness issue discussed above, nor does it account for forecast uncertainty. We therefore proceed, in a second step, to compute out-of-sample Mean Square Forecast Errors, while again focusing on the e ect of the ISM variables. Finally, we conduct a true real-time analysis (using the real-time vintages described in the previous section) of various quarters in isolation (corresponding to the period right before and during the Great Recession 7-9); for each of these quarters we consider di erent scenarios where we nowcast using information available in real time at the very beginning of a month, and then add one or more observations from ISM indices to our information set and see how this a ects the nowcasts of quarterly GDP growth rates. 5.1 The e ect of ISM indices on the in-sample t of the nowcasts We rst compute the "nowcasts", or more precisely, the tted values of GDP growth rates obtained using the factors F t, (estimated using the state-space approach of equations (1) and () ), over the entire historical period. We repeat this several times, each time leaving out one (or more) ISM indices from the raw data used to estimate the factors. These exercises produce a series of gures, two of which are included in Figure 4. These gures, which plot realized GDP growth, nowcasts obtained without using any ISM indices, and nowcasts obtained using one or more ISM indices 9 9 The full set of gures is available from the authors. Here we include two examples: The rst one includes all the ISM variables, and the second one includes nowcasts produced using the ve non-manufacturing ISM indices without their manufacturing counterparts, using manufactring indices without any non-manufacturing ones, and nally nowcasts obtained using both non-manufacturing indices and the PMI. 11

12 (in addition to the other variables of the data set) enable us to assess how well the nowcasts track realized GDP growth, in sample, with, or without ISM information. These gures lead us to two immediate conclusions: First, we con rm the GRS observation that their approach leads to good in sample ts - the nowcasts track the realized GDP growth rates well. Second, it seems that ISM variables don t make a big di erence, when it comes to in-sample t: the nowcast lines obtained with and without ISM indices are largely congruent 1. One might be tempted then to cast doubt on the importance of these ISM indices altogether. However, this would be premature: As discussed above, the exercise here ignores (being in-sample) the timeliness issue, the real-time nature of the data, as well as nowcast uncertainty. All of these issues of course are of paramount importance from the perspective of a policy maker or of an econometrician seeking to assess the evolution of GDP in real time and in a timely manner. In the exercises that follow, we thus seek to replicate more closely the actual environment faced by a nowcaster in real time: 5. The marginal impact of ISM indices on nowcast accuracy To assess the marginal impact of ISM indices on the accuracy of the nowcasts, we perform an exercise similar to the "pseudo" real-time exercise of GRS, replacing their 15 "stylized monthly" data releases by various ISM variables, considered either jointly or in isolation. More speci cally, we rst produce a nowcast 11 of current-quarter GDP growth, on the basis of information available at the very beginning of the month, that is even before the ISM manufacturing release of the rst business day of the month. Then we repeat this exercise using an expanded information set that includes the extra observation on the PMI that becomes available in the beginning of the month. Subsequently we repeat this exercise six times, each time augmenting the information set with an extra observation coming from one of the ve ISM indices that are part 1 As discussed earlier, the data are transformed into "quarterly equivalents" using the approximation introduced by the Mariano and Murasawa lter (Mariano and Murasawa 3). As a robustness check, we also computed in sample ts using factors estimated using data without applying the lter; the results were very similar and the conclusions reached (regarding the role of ISM variables in sample) are the same. Note that the RMSE from this exercise (using all three months factors), but using real time GDP data, was found to be 5% less than the RMSE using only the PMI series (see last column of Table 1). This is one way of illustrating the value of the GRS approach using many indicators. 11 Or forecast, if we re talking about the rst month of a quarter. 1

