Forecasting Consumption in Real Time: The Role of Consumer Confidence Surveys

Transcription

1 Forecasting Consumption in Real Time: The Role of Consumer Confidence Surveys Kajal Lahiri, George Monokroussos, Yongchen Zhao Department of Economics, University at Albany, SUNY, Albany, New York, 12222, USA Abstract December 2012 We study the role of consumer confidence surveys in forecasting real personal consumption expenditure. We reexamine existing models of consumption and consumer confidence using both quarterly and monthly data in real time. Additionally, we produce forecasts of consumption expenditures with and without consumer confidence measures using a dynamic factor model and a real-time, jagged-edge data set. We establish in a robust way that consumer confidence noticeably improves the accuracy of consumption forecasts. Furthermore, traditional macroeconomic theories seem to be unable to fully account for these results. JEL Classification: E21, E27, C53 Key words: Forecasting, Consumption, Consumer Sentiment, Factor Models, Kalman Filter, Real Time Data 1. Introduction The concept of animal spirits, in the standard Keynesian sense (Keynes (1936), Akerlof and Shiller (2010)) has influenced economic thinking for a long time and has received renewed and intense attention in the run up to the recent financial crisis and the ensuing recession. The confidence of economic agents and its importance to the economy occupy a central role in this discussion. Consumer confidence, in particular, is typically at the center of attention of the business press. It has also been studied extensively by academics as well as policy makers. This interest is certainly justified given the overwhelming importance of consumer spending for the US economy. addresses: klahiri@albany.edu (Kajal Lahiri), gmonokroussos@albany.edu (George Monokroussos), hzhao@albany.edu (Yongchen Zhao). 1

2 Several academic studies (see, inter alia, Bram and Ludvigson (1998), Carroll, Fuhrer and Wilcox (1994), Croushore (2005), Fuhrer (1993), Leeper (1992), Ludvigson (2004), Souleles (2004) ) investigate various aspects of the relationship between consumer confidence and consumer spending, at both the micro and the aggregate levels. Given the timing advantage of the standard measures of consumer confidence (see, for instance, Howrey (2001) or the following sections of this study for details) it is possible that such measures help to forecast consumption. This would of course have important implications for economic policy, as well as for key economic theories, such as the permanent income hypothesis. Consequently, a central preoccupation of the relevant literature, including the papers cited above, is to assess the forecasting power of consumer confidence for consumer spending. Some evidence to that effect is generally found in the literature, but it becomes more modest once a few additional variables that have traditionally been considered in studies of consumer confidence are added to the specification. However, and as discussed in some of the papers mentioned above, much of this existing literature has several limitations: First, quarterly data are commonly used. However, both of the most widely known measures of consumer confidence (the University of Michigan s Index of Consumer Sentiment (ICS) and the Conference Board s Consumer Confidence Index (CCI)) are available at a monthly frequency, and employing quarterly averages of these monthly indices in models of consumption expenditures may confound the effects of consumer confidence.¹ Moreover, consumer spending itself, and also many other relevant indicators are available on a monthly basis as well. Second, revised data on the relevant variables are employed, as opposed to the data that were actually available in real time, i.e., before any revision that only became available at subsequent points in time. Of course, for monetary policy purposes, or more generally, for the purpose of assessing the real-time forecasting power of consumer confidence, real-time data should be used. Third, the regression models used to assess the predictive power of consumer confidence typically include only a small number of additional variables, i.e., a rather small information set, whereas many more variables, possibly in the hundreds, are available that are potentially relevant to consumption forecasting. In this study, we provide what is arguably a more realistic assessment of the predictive power of consumer confidence on consumer spending by addressing all of 1 In what follows, we use the word sentiment and confidence interchangeably. 2

3 the three issues mentioned above. First, we extend some of the existing models using monthly and real-time data, in addition to quarterly and revised vintages. Then we employ a large real-time data set with close to two hundred explanatory variables at the monthly frequency in order to assess the marginal impact of confidence on consumer spending in the context of such a large information set in real time. To deal with the ensuing challenges, such as the proliferation of parameters, we use the dynamic factor framework of Giannone, Reichlin and Small (2008). In a manner similar to Lahiri and Monokroussos (2012), we gain insight on the marginal impact of consumer confidence on consumer spending by comparing consumption forecasts based on information sets with and without consumer confidence measures. Our results generally establish the importance of consumer confidence in forecasting consumption. The rest of the paper is organized as follows: Section 2 provides some discussion of the important aspects of the consumption and consumer confidence data. Section 3 revisits some models used in the existing literature on the predictive power of consumer confidence. Section 4 outlines the dynamic factor approach and assesses the predictive power of consumer confidence in real time when it is a part of a large information set. Section 5 relates the empirical results to macroeconomic theories. Section 6 concludes. 2. Consumption and Consumer Sentiment: A Closer Look at the Data Consumer spending accounts for about two-thirds of domestic final spending in the United States. The primary measure of consumer spending on various types of goods and services is real personal consumption expenditure (PCE). It covers purchases made by households and nonprofit institutions serving households (NPISHs). PCE data come from Personal Income and Outlays released by the Department of Commerce, Bureau of Economic Analysis (BEA). It can be measured by type of products or by function (health, recreation, communication, etc.). In this study, we examine the total PCE and PCE by main types of products: durable goods, nondurable goods, and services. PCE data are available at both the monthly and the quarterly frequencies. The quarterly series are released every month together with the GDP series, typically in the last week of the month. Similar to the GDP series, there is a one-quarter lag between the end of a period and the release of data covering that period. Advance 3

