1 Introduction

The National Institute of Economic Research each month asks representatives of Swedish firms and households about the present economic situation and the outlook for the near future. The information is compiled in the Economic Tendency Survey—a publication whose purpose is to be a timely available source of indicators for key economic variables.

The Economic Tendency Survey is the largest survey of its kind in Sweden. It is widely used for both forecasting and economic analysis concerning the Swedish economy. For those using the survey for such purposes, one problem is that most of the data presented are seasonally adjusted. In addition, a number of variables—primarily so called confidence indicators—are also standardised to have a mean of 100 and a standard deviation of 10. Such adjustments of the data mean that looking at a time series today, the value that it takes on for a particular point in time is likely to differ from what the value was according to an earlier publication/release. This problem is by no means unique to the Economic Tendency Survey; it is a well-known issue that researchers are aware that they may have to address when dealing with, for example, national accounts data (which can be substantially revised over time). The natural way to address the problem is to use data that reflect the information that an analyst, policymaker or forecaster would have had access to in real time; see, for example, Diebold and Rudebusch (1991), Croushore and Stark (2001), Orphanides (2001) and Herrmann et al. (2005) for important contributions on this topic. But while the usage of real-time data is the preferred solution to this type of problem, it is not always a feasible one. The typical obstacle to using real-time data is that they simply are not available. In some cases, real-time data can be created by going through historical records but often this is not possible.Footnote 1 For the Economic Tendency Survey, no real-time data set of relevant length exists, nor is it possible to construct one through the historical records.Footnote 2

In this paper we therefore document the creation of two quasi-real-time data sets for the Economic Tendency Survey which we have made publicly available.Footnote 3 By “quasi-real-time data” we mean data which are not actual real-time data but which have been created in order to provide a close approximation to real-time data. For several purposes, this will constitute a good enough approximation to the data that the analyst, forecaster or researcher is interested in. One application for which the quasi-real-time data are highly relevant is the evaluation of models built for nowcasting. We accordingly illustrate the potential usage of the data set with an out-of-sample nowcast exercise for Swedish GDP growth in which we rely on data from the Economic Tendency Survey as explanatory variables.

The remainder of this paper is organised as follows: In Sect. 2, we describe the Economic Tendency Survey. The construction of the quasi-real-time data sets is explained in Sect. 3. In Sect. 4, we illustrate the usability of the data by conducting an out-of-sample nowcast exercise for Swedish GDP growth. Finally, Sect. 5 concludes.

2 The Economic Tendency Survey

The Economic Tendency Survey of the National Institute of Economic Research is the largest survey of its kind in Sweden, including more than 6000 companies and 1500 households. Below we give a short description of the survey. For more details, the reader is referred to the user guide (National Institute of Economic Research, 2013).

2.1 Businesses

Stratified sampling of firms takes place through Statistics Sweden’s business register. The companies are divided into four main categories: manufacturing industry, construction industry, trade and private service sector. The questionnaires—which approximately half of the companies nowadays receive in electronic form—are addressed to upper management and the questions relate to the development in recent months, the present situation and the outlook for the near future regarding, for example, output, new orders, employment and prices. For a detailed description of each question, see Appendix 1.

Every third month—in January, April, July and October—the business survey contains more questions; for a detailed description of the questions, see Appendix 2. This is why we generate two data sets, one monthly and one quarterly.

2.2 Households

A random net sample of 1 500 individuals, between 16 and 84 years of age, are interviewed over the telephone. The questionnaire contains questions concerning both to the household’s own economic situation and the aggregate economy. For a detailed description of the questions in this part of the survey, see Appendix 3.

2.3 Presentation of the Results

For questions of binominal or multinomial type (where the answers, for example, are increase/unchanged/decrease), the summarised weights for each response alternative are standardised so that the percentages of the response alternatives add up to 100. In order to simplify the presentation of the data, the concepts “net figures” or “balances” are employed (and used equivalently). A net figure/balance is the difference between the share of respondents reporting an increase and a decrease for a particular question. For example, if 45% of respondents report that there has been an increase, 25% that there has been no change and 30% that there has been a decrease, the net figure/balance is 45 − 30 = 15.Footnote 4

As an illustration, consider the time series in Fig. 1 which shows the seasonally adjusted net figures for the output volume over the past three months in the manufacturing industry (question BTVI101, see Appendix 1) as given by the Economic Tendency Survey in December 2015. This shows, for example, how in April 2009—as the effects of the global financial crisis were seriously affecting the Swedish economy—there were substantially more companies which had decreased their production relative to those who had increased it. The net figure of almost −40 reflects that approximately 54% of the companies (weighted share) in the survey stated that they had decreased their production whereas 14% (weighted share) had increased it; the remaining 32% (weighted share) had left it unchanged.