13 of the PMI (New Orders, Production, Employment, Supplier Deliveries and Inventories) or from Prices. Finally, we produce another nowcast using all the new observations coming from all of the ve indices that enter the PMI together. Then we repeat, for all the subsequent months 1 ; that is, we compute a series of nowcasts recursively, and we measure (out-of-sample) nowcast uncertainty by estimating Mean Square Forecast Errors. The results of this recursive exercise are depicted in the top panel of Figure 5, which shows both the evolution of uncertainty across months through a quarter, and the marginal impact of ISM information within a month. Several conclusions emerge from that gure: First, we con rm the conclusion of GRS that forecast accuracy increases precipitously as the quarter progresses and more information (from accumulating data releases) gathers. Second, and for any month, one or more ISM variables help in increasing accuracy; while most ISM indices help, it is hard to identify a speci c index that clearly outperforms the rest (in reducing uncertainty). Third, incorporating the PMI and/or all the indices together always results in reduced MSFE. In a nutshell, information contained in the ISM always has a role in reducing uncertainty. However, this reduction is a little more pronounced (and thus the marginal e ect of ISM information is more signi cant) in the beginning of the quarter and more modest later in the quarter. We then focus on the role of the Non-Manufacturing ISM indices in a similar way: The bottom panel of Figure 5 shows Mean Square Forecast Errors for each of the months of the quarter 13 corresponding to nowcasts that are obtained as follows: First, before the ISM release at the beginning of the month; second using the new PMI observation only (i.e., as the sole addition to the information set available at the very beginning of the month); third using extra observation(s) coming from the non-manufacturing indices only, rst separately, and then jointly 14. The main observation that one can make using this exercise (in addition to the ones mentioned above) is that non-manufacturing information is useful in reducing nowcast uncertainty, and indeed more so than the PMI, as introducing the non-manufacturing indices together reduces the MSFEs more so than 1 We also include a fourth month in our gures; that is, we compute the nowcasts obtained using the information set available at the very end of a quarter, with and without the extra ISM information coming in the very beginning of the following quarter (and thus well before any GDP gures on the quarter that just ended become available by the BEA). 13 That is, the three months of the quarter plus the fourth month as discussed above. 14 Speci cally, these non-manufacturing di usion indices are: The non-manufacturing Business Activity Index, Employment, Inventories, New Orders, and Supplier Deliveries. 13

14 the PMI does 15. One important caveat that must be added here though is that the period over which the MSFEs are computed is much shorter, as, non-manufacturing indices, and in contrast to their manufacturing counterparts, are only available starting in July of Of course, and as discussed above, all these observations are made using a pseudo-real time data set. The exercise that follows aims to get even closer to actual nowcasting conditions by employing the sequence of real time data sets discussed in the previous section, thus also taking into account the role of data revisions. 5.3 Nowcasting with ISM in real time during the Great Recession We consider seven quarters in isolation starting in 7 Q4 and ending in 9 Q, that is the period right before and during the latest Recession, which is arguably a challenging period from a nowcasting perspective (Figure 6). For each of these quarters, and in a manner similar to before, we do the following exercises: First, we produce a forecast for quarterly GDP growth rate for the rst month of the quarter, say month, on the basis of the information set that would have been available in real time at the very end of month 1, or early in the morning of the rst business day of month ; that is, right before month s ISM manufacturing release becomes available 16. Second, and as before, we generate an additional ve nowcasts by augmenting the information set just described with the new observation from each of the ve di usion indices (one at a time) that are part of the PMI: New Orders, Production, Employment, Supplier Deliveries and Inventories. Third, we continue again as above: we repeat this exercise introducing all ve indices at the same time, and the PMI on its own. Thus, on the horizontal axis, "No ISM" denotes dynamic factor model forecasts generated in real time using the large set of predictors without any of the ISM indicators. Each of the other seven categories (e.g., Prod. only, Empl. only, All 5, etc.) 15 We also conducted an exercise where we compared MSFEs associated with nowcasts obtained by completely removing (or adding) ISM variables from the information set. We reached similar results and conclusions. 16 Then we produce similar nowcasts for months + 1, +, and + 3. Given that the survey solicits information for month-over-month changes, nowcasts for + 1, +, and + 3 relate directly to the quarter in question. The values for month are really nowcasts for the last month of the preceding quarter. As discussed earlier, we have a sequence of sets that contain real-time data as they are available on Fridays. We do some manual adjustments to these Friday data sets on a case by case basis and according to the calendar of data releases to create the information sets we need for our purposes. 14