4 estimates of PCE are released for the previous quarter at the end of the first month of each quarter. At the end of the second and the third months, the preliminary and the final estimates for the previous quarter are released, respectively. The monthly series are released one day after the release of the quarterly series. The publication lag for the monthly series is one month. Monthly PCE series are also subject to revisions. Such revisions are announced in the monthly releases. The two monthly PCE values for the first two months (released at the end of the second and the third month) of a quarter play an important role in forecasting the quarterly PCE for that quarter before the advance release of the quarterly value becomes available. However, because of data revisions, the first two monthly values, when combined with the advance release of the quarterly value, cannot be used to pin down the third monthly value. This implies that to someone forecasting in real time, the monthly PCE series and the quarterly ones contain different information. Still, to our knowledge, monthly PCE series have not been used in any extensive study of the relationship between consumer confidence and consumer spending. We thus consider using the monthly consumption series (in addition to their quarterly counterparts) as one of our main contributions. One of the most recognized measures of consumer confidence is the Index of Consumer Sentiment (ICS) from the Survey of Consumers administered by the Survey Research Center of the University of Michigan. The Survey started as an annual survey in It became a quarterly survey in 1952 and then a monthly survey in The ICS index can be separated into a present conditions index and an expectations index, based on the content of the questions used in constructing the index. Each month, about 500 households are interviewed by phone. The preliminary releases of the three indices come out around mid-month, based on the information gathered in the first half of the month, usually two-thirds of the full sample. The final releases are scheduled on the last Friday of each month. Public and media attention is usually concentrated on the final releases, which are quite timely, and subject to no further revision. The Survey of Consumers tracks many different aspects of consumer attitudes and expectations. About 50 core questions are asked in each survey. Five of these questions are used to construct the ICS, and they are as follows: (i) We are interested in how people are getting along financially these days. Would you say that you (and your family living there) are better off or worse off financially than you were a year ago? 4

5 (ii) Now looking ahead do you think that a year from now you (and your family living there) will be better off financially, or worse off, or just about the same as now? (iii) Now turning to business conditions in the country as a whole do you think that during the next twelve months we ll have good times financially, or bad times, or what? (iv) Looking ahead, which would you say is more likely that in the country as a whole we ll have continuous good times during the next five years or so, or that we will have periods of widespread unemployment or depression, or what? (v) About the big things people buy for their homes such as furniture, a refrigerator, stove, television, and things like that. Generally speaking, do you think now is a good or bad time for people to buy major household items? Among these questions, questions 1 and 5 are mainly about the present conditions of the household and the economy, while the remaining three questions are clearly about household expectations. For each of the five questions, a respondent can choose among three responses: favorable (e.g., situation getting better), neutral (e.g., situation is the same as before), and unfavorable (e.g., situation getting worse). The relative score for each question is calculated as the percentage of respondents giving favorable replies minus the percentage giving unfavorable replies plus 100. Then, scores for related questions are rounded and added up and then divided by the base period value from 1966 to form an index. Based on the content of the questions, the Index of Current Economic Conditions (ICC) is constructed using relative scores for questions 1 and 5. The Index of Consumer Expectations (ICE) is constructed using relative scores for questions 2 to 4. And finally, the overall index, ICS, is constructed using all five relative scores. Another widely used measure of consumer confidence is the Consumer Confidence Index (CCI) from the Consumer Confidence Survey administered by the Conference Board. The survey began in 1967 as a bi-monthly survey. Since June 1977, the survey has been administered monthly. Similar to the ICS, the Consumer Confidence Index can also be separated into two components, the present situation component and the expectations component. Each month, a mail survey is sent out and approximately 3,000 completed questionnaires are collected². Preliminary estimates are based on survey responses collected before the 18th of each month. Final 2 A new sample design was introduced effectively in November A discussion of historical comparability is available in the Consumer Confidence Survey Technical Note. 5