Fig. 1
figure 1

Output volume over the past 3 months in the manufacturing industry. Note Seasonally adjusted net numbers on the vertical axis. Data are from the Economic Tendency Survey of December 2015

Full size image

Not all series are presented using net figures though. For example, households’ inflation expectations at the twelve-month horizon (question Q063, see Appendix 3), which are shown in Fig. 2, are simply given in percent.

Fig. 2
figure 2

Households’ inflation expectations. Note Percent on the vertical axis. Data are from the Economic Tendency Survey of December 2015

Full size image

Having briefly described the survey, we next turn to the construction of the quasi-real-time data sets.

3 Quasi-Real-Time Data

The raw data that underlie indicators and individual questions in the Economic Tendency Survey are not revised, except on the relatively rare occasion when an error has to be corrected. This means that for the raw data, the latest data vintage is almost identical to true real-time data (differing only with respect to the corrected errors).Footnote 5 The quasi-real-time data can therefore be generated in a straightforward manner. For each vintage, we conduct the following three steps for individual questions:

  1. (1)

    Set the sample to the relevant time period.

  2. (2)

    Copy the raw data of the series.

  3. (3)

    If applicable: Seasonally adjust the series.Footnote 6

For indicators and other variables that are generated based on two or more series, the following additional steps are conducted:Footnote 7

  1. (4)

    If applicable: Standardise the series used in the calculation.

  2. (5)

    Weigh the (potentially standardised) series together using the appropriate weights.

  3. (6)

    If applicable: Standardise the weighted series. For example, confidence indicators have a mean of 100 and a standard deviation of 10.

Based on the above method, we generate vintages of time series which show how different variables in the Economic Tendency Survey would have looked in real time if today’s methods concerning seasonal adjustment, standardisation and weighing would have been employed. For recent vintages, our method generates data that generally are identical to what proper real-time data would have looked like. For older vintages, on the other hand, there will be differences. The difference will partly be due to methodological changes—such as different methods for seasonal adjustment and different choices concerning standardisation—but also to the above mentioned fact that errors may have been corrected. Whether this difference matters depends on the purpose behind using the data. For example, if one wants to conduct a study of how the Economic Tendency Survey has affected stock prices, the exact information content of the survey in real time is a key issue; the present quasi-real-time data set should accordingly not be used in such a case. If, on the other hand, one is interested in developing a new forecasting model based on variables from the survey and want to evaluate its out-of-sample forecast performance, the quasi-real-time data set should be very useful. In this case, the issues that we mainly are concerned might distort the analysis, namely seasonal adjustment and standardisation, have been addressed when creating the data set. For this purpose the quasi-real-time data should accordingly be close to a perfect substitute to actual real-time data.

3.1 A Monthly Quasi-Real-Time Data Set

The questions, indicators and other variables included in the monthly data set are provided in Appendices 1 and 3. We generate 192 vintages of monthly data. The first reflects the Economic Tendency Survey of January 2000. The last reflects the survey of December 2015. Since different questions have been included in the survey at different points in time, the number of variables included in the survey varies with the vintage. The vintage from which a particular question, indicator or other variable is available—as well as the starting point of the time series in question—can also be found in Appendices 1 and 3. For example, the Economic Tendency Indicator (KIFI) is available from the first data vintage, that is, that of January 2000. Its starting date is July 1996 which means that the first vintage has 43 observations. Households’ expectations on the variable home loan rate at the 1-, 2- and 5-year horizon (questions Q183, Q193 and Q203, see Appendix 3), on the other hand, are available only from the vintage of February 2010; this date also corresponds to the first observations for these three series.

As was described above, many time series have been seasonally adjusted. For individual questions, this is indicated with the suffix “S” in the name of the variable.Footnote 8 For example, net figures for the output volume over the past three months in the manufacturing industry are given by BTVI101 and the seasonally adjusted net figures are given by BTVI101S.Footnote 9 It should be noted though that while indicators such as the consumer confidence indicator (BHUSCON) and the Economic Tendency Indicator (KIFI) are based on seasonally adjusted data, they do not have the suffix “S”. The seasonal adjustment is conducted on no less than three years of data. This means that in some cases, the first vintage in which a variable is included will differ between the original series and the seasonally adjusted one; in the cases where the first vintage differs, this is indicated with two dates for the first vintage in which the series in question is included. The date for the seasonally adjusted data is indicated with the suffix “S”. For example, January 2000 is the first vintage in which BTVI107 is included. The first vintage in which the seasonally adjusted series BTVI107S is included is July 2002 (since the first observation of BTVI107 is July 1999 and seasonal adjustment is conducted using no less than three years of data); see Appendix 1 for details.