15 denotes forecasts generated by No ISM plus ISM production, employment, all ve ISM indicators together, etc. Successive panels also indicate how these large set of indicators were performing in nowcasting quarterly GDP growth in real time as the recession was progressing. We thus have eight nowcasts in total, which are to be assessed in terms of how far they are from the revised or real time GDP series 17. The results for the seven quarters are summarized in the rst seven panels of Figure 6, which contain two straight lines (the real time and revised GDP gures) and four additional lines for the four months corresponding to a quarter. Looking at these gures, we can reach some tentative conclusions 18, several of which however are qualitatively similar to the ones we reached above: In every case there is helpful information in ISM indices for nowcasting; but it s not always all the indices that help. Furthermore, and on the basis of these seven quarters, it s not possible to securely identify any individual indices that consistently overperform or underperform. On balance, however, production, all 5 and the PMI of the ISM manufacturing indicators seem to have marginal value over the factor nowcasts without ISM. The ISM indices taken together (either with free weights or the PMI) do tend to improve nowcasts most of the time. On the other hand, clearly, the successive nowcasts get progressively closer to the real time GDP gure as we go from the rst month of the current quarter to the rst month of the next quarter. During the depth of the recession i.e., 8Q4-9Q1, the large negative GDP growth gures were very successfully nowcasted at the end of the quarter as nowcasts from rst month of the quarter were very rapidly revised over the next three successive nowcasts. Only exception was the third quarter of 8 (when the real time GDP growth became negative for the rst time during this recession), where the successive nowcast revisions almost stayed unchanged and refused to produce a negative value even at the end of the nowcasting round for the quarter. After this quarter though, the nowcasts quickly learnt to adapt to the rapidly deteriorating economy and also the rebound in 9Q. As we progress to later months within a quarter, nowcasts tend to move closer to the target number (revised or initial GDP growth estimate), but it also tends to get harder for ISM information - the ve indices taken together - to have a signi cant marginal impact on the nowcasts 17 For the revised GDP growthseries we use gures available on Friday April Realizing of course that they are based on information obtained from a small number of quarters only. 15

16 (obtained with no ISM information). It is also noteworthy that during the depth of the recession (8Q4-9Q1), ISM indicators was particularly useful in nowcasting current quarter GDP in relation to the factor model nowcasts without ISM. We also investigate the role of non-manufacturing ISM indices in a similar manner: We generate nowcasts using information at the beginning of the month with no ISM data, and then we see how these nowcasts change when we introduce new observations coming from non-manufacturing ISM indices - or the new observation coming from the PMI. Results for the 7 quarters are provided in the latter seven panels of Figure 6. The conclusions reached are similar to the ones already discussed in the previous case - both non-manufacturing index (NMI) and the PMI tend to help nowcasts, if only marginally, over the factor model nowcasts without ISM variables. We also observe that in certain quarters (e.g., 8:Q1 and Q, 9:Q1 and Q) NMI acted slightly di erently. The relative contribution of NMI over PMI is not consistent over all quarters examined. But we can conclude that NMI nowcasts are at least as good as those based on PMI. Like PMI, the four successive nowcasts for 8:Q3 with NMI also did not adapt quick enough to foresee the rst negative quarterly real GDP growth value of the recession. Another conclusion that emerges from these gures is that all forecasts - factor forecasts without ISM indicators, and factor forecasts with manufacturing and non-manufacturing indicators - are generally closer to GDP growth gures based on initial announcements than those based on revised gures. 6 Summary The di usion indices produced by the Institute for Supply Management are well known and highly publicized as they are generally thought to contain useful information about the economy s direction. As such, they have been studied extensively by various researchers. Much of the existing literature, nevertheless, tends to consider the ISM indices, and the PMI in particular, in isolation. However, what may be more interesting from, say, a policy maker s perspective who tries to assess the usefulness of ISM indices in nowcasting GDP (given the plethora of variables that could be used for that purpose), is the marginal impact of the ISM variables given all the other information that s available in real time. In this paper, we focus on this issue: we study the importance of ISM variables within the context of a large data set that is used to nowcast GDP growth. For 16

17 this purpose, we employ the approach to nowcasting from large data sets developed by Giannone, Reichlin, and Small (8). We nd some evidence that the emphasis on ISM indices (both manufacturing and nonmanufacturing) is in a sense well placed: Even in a context where many other variables are available, ISM indicators help us improve the nowcasts of current quarter GDP growth. This is primarily because they become available rst thing in the month and ahead of other indicators. The improvement tends to be more signi cant early in the quarter, but the three successive monthly nowcasts consistently improve upon the earlier ones, and approach rapidly towards real time GDP growth gures. References Barhoumi, K., G. Runstler, R. Cristadoro, A. Den Reijer, A. Jakaitiene, P. Jelonek, A. Rua, K., Ruth, S. Benk and C. Van Nieuwenhuyze (8): "Short-Term Forecasting of GDP Using Large Monthly Datasets: A Pseudo Real-Time Forecast Evaluation Exercise", NER - E #15, Banque De France. Banbura, M., D. Giannone, and L. Reichlin (1): "Nowcasting", in Michael P. Clements and David F. Hendry, editors, Oxford Handbook on Economic Forecasting, forthcoming. Camacho, M. and G. Perez-Quiros (1): "Introducing the EURO-STING: Short term indicator of Euro area growth," Journal of Applied Econometrics (forthcoming). Cho, D.I., Ogwang, T., 6, Conceptual Perspectives on Selecting the Principal Variables in the Purchasing Managers Index, The Journal of Supply Chain Management, vol. 4, no., pp. 44-5(9). Cho, D.I., Ogwang, T., 7, A Conceptual Framework for Computing U.S. Non-manufacturing PMI Indexes, The Journal of Supply Chain Management, vol. 43, no.3, pp Dasgupta, S., and K. Lahiri (199): "A Comparative Study of Alternative Methods of Quantifying Qualitative Survey Responses Using NAPM Data," Journal of Business and Economic Statistics 1(4),