6 estimates are published with the release of the following month s data, scheduled on the last Tuesday of each month. The Consumer Confidence Index and its two components are also based on five questions. The first two are used to construct the present situations index and the rest are used to construct the expectations index. All five questions are used to construct the Consumer Confidence Index. These questions are as follows:³ (i) How would you rate the present general business conditions in your area? (ii) What would you say about available jobs in your area right now? (iii) Six months from now, do you think general business conditions will be better, the same, or worse? (iv) Six months from now, do you think there will be more, the same, or fewer jobs available in your area? (v) How would you guess your total family income to be six months from now? Answers: higher, the same, or lower? To each of the five questions, three response options are available: positive, neutral, or negative. The proportion of respondents giving positive responses among those who do not give neutral responses for each question is computed first. A corresponding index is produced for each proportion with the average value for all months in 1985 as the benchmark. Finally, relevant indices are averaged to produce the Consumer Confidence Index and its two components. Seasonal adjustment is performed where needed. In this study, we focus on consumption and consumer confidence data between January 1982 and September 2011 a total of 357 months (119 quarters). Figure 1 compares the University of Michigan s consumer sentiment measures and the Conference Board s consumer confidence measures with the 12-month moving average of (advanced estimates of) annualized monthly growth of real personal consumption expenditure. The shaded areas are periods of recession. The top panel shows the two overall indices; the middle panel shows the two current conditions components; and the lower panel shows the two expectations components. It is clear from the figure that the confidence measures evolve in a similar way as consumption growth does. However, the two measures themselves are not always that closely correlated to each other, especially when it comes to the current conditions component. As mentioned above, Figure 1 plots real-time data of consumption. The importance of using real-time data for tasks such as assessing monetary policy or evaluating 3 See Croushore (2005) for more detailed discussions. 6

7 forecasts in general cannot be understated (see inter alia, Orphanides (2001) and Croushore and Stark (2001)). More specifically to our context, and as Ludvigson (2004) discusses, it is essential to not use revised data when assessing the real-time forecasting power of consumer confidence for consumption. However, much of the existing literature on consumer confidence and consumer spending has employed revised data. Therefore, in the estimation and forecast evaluation exercises that follow, we employ both current/latest and first vintages (obtained from the Federal Reserve Bank of Philadelphia s Real-Time Data Research Center). Furthermore, and as discussed in the introduction, the existing literature employs quarterly data, which can mask important information available at the monthly frequency. Thus, in the exercises that follow, we use both quarterly and monthly data. 3. Consumer Confidence and Consumption: Existing Models Extended A number of attempts have been made in the literature to quantify the importance of consumer confidence measures in explaining and predicting quarterly consumption expenditures. In this section, we first re-examine some of the main models used in previous studies (e.g., Carroll, Fuhrer, and Wilcox (1994), Bram and Ludvigson (1998), and Ludvigson (2004) ), and then extend them to model monthly consumption expenditures. In the process, we focus on the significance of the confidence measure and the change in the model s explanatory power due to the addition of this measure A Univariate Approach We first consider a simple model of consumption expenditure where consumer confidence is the only predictor. For time period t, let C t be (a type of) consumption expenditure, and let S t be (a measure of) consumer confidence. We estimate the following model: τ τ ln(c t ) = α 0 + α i ln(c t i ) + β i S t i + ε t (1) i=1 i=1 7

8 where, following the literature, the number of lags τ is set to 4. We estimate this model for monthly consumption and for quarterly consumption separately.⁴ In the model, the consumption expenditure is one of the following four: total personal consumption expenditure (Total), expenditure on durable goods (Durable), expenditure on non-durable goods (Non-Durable), and that on services (Services). The confidence measure is one of the following six: the three sentiment indices from the University of Michigan and the three confidence indices from the Conference Board. Quarterly sentiment measures are averages of all monthly values within a quarter. Standard error estimates are robust to heteroskedasticity and serial correlation. Table 1 presents the estimation results.⁵ In general, the model s explanatory power is similar to that found in the existing literature (see Ludvigson, 2004). In the table, incremental R 2 s are the differences between the above models (based on Equation (1)) R 2 s and the R 2 of the benchmark model with only four lags of consumption and a constant. The joint significance of coefficients of all lags of the consumption expenditure and of the confidence measure (as well as their corresponding sums of coefficients of both groups of lags) are also reported. The table shows that the consumer confidence measures do indeed explain variations in consumption, but the amount varies. As high as 17% of additional explanatory power can be obtained by adding a confidence measure to the benchmark model, as in the case of adding the expectations component of CCI to the model of real time quarterly consumption expenditure on services. On average (over 12 cases 3 confidence measures 4 types of consumption), when using the UM sentiment measures, a 7.3% increase in R 2 is observed in models of quarterly consumption expenditures, while a 2.7% increase is observed in models of monthly expenditures. In models using the CB confidence measures, an average of 5.4% incremental R2 is observed in quarterly models; and an average of 2.0% increment is observed in 4 We keep the number of lags at four for the monthly regressions as well, since higher lags are insignificant. 5 Table 1 presents quarterly results based on real-time data and monthly results based on revised data. Quarterly results based on revised data are omitted as they are often reported in existing studies. Monthly results based on real-time data are omitted as well because they are not directly comparable with respective results from Section 3.2. However, in these cases of monthly real-time data, the respective average improvements (starting in Sep. 1998, the earliest month for which real-time consumption data are available) are higher than in the cases with revised data: when using the UM sentiment measures, a 6.3% increase is observed, whereas in models using the CB confidence measures, the average improvement is 4.8%. All omitted results here and in the rest of this study are available upon request. 8

