# A system of rapid estimates to improve real-time monitoring of the economic situation: the case of the Euro Area.

Timely availability of relevant economic information is from the
users' point of view a crucial pre-requirement for effective
economic and monetary policy and for accurate analysis and reliable
forecasting. Unfortunately, official statistics are not always able to
meet such needs. For this reason, particularly over the recent past,
advanced statistical and econometric techniques have been used to
complement the classical production process in order to increase the
timeliness of the relevant economic information. Furthermore, official
statistics do not necessarily show explicitly relevant phenomena such as
the cyclical situation and the location of turning points. In order to
fill this gap, indicators have to be built up to extract such relevant
signals and to make them available in an easily readable and
comprehensible way to all users. This paper gives a general overview of
ongoing Eurostat projects, aiming to supply timely information to all
users based on transparent and statistically sound methodologies. In
particular this paper presents results related to the construction of
flash estimates in order to increase the timeliness of official
statistics as well as the construction of a timely statistical system
for business cycle analysis based on a set of turning points indicators
and cyclical estimates. The results obtained so far are very
encouraging, especially in terms of the overall quality of the estimates
and the statistical robustness of the proposed methods. Nevertheless,
some improvements, particularly related to the precision of flash
estimates, are needed, especially under an uncertain economic situation
and regime changes.

Keywords: Coincident indicators; cyclical estimates; flash estimates; monthly GDP; turning points detection JEL Classifications: C53; E32; E37

I. Introduction

Regular monitoring of the economic situation requires, especially under uncertain conditions, a complete, timely and reliable set of statistics and statistical indicators giving a clear picture of economic movements.

The Euro Area system of short-term statistics was built up in the 1990s with the adoption of a set of new legal acts. Since the beginning of the new century, when a set of Principal European Economic Indicators (PEEls) were defined, Eurostat has launched an improvement strategy focusing mainly on the reduction of publication delays and on the progressive harmonisation of data production processes at national and European level.

In addition, action has been taken to fill the availability gap, particularly in the areas of services and labour-market statistics. Nowadays, almost all PEEls are available and tangible results have been achieved concerning timeliness, especially due to the use of flash estimation techniques (for GDP and HICP) or to earlier estimates based on an incomplete set of information (for Retail Trade).

Nevertheless the situation is not yet satisfactory from the user's point of view: 1) further improvements to timeliness are required without a significant decrease in accuracy; 2) the length of Euro Area series is not as long as needed for economic modelling and business cycle analysis purposes; 3) the frequency at which data are available is, in some cases, not appropriate to ensure a regular follow-up of the cyclical situation and an effective implementation of monetary policy actions. This last drawback is particularly relevant and requires some additional thought. In the past, the Industrial Production Index was considered to be the reference variable for short-term analyses. With the continuous growth of the services sector and the associated decrease of the industrial sector, the role of the Industrial Production Index has also been substantially re-thought. In addition, the empirical evidence of cyclical movements in the services sector has convinced business cycle analysts of the need to find a new and more comprehensive measure of economic activity. This belief has been reinforced by the fact that in the recent years, some industrial cycles have not generated fluctuations for the whole economy. Finally, the high degree of volatility of the Industrial Production Index, which prevents a clear identification of signals, has also played a role in diminishing the relevance of this variable. For these reasons, ongoing studies aim to produce a more significant monthly indicator for business cycle purposes.

Another issue is that statistics themselves cannot answer the needs of all users because data give a general measure of a phenomenon. Analysts often want to extract specific signals from data in order to emphasise specific cyclical features. Gaps in statistics can be filled by specific indicators to complete the information set to be put at the disposal of analysts.

Thus, statisticians are required, starting with official statistics, to: a) extract from statistics some special signals to help improve our understanding of the economic situation; b) combine official statistics to derive new indicators filling specific gaps in data availability.

Improvements in statistics using statistical and econometric techniques, as well as compiling new indicators using existing statistics, are not traditionally considered part of official statistics. In fact, in recent years, more and more statistical agencies have been involved in such activities. It is also clear that statisticians, having comprehensive knowledge of data and of production systems, are in a privileged position to produce reliable and statistically sound estimates and indicators. Users could benefit from the availability of such estimates and indicators, using them in their analyses as well as in the decision-making process.

This paper briefly presents some current activities going on in Eurostat related to the improvements of PEEIs and to the construction of new indicators in order to help analysts and policymakers interpret correctly the economic situation.