18 Dasgupta, S., and K. Lahiri (1993): "On the Use of Dispersion Measures from NAPM Surveys in Business Cycle Forecasting," Journal of Forecasting 1(3&4), Evans, M.D.D. (5): "Where are we now? Real-time estimates of the macro economy", International Journal of Central Banking 1() Giannone, D., M. Modugno, L. Reichlin, and D. Small (1): "Nowcasting in Real-Time", mimeo. Giannone, D., L. Reichlin, and D. Small (8): "Nowcasting: The real-time informational content of macroeconomic data", Journal of Monetary Economics (55), Kau man, Ralph (1999): "Indicator Qualities of the NAPM Report on Business", The Journal of Supply Chain Management 35(), Klein, P.A. and G.H. Moore (1991): "Purchasing Management Survey Data: Their Value as Leading Indicators", in Lahiri, K., and G.H. Moore (Eds.) Leading Economic Indicators: New Approach and Forecasting Records, Cambridge University Press, Koenig, Evan (): "Using the Purchasing Managers Index to Assess the Economy s Strength and the Likely Direction of Monetary Policy", Economic and Financial Policy Review, Federal Reserve Bank of Dallas 1(6). Koenig, E.F., S. Dolmas, and J. Piger (3): "The use and abuse of real-time data in economic forecasting", The Review of Economics and Statistics 85(3) Kuzin, V., M. Marcellino, and C. Schumaker (forthcoming) MIDAS vs Mixed-Frequency VAR for Nowcasting GDP in the Euro Area, International Journal of Forecasting (forthcoming). Lindsey, M.D., and R. Pavur (5): "As the PMI turns: A Tool for Supply Chain Managers", The Journal of Supply Chain Management 41(3), Mariano, R., and Y. Murasawa (3): "A new coincident index of business cycles based on monthly and quarterly series," Journal of Applied Econometrics, 18, pp McConnell, M. and G. Perez-Quiros (): " Output Fluctuations in the United States: What has Changed since the Early 8s? American Economic Review, 9(5) 18

19 Orphanides, Athanasios (): "Monetary-Policy Rules and the Great In ation", American Economic Review 9(), Pelaez, R.F., 3a, A New Index Outperforms the Purchasing Managers Index, Quarterly Journal of Business and Economics, vol. 4, no.1, pp Pelaez, R.F., 3b, A Reassessment of the Purchasing Managers Index, The Journal of National Association for Business Economics, vol. 38, no. 4, pp Torda, T.S., 1985, Purchasing Management Index Provides Early Clue on Turning Points, Business America, U.S. Department of Commerce, pp

20 Table 1: Regression of GDP Growth on PMI Components, 1948:1 1:1 First Month PMI First Two Month PMI All Three Months PMI Adj R AIC DW RMSE MAE Theil's U Naïve AR(1) Table : Regression of GDP Growth on NMI Components, 1997:7 1:1 First Month NMI First Two Month NMI All Three Months NMI Adj R DW RMSE

21 Figure 1: From Top to Bottom: New Orders, Production, Employment, Supplier Deliveries, Inventories, PMI Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr Jan-48 Oct-49 Jul-51 Apr-53 Jan-55 Oct-56 Jul-58 Apr-6 Jan-6 Oct-63 Jul-65 Apr-67 Jan-69 Oct-7 Jul-7 Apr-74 Jan-76 Oct-77 Jul-79 Apr-81 Jan-83 Oct-84 Jul-86 Apr-88 Jan-9 Oct-91 Jul-93 Apr-95 Jan-97 Oct-98 Jul- Apr- Jan-4 Oct-5 Jul-7 Apr-9 1