9 monthly models Models with Additional Variables Despite clear evidence of the importance of consumer confidence measures in models of personal consumption expenditures from the exercise above, it remains a question whether sentiment measures contain unique information, that is not available in other aggregate measures of economic activity. To investigate this question, we consider the following specification: τ τ τ ln(c t ) = α 0 + α i Z t i + α i ln(c t i ) + β i S t i + ε t (2) i=1 i=1 where in the benchmark model, apart from lagged values of consumption expenditures, there is a set of baseline macroeconomic variables that are typically included in the existing literature (see, inter alia, Carroll, Fuhrer and Wilcox (1994), and Ludvigson (2004)). We include in Z t the S&P500 index, the 3-month Treasury-Bill rate, and labor income, which is wages and salaries plus transfers minus personal contributions for social insurance⁶. Similar to Section 3.1, we estimate the quarterly models using real-time data and the monthly models using revised series.⁷ The results for the real time quarterly models are reported in Table 2. In addition to the statistics reported in Table 1, we report the sum of coefficients and the joint significance of all lags of each and every additional explanatory variable. Similar to the results in the previous subsection, in all 24 models (corresponding to 4 types of consumption and 3 measures of confidence for both CB and UM), including consumer confidence measures increases the R 2, to varying degrees. But with these additional explanatory variables, the magnitude of such increases is in general smaller than that observed in models with the consumer confidence measure being the only predictor. On average, in models using the UM sentiment measures, a 5.2% increase in R 2 is observed 2% less than that of the models without additional variables. In models using the CB confidence measures, an average increase of 4.8% is observed a 0.6% decrease from the previous set of models. i=1 6 Nominal labor income and the S&P500 index are converted to real terms using the PCE price index. 7 Real time data for the components of labor income are only available at a quarterly frequency, thus our monthly models are estimated using only the latest vintage. Quarterly models with the latest vintage are also estimated; the results are qualitatively similar to the ones reported here. 9

10 Table 3 presents the results for the monthly models. Here we also observe an across-the-board increase in explanatory power when a confidence measure is included. However, and in contrast to the quarterly regressions, where the contributions of confidence measures are somewhat damped in the presence of additional explanatory variables, in these monthly regressions we see that the explanatory power is about the same as before. On average, in models using the UM sentiment measures, a 2.7% increase in R 2 is observed a mere 0.02% more than that of the models without additional variables. In models using the CB confidence measures, an average increase of 2.3% is observed a 0.3% increase from the previous set of models. These results suggest that the contributions of consumer confidence measures in explaining consumption expenditures are important.⁸ There is little doubt that confidence does contain a significant amount of information that is not found in a few standard macro-economic variables. However, while the explanatory variables used here are the standard choices in the literature, there are many other variables with potentially significant explanatory power for consumption. Therefore, the findings reported above could be an artifact of having employed an information set that is too restrictive. It is conceivable that the information contained in consumer confidence measures could simply be a combination of the information found in a large number of macroeconomic indicators not included in the above models. 8 Croushore (2005) finds that confidence is not significant in general in forecasting consumption in a specification that includes explanatory variables similar to the ones we used in this section. There are several important differences between his exercises and the ones presented here. First, our sample covers the most recent years. Second, he may be using an income explanatory variable which is different from ours, as he uses nominal personal income deflated with the personal consumption expenditures price deflator. Third, while he uses real-time data for estimation and forecasting, for forecast evaluation, he uses not the consumption data that were available in real time, but vintages available just prior to a benchmark revision these benchmark revisions are made about every five years. One may say that evaluating forecasts generated in real time using figures that only became available after several revisions may be a very stringent test that confidence measures, or many other variables for that matter, are unlikely to pass. This should not necessarily, in our view, be perceived as a prescription to not use confidence when forecasting consumption. For monetary policy and other real-time forecasting purposes, identifying variables such as confidence that give you an edge when forecasting macro-aggregates such as consumption may be a very useful exercise, especially when official government estimates, even advance ones, are not available. 10