Figure 3 shows three different vintages of the Economic Tendency Indicator (KIFI) from the monthly quasi-real-time data set. As can be seen, the three vintages look fairly similar but they are not identical. As an example, we can note that between September 1996 and February 1997, the difference between the January 2006 and December 2015 vintages is small (always less than 0.8 index units). However, between March and August 2000, the difference between the same two vintages is never less than 4 units.

Fig. 3
figure 3

Different vintages of the Economic Tendency Indicator. Note Index numbers on the vertical axis. “KIFI” is the Economic Tendency Indicator. Data are from the quasi-real-time data vintages of January 2000, January 2006 and December 2015

Full size image

In Fig. 4, different vintages are shown of the seasonally adjusted figures corresponding to the (weighted) share of companies in the construction industry that answered that the weather currently was the main obstacle to increased activity (BBOA1075S). Note that the sample in the figure is January 1988 to December 2000 in order to make differences clearer.Footnote 10 Looking at the figure, we see that there is no difference at all—at any point in time—between the January 2006 and December 2015 vintages.Footnote 11 Comparing these series to the January 2000 vintage though, it is clear that the seasonal adjustment matters from a real-time perspective, even if the differences by no means are dramatic. For example, the seasonally adjusted value for June 1999 is 3 in the January 2000 vintage but 9 in the January 2006 and December 2015 vintages.

Fig. 4
figure 4

Different vintages of the share of companies in the construction industry whose main obstacle to increased activity was the weather, January 1998 to December 2000. Note Percent on the vertical axis. Seasonally adjusted data are from the quasi-real-time data vintages of January 2000, January 2006 and December 2015

Full size image

Above we have shown just a few examples of how time series look different depending on data vintage. We will not illustrate this issue further since it now should be clear that the real-time aspect of data could matter when using the Economic Tendency Survey for analysis. The full monthly data set can be downloaded from www.konj.se/quasi-real-time-data.

3.2 A Quarterly Quasi-Real-Time Data Set

Turning to the quarterly data, the questions, indicators and other variables included in the data set are given in Appendix 2. As was the case for the monthly data set, both raw and seasonally adjusted data have been made available. And just like above, the seasonally adjusted series have the suffix “S”. It should be noted that the quarterly data set only contains data from the business survey (since it is this part of the survey that differs in January, April, July and October). We generate 64 vintages of quarterly data. The first reflects the Economic Tendency Survey of January 2000. The last reflects the survey of October 2015. Similar to the case of the monthly survey, different questions have been included in the survey at different points in time; the number of variables included in the survey therefore varies with the vintage also in this case. The vintage from which a particular question, indicator or other variable is available—as well as the starting point of the time series in question—can also be found in Appendix 2. The full quarterly data set can be downloaded from www.konj.se/quasi-real-time-data.

4 Empirical Illustration: Nowcasting GDP Growth Using Indicators

The main purpose for developing the data sets presented above is to allow for model-based out-of-sample nowcast or forecast evaluations to be conducted in the best possible manner. Variables from the Economic Tendency Survey are often used as explanatory variables in models developed for this purpose since it is assumed that they provide timely information on the economic situation which is relevant to nowcasters/forecasters. However, until recently such evaluations have typically relied on ex post data from the survey since these are the ones that have been available.Footnote 12 For variables that are seasonally adjusted and/or standardised, this introduces an error which could cause, for example, forecast precision being over- or understated. If one is unfortunate, this could lead to the wrong conclusions being drawn. Having developed the quasi-real-time data sets for the Economic Tendency Survey, one can now minimise the errors when conducting this type of analysis. We next illustrate this issue in an empirical application where we nowcast Swedish GDP growth.