22 Figure : From Top to Bottom: Business Activities, New Orders, Employment, Supplier Deliveries, NMI Jul-97 Dec-97 May-98 Oct-98 Mar-99 Aug-99 Jan- Jun- Nov- Apr-1 Sep-1 Feb- Jul- Dec- May-3 Oct-3 Mar-4 Aug-4 Jan-5 Jun-5 Nov-5 Apr-6 Sep-6 Feb-7 Jul-7 Dec-7 May-8 Oct-8 Mar-9 Aug-9 Jan-1 Jun Jul-97 Dec-97 May-98 Oct-98 Mar-99 Aug-99 Jan- Jun- Nov- Apr-1 Sep-1 Feb- Jul- Dec- May-3 Oct-3 Mar-4 Aug-4 Jan-5 Jun-5 Nov-5 Apr-6 Sep-6 Feb-7 Jul-7 Dec-7 May-8 Oct-8 Mar-9 Aug-9 Jan-1 Jun Jul-97 Dec-97 May-98 Oct-98 Mar-99 Aug-99 Jan- Jun- Nov- Apr-1 Sep-1 Feb- Jul- Dec- May-3 Oct-3 Mar-4 Aug-4 Jan-5 Jun-5 Nov-5 Apr-6 Sep-6 Feb-7 Jul-7 Dec-7 May-8 Oct-8 Mar-9 Aug-9 Jan-1 Jun Jul-97 Dec-97 May-98 Oct-98 Mar-99 Aug-99 Jan- Jun- Nov- Apr-1 Sep-1 Feb- Jul- Dec- May-3 Oct-3 Mar-4 Aug-4 Jan-5 Jun-5 Nov-5 Apr-6 Sep-6 Feb-7 Jul-7 Dec-7 May-8 Oct-8 Mar-9 Aug-9 Jan-1 Jun Jul-97 Dec-97 May-98 Oct-98 Mar-99 Aug-99 Jan- Jun- Nov- Apr-1 Sep-1 Feb- Jul- Dec- May-3 Oct-3 Mar-4 Aug-4 Jan-5 Jun-5 Nov-5 Apr-6 Sep-6 Feb-7 Jul-7 Dec-7 May-8 Oct-8 Mar-9 Aug-9 Jan-1 Jun-1

23 Figure 3: GDP and ISM Indicators q1 1997q3 1998q1 1998q3 1999q1 1999q3 q1 q3 1q1 1q3 q1 q3 3q1 3q3 4q1 4q3 5q1 5q3 6q1 6q3 7q1 7q3 8q1 8q3 9q1 9q3 1q1 Recession NMI(Koenig) GDP Growth (%) q1 1997q3 1998q1 1998q3 1999q1 1999q3 q1 q3 1q1 1q3 q1 q3 3q1 3q3 4q1 4q3 5q1 5q3 6q1 6q3 7q1 7q3 8q1 8q3 9q1 9q3 1q1 Recession PMI1 NMI1 GDP Growth (%) q1 1997q3 1998q1 1998q3 1999q1 1999q3 q1 q3 1q1 1q3 q1 q3 3q1 3q3 4q1 4q3 5q1 5q3 6q1 6q3 7q1 7q3 8q1 8q3 9q1 9q3 1q1 Recession PMI NMI GDP Growth (%) q1 1997q3 1998q1 1998q3 1999q1 1999q3 q1 q3 1q1 1q3 q1 q3 3q1 3q3 4q1 4q3 5q1 5q3 6q1 6q3 7q1 7q3 8q1 8q3 9q1 9q3 1q1 Recession PMI3 NMI3 GDP Growth (%) 3

24 Figure 4: In Sample Results In Sample Results: Manufacturing ISM Variables vs no ISM Variables Year GDP All No ISM In Sample Results: Non-Manufacturing vs Manufacturing ISM Variables Year GDP No ISM Mnf only Non-Mnf only Non-Mnf&PMI 4

25 Figure 5: Mean Squared Forecast Erros MSFE for Various Manufacturing ISM Variables 1st month nd month 3rd month 4th month NO ISM PMI PROD EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI PROD EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI PROD EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI PROD EMPL INV NO SUPPL DELIV PRICES ALL MSFE for Non-Manufactuting ISM Variables and the PMI.9 1st month nd month 3rd month 4th month NO ISM PMI BUS_ACT EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI BUS_ACT EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI BUS_ACT EMPL INV NO SUPPL DELIV PRICES ALL NO ISM PMI BUS_ACT EMPL INV NO SUPPL DELIV PRICES ALL 5