11 4. Forecasting Consumption with Consumer Confidence in a Large Data Set Given the above discussion, we want to explore the role of confidence in forecasting consumption, when the forecasts are generated using a wide information set. This requires using a large number of additional explanatory variables and one immediate problem that arises with this type of exercise is the lack of degrees of freedom. Even when we use monthly consumption data, the number of available observations is limited to the hundreds. Yet, potentially useful variables also come in the hundreds, which of course makes the classical linear regression model a poor choice. So we employ here an approach that allows us to address this challenge. The model we use in this section is derived from the dynamic factor model of Giannone, Reichlin, and Small (2008, henceforth GRS). We first introduce the model, the explanatory variables, and discuss associated issues. Then we explore the effect of sentiment on the accuracy of consumption forecasts through out-of-sample pseudo-real-time exercises. Finally, we assess the marginal contribution of sentiment to consumption forecasts again, but now in the context of real-time data e Dynamic Factor Framework This approach utilizes a dynamic factor model in a state-space form to summarize the common information from a large number of explanatory variables with potentially mixed frequencies and varying patterns of missing data. Let x t be a N 1 vector of observed independent variables for time period t, and let F t be a r 1 vector of latent factors representing the state of the economy. The latent factors drive both the concurrent evolution of the explanatory variables and the future evolutions of the latent factors themselves. This relationship is summarized in a state-space model as follows: x t = µ + Λ F t + ξ t (3) F t = A F t 1 + B u t (4) where Λ is an N r matrix of factor loadings, ξ t is an N 1 vector of variable-specific innovations, A is an r r matrix with all roots of det(i r Az) outside the unit circle, B is an r q matrix of rank q, and u t is a q 1 vector of common shocks. As is standard in the literature, we set r = q = 2. Given the jagged-edge nature of the data, i.e., the varying missing data patterns 11

12 in the large number of explanatory variables, especially toward the end of the sample period, estimation of the model is performed in two steps. In the first step, a fully balanced panel of the explanatory variables is created by discarding any observation toward the end of the sample period for which at least one variable is not observed. This is used to obtain preliminary estimates of the latent factors by principal components. These estimates are in turn used to estimate the parameters of the model. Given these estimates, in the second step, the Kalman smoother is used to compute the latent factors for the entire sample period, including those periods discarded in the first step. In this process, the Kalman smoother forecasts the latent factors for periods when the observations for certain variables are unavailable. Our variable of interest, a type of consumption expenditure, is assumed to be determined by the latent factors and possibly a lagged measure of consumer confidence. A simple OLS regression can be used to establish the link between consumption expenditure and these predictors⁹, and to forecast future consumption expenditures given consumer confidence and forecasts of the factors¹⁰. This two-step procedure is necessary to deal with the jagged-edge data structure, which is caused by varying data release schedules and publication lags across all the explanatory variables. Such data structure is unavoidable if no information is to be discarded when forecasting. For example, at the end of each month, the variables with a publication lag of one month will have one more observation in the data set than the variables with publication lag of two months. With the publication lag affecting the structure of the data set at any given time, there are three main determinants of the value of an explanatory variable. The first determinant is the information content of the variable. If it contains only information that comes from other variables, in the sense that it is highly collinear with those variables, then the addition of this variable will not affect the estimates of the latent factors too much, and thus the forecasts made using these dynamic factors. The second determinant is the timeliness of the release of a variable. The shorter the publication lag, or the earlier its release date is within a month, the more useful the variable is likely to be. The last determinant is data revisions. Even though its effect 9 In all of the exercises that follow, we regress consumption on the factors and confidence measures and lags. As a robustness check, we have conducted these exercises using an alternative specification where the factors are extracted from a large data set including confidence, and consumption is regressed on these factors alone. Our conclusions from these exercises remain the same. 10 To avoid the need to forecast consumer confidence before we can forecast consumption expenditures, we use lagged confidence measures in our model (with the minimum amount of lagging). 12

13 on forecast accuracy is unclear, there should be no doubt that the role a variable plays in forecasting in real time could not be fully revealed without considering the effect of data revisions. Corresponding to the three determinants above, we conduct three exercises. In the first exercise, we examine the in-sample fit of the model with and without the entire series of a confidence measure. This exercise focuses mostly on the first determinant, the information content. In the second exercise, we attempt to reconstruct snapshots or vintages of the jagged-edge data set based on a stylized calendar of data release schedules and publication lags. Using this data, we examine the accuracy of consumption forecasts made with and without the latest release of a confidence measure while all the historical values of this measure are always in the data set. This exercise assesses the value of the timeliness of the data release and of the short publication lag of consumer confidence measures. In the last exercise, we construct real-time data sets as they were actually available to forecasters in the past, and use these data sets to examine the role of consumer confidence measures in real-time forecasting. In addition to the two determinants considered above, this exercise accounts for the effect of data revisions on forecast accuracy. As before, we repeat all three exercises using both monthly and quarterly data. Most of our explanatory variables come from the data set used by GRS. This data set consists of nearly 200 macro variables for the US economy starting in January, These variables, most of which are at the monthly frequency, include real and monetary quantities, prices, and surveys. We construct all the variables used in the exercises in Section 3 and add them to their data and we also remove the UM sentiment indices, the consumption expenditures, and real GDP. Real time data for personal consumption expenditure and its components come from the Federal Reserve Bank of Philadelphia s Real-Time Data Research Center. For quarterly consumption expenditures, monthly vintages are used instead of the quarterly vintages, since we produce forecasts of quarterly consumption expenditures at a monthly frequency Marginal Impact of Consumer Confidence on Forecast Accuracy Using in-sample fitted values, we first measure the difference in mean squared errors (MSE) between models of consumption expenditure with and without the entire series of a confidence measure. This is the first exercise discussed above. We do so for all possible pairs of confidence measures and types of consumption ex- 13