In section 2 we discuss alternative ways to increase data timeliness and we present some preliminary results. In section 3 we describe our approach for the construction of a Euro Area monthly indicator of economic activity, which can be viewed as a very good proxy of GDP at the monthly frequency. Section 4 presents the statistical framework for business cycle analysis developed at

Eurostat. This framework consists mainly of two parts: a turning points dating and detection system and a set of growth cycle estimates. The most recent results are also presented. Section 5 presents some conclusions and lines of enquiry for further investigation and improvement.

2. Increasing data timeliness

Users, particularly institutional ones, consider the timeliness of statistics as one of their most important characteristics. In theory they should welcome real-time or quasi real-time information on main economic indicators. However, data producers are often reluctant to anticipate the release date because of the loss of precision that such anticipation could afford.

The trade-off between quality and timeliness is a key consideration when deciding on the timing of macroeconomic statistics releases. For the Euro Area the timeliness is not only relevant in absolute terms but also in comparative terms with respect to US release dates. Despite significant improvements in Euro Area release dates observed over the past 5-6 years, the gap between Euro Area and US releases is still very relevant. The only exception to this is in the case of Consumer Prices where we anticipate the US release by means of our flash estimates.

An increase in timeliness can be achieved using various strategies:

1) speeding up of the data production process, using advanced survey techniques, simplied questionnaires, use of administrative information;

2) the use of statistical and econometric techniques to extrapolate indicators by using an incomplete information set;

3) the use of so called EU sampling techniques, which consist in the construction of European samples which do not necessarily generate nationally significant ones;

4) the construction of coincident indicators for the main economic variables, which give an idea of their latest trends.

The first strategy is clearly a structural one which can be feasible in the medium-to-long term and which can be considered, from the producers' point of view, to be the most suitable approach to this issue.

The third strategy implies defining a European sampling scheme applicable to the most relevant surveys on which key economic variables are based (e.g. the industrial production index, quarterly level cross surveys, etc). This solution might be difficult when applied to National Accounts which are mainly a synthesis of various macroeconomic information. Until now, Eurostat has obtained interesting results based, at least partially, on this approach in the field of Retail Trade Turnover. In this paper we concentrate mainly on current studies related to the second and fourth strategies.

2. I Use of statistical and econometric techniques

The use of statistical and econometric techniques can contribute significantly to the increase of timeliness in the short and medium term. In this context, the key tool is represented by a set of forecasting techniques adapted to estimate the recent past or the present instead of the near future. Since 2006, we have been working on a simulation study aimed at producing flash estimates for the Producer Price Index at t+16, Employment at t+30 and, more recently, GDP at t+0, t+15 and t+30. The flash estimates, which we are working on, are based on the following main principles agreed with Eurostat Production Units:

* whenever partial information on the target variable, either at geographic or sectoral level is available, this has to be included in the flash estimation model;

* soft data (e.g. business and consumer surveys) may be integrated into the model on condition that a minimum amount of hard data is available;

* in order to increase forecasting accuracy, statistically related indicators {e.g. conventional earnings in the case of flash estimates of the Labour Cost Index) may be used in the model either in the case of the unavailability of significant partial information on the target variable or to complement such partial information;

* any hypothesis based on economic theories has to be avoided in the model specifications;

* purely univariate models should not be taken into account;

* the selected model should be as simple as possible, statistically sound, easy to use in the regular production process and characterised by a very simple dynamic specification where needed.

In the simulation exercise we pay particular attention to the identification of the most appropriate variable selection strategies and to the comparison of alternative model specifications to achieve the best performance of our estimates. Forecast combination techniques are also tested as a means of increasing the reliability of estimates.

Table 1 shows the real-time simulation for the period April 2007-March 2009 of the flash estimates for the Producer Price Index against the Eurostat first and final estimates for the Euro Area. The selected model is based on the German Producer Price Index and on oil price data according to the BIC selection criterion. Results appear quite encouraging; despite the presence of two cases of sign discordance, often the flash estimates are quite close to Eurostat first estimates. The model reacts fairly well to the change of regime even if a little slowly. Our opinion is that this constitutes a very good starting point, which could be easily improved either by adding more national information or by including Euro Area Import Prices which were not available at the start of the simulation period.

2. 2 Coincident indicators

Coincident indicators aim to forecast the evolution of economic variables during the reference period or just after it. For this purpose, they are based on the same principles as leading indicators. The main advantage of coincident indicators is that they are subject to fewer constraints than flash estimates (see section 2.1 above), even if, once again, the use/imposition of economic relationships is not recommended. In recent years, we have investigated alternative model specifications for GDP, IPI and more recently Employment. The results for the IPI are not very satisfactory mainly because of the high degree of volatility of this indicator. In this paper we briefly present our approach to constructing a coincident indicator of GDP. We have compared four different models: 1) a first bridge model containing only Business and Consumer Surveys data and financial variables, 2) a second bridge model which differs from the first by the inclusion of the IPI, 3) a bridge model based only on financial variables 4) a factor based model, where data are selected by the LARS (Least Angle Regression) algorithm. In the real-time simulation we have used, for the first part, an estimate based on combining forecasting techniques with equal weights of the three bridge models and, for the latest part, the factor model which seems to perform slightly better than the others.