By “nowcast”, we mean that we are trying to predict the GDP growth associated with a certain quarter when standing partway through the quarter in question.Footnote 13 Seeing that GDP is a national accounts variable which tends to be revised as time passes, we follow standard practice and use real-time data on GDP growth for our analysis.Footnote 14 We compare the nowcasting performance of models using the Economic Tendency Indicator and six confidence indicators from the Economic Tendency Survey for Swedish GDP growth in two cases. First, we use ex post data from the Economic Tendency Survey; these are given by the December 2015 vintage. One quarterly time series per variable is constructed by using the February observations for Q1, the May observations for Q2, the July observations for Q3 and the November observations for Q4.Footnote 15 Second, we use the monthly quasi-real-time data described above. In a similar manner to the ex post data, quarterly time series are generated based on the monthly data. However, in this case, one quarterly time series per variable is constructed for each data vintage.

The out-of-sample nowcast exercise is conducted the following way: A number of models (specified below) are used. The first nowcast employs data on GDP growth from 1996Q3 to 2003Q4 which reflect national accounts data that were released in early March 2004. At this point in time, the survey data that would have been available for nowcasting correspond to the Economic Tendency Survey published in February 2004 (where the value for this month is used for 2004Q1 as described above). Based on these data, we estimate the models and a nowcast for 2004Q1 is generated from each model. The sample is then expanded one quarter, the models re-estimated and new nowcasts are generated. This process continues until the end of the sample period; the final nowcast is based on GDP data from 1996Q3 to 2015Q2, which are used to generate a nowcast for 2015Q3. We accordingly evaluate 47 nowcasts for each model.

Two simple models are used as reference points. First, we rely on the model:

$$ g_{t} = c + e_{t} $$
(1)

where g t  = 100(Y t  − Y t − 1)/Y t − 1, where Y t is GDP in period t and e t is an error term. Since this model’s only explanatory variable is a constant, its nowcast will be the estimated historical average. It accordingly provides a natural benchmark for other models to outperform. A model which does not have higher forecast precision than such a simple alternative appears to be of limited usefulness. Concerning the models which do have higher forecast precision than the model in Eq. (1), we say that they have positive nowcast content.Footnote 16 Second, we also employ an AR(1) model:

$$ g_{t} = c + \rho g_{t - 1} + e_{t} $$
(2)

Due to its simplicity and flexibility, the AR(1) model is a frequently used benchmark in applied macroeconomic work.Footnote 17

Apart from these two models, we also estimate six models for each set of survey data (that is, ex post and quasi-real-time). Each model has the form

$$ g_{t} = c + bS_{t} + e_{t} $$
(3)

where S t is an indicator based on the survey data.

In addition to these models, we finally report the forecast precision from a naïve forecast. This simply states that GDP growth in the present quarter will be the same as that of the previous quarter:

$$ \hat{g}_{t|t} = g_{t - 1} $$
(4)

where \( \hat{g}_{t|t} \) is the nowcast of GDP growth at time t made at t.

We rely on the mean absolute error (MAE) and root mean square error (RMSE) of the nowcast as our criteria to evaluate the nowcast performance of the models. Defining the nowcast error as \( v_{t + i|t + i} = g_{t + i} - \hat{g}_{t + i|t + i} \), where g t+i is the outcome and \( \hat{g}_{t + i|t + i} \) the nowcast, these measures are given as

$$ MAE = \left( {1/n} \right)\mathop \sum \limits_{i = 0}^{n - 1} \left| {v_{t + i|t + i} } \right| $$
(5)

and

$$ RMSE = \sqrt {\left( {1/n} \right)\mathop \sum \limits_{i = 0}^{n - 1} \left( {v_{t + i|t + i} } \right)^{2} } $$
(6)

Results from this exercise can be found in Table 1. We first turn to the analysis using ex post data which can be found to the left. As can be seen, the MAEs indicate that all versions of Eq. (3) except that relying on BBOACON (the confidence indicator for the construction industry) have higher forecast precision than Eq. (1). Put differently, all investigated variables except BBOACON appear to have positive nowcast content. Comparing the MAEs to each other, we find that the lowest MAE is found for BTVICON (the confidence indicator for the manufacturing industry), closely followed by BHUSCON (the consumer confidence indicator) and KIFI (the Economic Tendency Indicator). Turning to the RMSEs, the results are quite similar. We again find that all versions of Eq. (3) except that relying on BBOACON have higher forecast precision than Eq. (1). In terms of the ranking of the models, this is slightly different from when using the MAEs for evaluation. We now we find that the lowest RMSE is found for BHUSCON, followed by KIFI and BTOTCON (the confidence indicator for the total business sector).