14 penditure. The results are presented in Table 4, where a relative MSE value that is smaller than 1 means that forecasts using consumer confidence too have a smaller MSE. Table cells like this are shaded. In this and all subsequent exercises, we test the reported differences in MSEs using the Diebold and Mariano (1995) test with its small sample modification by Harvey, Leybourne and Newbold (1997).¹¹ Whenever the difference between two competing MSEs with and without using a consumer confidence measure is statistically significant at 10%, we report the relative MSE in bold. We first confirm that the overall fits of our models are satisfactory. Adding a confidence measure improves the fit of all 48 models, and more often than not, this improvement is statistically significant. On average, adding a consumer confidence measure reduces the in-sample MSE by 7%. The models of services and total consumption benefit the most from this addition, with reductions in MSE at 12% and 10% respectively. We also find that the improvements due to the addition of a UM sentiment measure and a CB confidence measure are similar in monthly models. But in quarterly models, adding a UM sentiment measure leads to an average of 15% reduction in MSE 5% more than the average resulting from adding a CB confidence measure. Quite clearly, the message from the previous section is not only confirmed but also reinforced here: Even when considered in a rich information context, confidence matters when forecasting consumption. We then proceed with the second exercise, where we forecast consumption expenditure using reconstructed data sets that reflect the varying data release schedules and publication lags across explanatory variables. Each month, for each pair of consumption and confidence, we make two sets of forecasts, one before we observe the confidence measure for that very month, and one after we observe it. By comparing the MSEs of the two sets of forecasts, we reveal the value of release timing and publication lag of the consumer confidence measures in forecasting consumption. The results of this exercise should be interpreted differently depending on the frequency of the data. 11 The standard Diebold Mariano test gives conservative results in MSE comparisons between nested models. Therefore, the true contribution of confidence measures could be more significant than that indicated by the p-values reported in Table 4 and all subsequent tables. See Clark and McCracken (forthcoming, chap. 14). 14

15 With quarterly data, for each quarter, we place ourselves in three different points in time, namely the day of each of the three months of the quarter when the confidence measure is released. On each of the three days, we make ten forecasts. Five of them (horizons 1 to 5 quarters ahead with horizon 1 being the earliest quarter for which we do not have any official release of consumption) are based on the information set that includes this latest release of confidence measure, and the other five are based on the information set that does not. With monthly data, things are simple. We make ten forecasts at the end of each month, five with the latest release of the confidence measure, and five without. The forecasts labeled horizon 1 are always for the month we are at (the current month). The actual value for that month becomes available before we make the forecast at the end of the next month. The results from this exercise are summarized in Table 5, where a relative MSE value that is smaller than 1 means that forecasts with the latest value of confidence have a smaller MSE. Table cells like this are again shaded. From the table, we observe that in 55% of the cases with the UM sentiment measures, the forecasts made with the latest value have lower MSEs, with the differences averaging to about 4.4%. Similarly, in 64% of the cases with the CB confidence measures, adding the latest value decreases the MSE by an average of 3.4%.¹² It should be noted here that the improvement that we get when we include the latest value of confidence is often not statistically significant. However, even if our focus is on statistical significance, a case can be made for including the latest observation of confidence when forecasting consumption: The number of times that a statistically significant improvement is encountered is 41.4% higher than the number of times with a significant deterioration. Conversely, the chance that by including the latest observation of confidence we obtain forecasts that are less accurate in a statistically significant way is only 6%. The results from this and the previous subsection suggest that confidence measures lead to a noticeable improvement in the accuracy of consumption forecasts. Such improvement can be attributed to both the information content of confidence measures and the timeliness of their releases. However, the effect of data revisions, i.e., the third determinant as discussed before, remains unaccounted for. 12 In most cases where an improvement in out-of-sample forecasting performance is observed, the dependent variable is either services or total consumption. This is consistent with the results from the previous exercise. 15