For each quarter we produce three estimates of GDP: the first at the end of the second month (t-30), the second at the end of the quarter (t+0) and the last at the end of the first month of the current quarter (t+30).

Table 2 shows the real-time results obtained by using the factor model for data selection over the period 2007Q1-2009Q1, complemented by the LARS algorithm from 2009. We compare the three estimates of the coincident indicator to the Eurostat flash estimate. The results appear very encouraging. We observe only one case of sign discordance in 2008Q2, where the growth is around zero. Furthermore, the coincident indicator anticipates GDP growth correctly in most cases. In the fourth quarter of 2008, we note an unusually sized error between the estimates of the coincident indicator and the Eurostat flash one. Obviously, the simulation period presented in the table is quite short and further simulations are required, but the coincident indicator presented in this section appears to be robust and reliable enough.

3. Monthly indicator of economic activity

GDP is obviously the ideal candidate as the reference variable for short-term and business-cycle analysis but, unfortunately, it is only available on a quarterly basis and the production of monthly GDP, completely based on National Accounts standards, still appears infeasible. For this reason, several studies have been conducted recently investigating alternative ways to construct monthly proxies of GDP. Examples of such indicators are available in Sweden, Finland, Estonia and the UK, as well as in Canada. From 2006 onwards, we have investigated the possibility of constructing a Euro Area monthly indicator of economic activity as much as possible consistent with GDP.

3.1 Euro-MIND: a Euro Area monthly indicator of economic activity

The methodology proposed here is presented in more detail in Frale, Marcellino, Mazzi and Proietti (2008). It may be described synthetically as follows:

1) We base the construction of the monthly indicator of economic activity on a disaggregate approach represented by the output and expenditure breakdowns of GDP at the quarterly base.

2) For each disaggregate GDP component, a set of monthly indicators is carefully selected, including both macroeconomic variables and survey answers.

3) The indicator is based on information at both the monthly and quarterly level, rather than the monthly only, modelled with a dynamic factor specification cast in state-space form.

4) The state space methodology has the flexibility of handling data with different frequencies of observation. This is achieved by suitably defining the states of the system so as to convert temporal aggregation into a systematic sampling problem.

5) Since estimation of the multivariate dynamic factor model can be numerically complex, computational efficiency is achieved by implementing univariate filtering and smoothing procedures.

6) Special attention is paid to chain-linking and its implications for the construction of a monthly indicator of economic activity, via a multistep procedure that exploits the additivity of the volume measures expressed at previous year prices.

7) The estimate of the Euro Area monthly indicator of economic activity is obtained by combining the estimates from the output and expenditure sides, with optimal weights reflecting their relative precision.

8) The resulting pooled estimator is more precise than each of its two components, paralleling the results on the usefulness of pooling in the forecasting literature. The resulting estimates are benchmarked to quarterly national accounts produced by Eurostat so that the full consistency between monthly and quarterly estimates is achieved.

9) We provide an explicit measure of uncertainty around the indicator, which is particularly relevant in a decision-making context and for evaluation purposes.

[FIGURE 1 OMITTED]

A generalisation of this model using two factors, where the second one contains business and consumer surveys data is presented in Frale, Marcellino, Mazzi and Proietti (forthcoming). This extension increases the forecasting ability of the model at one-two-three steps ahead. A real-time simulation of the one factor-based model has been carried out since 2006 with very encouraging results. In this simulation we are producing estimates at t+45 each month, so that, in month t, we produce the estimate for the month t-2. At the same point in time, estimates for months t-1, t, t+1 can be obtained by using the two-factor version of the model.

Figure 1 presents the growth rate of Euro-MIND from January 2005 to February 2009 as estimated in April 2009, together with confidence intervals at 95 per cent. Looking at the graph, it is important to note that the evolution of the indicator is quite regular and it follows cyclical movements very well. The estimates appear very stable and not volatile, which is also confirmed by analysing subsequent vintages for the same period. The main point on which the indicator still needs some improvement is represented by the way it estimates the months of the current quarter, especially in the recession phase. Our indicator delivers negative growth rates (e.g. in January and February 2009), which appear too optimistic in comparison with the expected results. A more accurate specification of the model for the financial services sector and for the demand side component will probably improve the ability of the mode[ to estimate the most recent months.