Table 1 RMSEs when nowcasting GDP growth
Full size table

Overall though, differences in forecast precision between most models using survey data are small and should not be over-interpreted. In order to illustrate this, consider the fact that judging purely by the MAEs and RMSEs—as we did above—all investigated variables from the Economic Tendency Survey except one appear to have positive nowcast content. But what does a formal test have to say about this issue? We assess this by testing whether the differences in nowcast precision between the model in Eq. (1) and the other models are statistically significant. This is done by conducting Diebold-Mariano tests (Diebold and Mariano 1995) under an assumption of a quadratic loss function.Footnote 18 That is, we run the regression

$$ \left( {v_{t|t}^{1} } \right)^{2} - \left( {v_{t|t}^{comp} } \right)^{2} = c + \omega_{t} $$
(7)

where v 1 t|t is the nowcast error from Eq. (1) and v comp t|t the nowcast error from the competing model; ω t is an error term. The nowcasts from the competing models are based on Eqs. (2), (3) and (4). We test the null hypothesis that the parameter c in Eq. (7) is equal to zero using a t test based on Newey–West standard errors (Newey and West 1987). As can be seen from the table, the Diebold–Mariano test is statistically significant (at the 10% level or less) for BHUSCON, KIFI, BTOTCON and BTVICON. For BTJACON and BHANCON though, the difference is not statistically significant; this is despite the fact that the model based on BTJACON has an RMSE which is 12% lower than that of the model in Eq. (1).Footnote 19

The results just discussed illustrate conclusions one might draw from a very simple out-of-sample nowcast exercise where ex post data for the Economic Tendency Survey have been used. So how sensitive are the conclusions drawn with respect to the fact that we relied on ex post data? Were MAEs and/or RMSEs over- or understated? Were any incorrect conclusions drawn? In order to answer such questions, we next turn to the analysis using quasi-real-time data.

As can be seen from the right part of Table 1, it can first be noted that the MAEs and RMSEs using the quasi-real-time data in all cases are close to those when ex post data were used. In general, both measures tend to be lower when using the quasi-real-time data; the MAE and RMSE are higher only for BTJACON (the confidence indicator in the service sector). Overall though, it seems fair to say that no serious over- or understatement of forecast precision was introduced by using the ex post data. If we look at the relative performance of the different indicators, we see that the ranking of the models is similar; it can be noted though that the best indicator judging by the MAE is now found to be BHUSCON rather than BTVICON (as we found when using the ex post data above).Footnote 20 It should, however, be pointed out once again that the differences in forecast precision between most models are small. Finally, turning to the results from the Diebold–Mariano tests, we see that these are almost identical to when the ex post data were used. One minor difference can be found, namely that for KIFI the null hypothesis can be rejected at the 5% level (rather than the 10% level that was the case when employing the ex post data).

Summing up this exercise, we established that there were only minor differences in the results when using quasi-real-time data instead of ex post data. Our findings suggest that for the simple models studied here, one would not have been seriously mislead by using the ex post data. This conclusion should, however, not be interpreted as a reason not to use the quasi-real-time data. Because while we have established that the error committed by using ex post data in this particular application was small, we had to conduct the above analysis to reach this insight. If this had not been done, we simply would not have known this. Using larger and/or more complicated models, results may very well look different. In practice, the researcher never will know exactly how big the error from using ex post data is in a given application unless analysis similar to that in this paper is conducted. To conclude, we note that real-time data have become the benchmark when it comes to certain analysis relying on national accounts data, such as forecasting and nowcasting. We recommend that the quasi-real-time data sets presented in this paper should be employed in a similar manner for analysis relying on the Economic Tendency Survey.

5 Conclusions

In this paper we have documented two quasi-real-time data sets of the Economic Tendency Survey. The data sets consist of monthly/quarterly vintages of the most important series of the survey. A natural usage of these data sets is evaluations of model-based nowcasts and forecasts. We have accordingly illustrated how the data sets can be employed by conducting an out-of-sample nowcast exercise for Swedish GDP growth in which data from the Economic Tendency Survey act as explanatory variables. This shows that several of the studied indicators from the Economic Tendency Survey appear to have positive nowcast content for GDP growth.

While the quasi-real-time data do not solve all problems that the applied researcher using the Economic Tendency Survey might face, it should help address a few of them. The quasi-real-time data should be the natural starting point when conducting, for example, an out-of-sample nowcast or forecast exercise; it is after all the best approximation to real-time data that we have. The data sets should also be useful when studying a range of other questions, including issues concerning the survey itself. For example, one could investigate quality aspects of seasonal adjustment. Such analysis could contribute not only to our understanding of the properties of the time series in the Economic Tendency Survey in particular, but also to a widening regarding our knowledge of the behaviour of data based on similar surveys in general.