16 4.3. Forecasting Consumption in Real Time In this exercise, we consider a more realistic setup, which allows us to examine the effect of all three determinants affecting the role of consumer confidence in forecasting consumption expenditure, i.e., the information content, the timing and lag of data releases, and data revisions. So, in contrast to the pseudo-real-time exercises of the previous subsection, here, we create a series of true real-time data sets from March 2005 to September 2011 for the last Friday of each month for the UM sentiment measures, and for the last Tuesday of each month for the CB confidence measures, using the stylized data release schedule and publication lags, as well as information on data revisions associated with the GRS data set.¹³ We repeat this exercise for both quarterly and monthly consumption data, just like with all previous exercises. On each of the 79 Fridays (or Tuesdays) of our sample, we produce ten forecasts, similar to the second exercise in the previous subsection. The interpretation of results is the same as in the second exercise. Table 6 presents the results. Similar to Table 5, shaded cells are ones in which the relative MSE is lower than 1, i.e., adding a consumer confidence measure improves the model s performance. MSE numbers reported in bold are those that are statistically significant at 10% based on the modified Diebold-Mariano test. In the vast majority of cases (84% of the time for the UM and 95% of the time for the CB measures), adding a confidence measure notably improves the real-time forecasting performance of the model. On average, adding the UM sentiment measure leads to a 6.8% reduction in the MSE of real-time forecasts; adding the CB measure leads to a 14% reduction. This improvement is often statistically significant, specifically, in 30.4% of all cases. Conversely, the instances in which adding a confidence measure leads to a statistically significant deterioration are quite rare only 1.5% of the time. Three additional observations are in order here, as they relate to patterns that repeat in both this and previous exercises. First, looking across components of personal consumption expenditures, we note that most of the sizable improvements to the forecasting performance (resulting from the addition of the latest confidence information) are for models of services and total 13 All the details on the schedule of data releases and publication lags are available in GRS. The real time data for earlier than March 2005 is not readily available for all variables. The data, which were kindly provided to us by David Small, are the series used by Giannone et al. (2010), minus a few proprietary series. 16

17 consumption. For example, when the CB confidence measures are used, improved forecast accuracies are noted for all models of services and total consumption (for all but one, models of non-durable goods consumption are improved as well). Second, across measures of consumer confidence, the expectations component outperforms the current conditions component in contributing to the forecast accuracy. Third, regarding the effect of data frequency on forecast accuracy, quarterly models benefit more from the addition of the latest observation on confidence than the monthly models do. The average reductions in MSE across all models, including ones to which adding sentiment leads to deteriorated performance, is about 5.5% for monthly models and 12% for quarterly models. Overall, the main results from this real-time exercise, which is also the most realistic of the three, strongly establish the positive and significant effect of consumer confidence measures on the accuracy of consumption forecasts. In an overwhelmingly large proportion of cases, including only one additional number, i.e., the latest release of a confidence measure, reduces the MSE of the resulting forecasts by about 5% to 10%. 5. e Role of Consumer Confidence: Macroeconomic Perspectives Given this clear evidence from the two sections above establishing that consumer confidence plays an important role in predicting consumption, a question that naturally arises is whether influential macroeconomic theories can account for this strong effect. Two related theories that have been considered in the literature (see, inter alia, Ludvig- son (2004)) in the context of the relationship between confidence and consumption are the precautionary savings theory and more generally the permanent income hypothesis and other departures from it. So, in what follows, we consider each of them in turn A Precautionary Savings Perspective One way to understand the role of consumer confidence in predicting consumption is to look at it from a precautionary savings perspective. If consumer confidence represents consumers expectation of future uncertainty, then a higher value of confidence, i.e., lower level of uncertainty, should be associated with less precautionary savings. Therefore, consumer confidence should be positively correlated with 17

18 consumption expenditures contemporaneously. This also means that the growth of consumption, from the current period to the future period, should be negatively correlated with consumer confidence. In a model of consumption growth where consumer confidence is a predictor, the coefficient of consumer confidence should therefore be negative. This is an ex-ante plausible hypothesis, for which Souleles (2003) provided evidence at a micro level using household data from the Survey of Consumers. However, at the aggregate level, previous studies have found no evidence of the negative relationship between the level of confidence today and consumption growth tomorrow (see Carroll, Fuhrer and Wilcox (1994), Bram and Ludvigson (1998), and Ludvigson (2004)). In this study, we find the relationship between the two to be positive almost always. For instance, for the univariate approaches of Section 3.1 we find negative sums of coefficients of lagged confidence in 6 out of 48 cases (12 each for quarterly and monthly frequencies, for both UM and CB measures). For the model with additional explanatory variables of Section 3.2, the respective count of negatives is 0 out of 48, and for the in-sample exercise of Section 4.2 the count is again 0 out of 48. We thus confirm in a robust way in our setting the findings of the aggregate-level studies cited above: A precautionary savings approach generally cannot account for the contribution of consumer confidence to the accuracy of consumption forecasts Permanent Income Hypothesis and Departures If we think of the relationship between consumer confidence and consumption through the permanent income hypothesis in general, this implies that consumption can only change because of unexpected changes in permanent income. Of course it could very well be the case that the findings of this and other studies on the significance of confidence for forecasting consumption arise through an income channel ; that is confidence forecasts income, which is in turn closely tracked by consumption. Indeed, Ludvigson (2004) provides evidence that confidence measures have statistically significant predictive power for labor income growth, and for non-stock market wealth growth. To investigate this issue, we adapt the approach of Carroll, Fuhrer and Wilcox (1994) and Ludvigson (2004) to our context. They employ a two-step procedure. In the first step, they predict labor income growth and non-stock wealth growth using 18

19 four lags of the dependent variable, consumption growth, returns to the S&P500 index, the 3-month Treasury bill rate, and consumer confidence. In the second step, they regress consumption expenditure on predicted labor income growth, or predicted non-stock wealth growth, and four lags of consumer confidence. If confidence forecasts consumption only through the income channel, the four lags of confidence in the second-step regression should not be jointly significant. We extend their approach by replacing their first-step explanatory variables (except for confidence) by the estimated dynamic factors from the model used in the previous section. Tables 7 and 8 present the results from these exercises on the significance of various measures of confidence in the second-stage regression when labor income growth (Table 7) or non-stock wealth growth (Table 8) are used. They show that consumer confidence is often significant in these second-step regressions, and thus they show that it has clear predictive power beyond that provided by predicted income growth or wealth growth. We again confirm the findings of the studies cited above: The role of confidence in forecasting consumption cannot be adequately explained by the income channel. Thus the direct relationship between confidence and consumption growth cannot be fully accounted for by a framework centered on the permanent income hypothesis and departures from that hypothesis, such as liquidity-constrained consumers. 6. Concluding Remarks In this paper, we examine the role of consumer confidence surveys in forecasting personal consumption expenditure. Existing models in the literature rely mostly on simple regressions and are limited in terms of data frequency, data vintages, and number of predictors used. So, in a first step, we revisit and extend these models using both quarterly and monthly data, both in real-time and using revised data vintages. Our exercises provide concrete evidence, in a more realistic and general context, of the notable contribution of confidence measures to the in-sample fit of personal consumption expenditure. We then further consider the ability of confidence to forecast consumption in a even more robust way that accounts for data frequency and vintage issues in a rich information context. We use a dynamic factor model with nearly 200 explanatory variables in a real-time jagged-edge data set. In this framework, we first examine the effect of the whole confidence series on the in-sample fit. Then, we examine 19

20 the contribution of the latest release of consumer confidence measures on forecast accuracy, accounting for varying release schedules and publication lags for all explanatory variables. Lastly, we perform our most realistic forecasting exercise that helps to unveil the real-time effect of consumer confidence on the consumption forecasts. The results from our analysis establish that measures of consumer confidence in general make a notable and positive contribution in forecasting personal consumption expenditure. And this is a finding that standard macroeconomic theories cannot fully account for. References Akerlof, G.A. and R.J. Shiller (2010): Animal Spirits, New Jersey: Princeton University Press. Bram, J. and Ludvigson, S. (1998): Does Consumer Confidence Forecast Household Expenditure? A sentiment index horse race, Federal Reserve Bank of New York. Carroll, C. D., Fuhrer, J. C., & Wilcox, D. W. (1994): Does Consumer Sentiment Forecast Household Spending? If So, Why? e American Economic Review, 84(5), Clark, T. E., & McCracken, M. W. (forthcoming). Testing for unconditional predictive ability. In G. Elliott & A. Timmerman (Eds.), Handbook of economic forecasting, Vol. 2. Amsterdam: Elsevier. Croushore, D. (2005): Do Consumer Confidence Indexes Help Forecast Consumer Spending in Real Time? North American Journal of Economics and Finance, 16, Croushore D. and T. Stark. (2001): A Real-Time Data Set for Macroeconomists, Journal of Econometrics 105(1), Diebold, F.X. and R.S. Mariano (1995): Comparing Predictive Accuracy, Journal of Business & Economic Statistics, 13, Fuhrer, J.C. (1993): What Role Does Consumer Sentiment Play in the U.S. Macroeconomy? New England Economic Review, January/February, Giannone, D., Modugno, M., Reichlin, L., & Small, D. (2010). Nowcasting in real-time. Mimeo. July 13. Giannone, D., L. Reichlin, and D. Small (2008): Nowcasting: The Real-Time Informational Content of Macroeconomic Data, Journal of Monetary Economics (55), Harvey, A., S.J. Leybourne, and P. Newbold (1997): Testing the equality of prediction mean squared errors, International Journal of Forecasting (13), Howrey, E.P. (2001): The Predictive Power of the Index of Consumer Sentiment, Brookings Papers on Economic Activity, 1, Keynes, J.M. (1936): e General eory of Employment, Interest, and Money, New York: Harcourt Brace. Lahiri, K. and G. Monokroussos (2012): Nowcasting US GDP: The role of ISM Business Surveys, International Journal of Forecasting, forthcoming. 20