4. A statistical framework for business cycle analysis

The set of macroeconomic statistics regularly compiled by a statistical office represents a very useful instrument available to all users and analysts. Nevertheless, we have to recognise that not all the information needed by analysts is explicitly available from an investigation of statistics. Some signals need to be extracted in order to have a clearer picture of the cyclical evolution of the economy, complementing the information supplied by statistics. In this context, we have decided to launch several activities aiming at defining a coherent statistical framework for business cycle analysis. They include the construction of statistical turning point chronologies, the development of turning point coincident indicators and estimates of the growth cycle (i.e. the output gap in the case of GDP), which can support economic monitoring and decision-making processes.

4.1 Euro Area turning point coincident indicators The methodology for the construction of a Euro Area turning point chronology and a system of coincident turning point indicators is presented in Anas, Billio, Ferrara and Mazzi (2008). The methodology can be synthetically expressed as follows:

1) simultaneous analysis of the classical business cycle and the growth cycle in the so-called ABCD framework;

2) statistical dating of Euro Area turning points by means of a simple non-parametric dating rule;

3) comparison of Euro Area and Member States dating to achieve a final statistical chronology ensuring the maximum degree of consistency between the two approaches;

4) preliminary investigation of alternative models for the construction of coincident turning point indicators for the classical business cycle and the growth cycle, including the identification of appropriate number of regimes and thresholds;

5) variables selection for the growth cycle coincident indicators (Employment expectations, Construction confidence indicator, Financial situation of the last 12 months, IPI, Imports of intermediate goods);

6) construction of the growth cycle coincident indicators (GCCI) as a weighted mean of the transition probability returned by five univariate two regime Markov Switching models fitted on each variable. An equal averaging weighting scheme is used.

7) variable selection for the business cycle coincident indicators (IPI, New cars registration and Unemployment rate);

8) construction of the business cycle coincident indicators (BCCI) as a weighted mean of the transition probability returned by three univariate regime Markov Switching models fitted on each variable. The following weighting scheme is used, IPI = 0.34, Unemployment rate = 0.46, New car registrations = 0.20.

[FIGURE 2 OMITTED]

[FIGURE 3 OMITTED]

Filtered probabilities may be viewed as the probabilities of being in a recession phase delivered by each component of the indicators. Indicators deliver the joint recession probabities. A higher value of the indicator corresponds to a high probability of being in a recession. The threshold (set at 0.5) corresponds to a decision rule; values exceeding the threshold indicate a recession phase, values below the threshold correspond to an expansionary phase.

A real-time simulation of the two indicators against respectively the business cycle and the growth cycle chronology has been carried out to check the reliability of the models as well as to discover possible false signals. The main results are that the two indicators do not show any significant evidence of false signals and that they are slightly lagging with respect to the corresponding chronologies. Each month we produce estimates of business cycle coincident indicators and growth cycle coincident indicators for the month t-2, based on filtered probabilities. Estimates for the month t-1 and t are based on forecast probabilities. Figures 2 and 3 show the behaviour of the two indicators GCCI and BCCI as estimated in April 2009. In both graphs the black bold line is the constant threshold equal to 0.5. When the indicators deliver values higher than 0.5 we are respectively in a growth cycle or business cycle recession phase. On the other hand, when the indicators deliver values below 0.5, we are in an expansion phase for both cycles. The black lines show the values of the two indicators obtained by averaging the filtered probabilities of the components. The red dashed line at the end corresponds to the value obtained by averaging forecasting probabilities instead of filtered ones. Looking at the indicators, the negative phase for the growth cycle started in April 2007 and still continues (see figure 2). Concerning the business cycle, the recession started in October 2008 and still persists (see figure 3). As already mentioned, both indicators appear to be slightly lagging and this is particularly true for the BCCI. In fact, we now think that the business cycle recession started in the first half of 2008. From this point of view it is obvious that BCCI still needs some improvement. Nevertheless, it has to be noted that it is preferable to have indicators detecting later turning points than ones delivering false signals or anticipating too many turning points.

4.2 Growth cycle estimates

The accurate and timely identification of turning points is a very important source of information for policymakers and analysts. Unfortunately, even the most accurate system of detecting turning points fails to supply details on the shape and the main features of the cycle. This is especially true for the growth cycle, which needs to be extracted for the seasonally adjusted version of indicators by means of &trending techniques. From the policymakers' and analysts' point of view, an accurate estimate of the growth cycle is of crucial importance, particularly for monitoring inflationary pressures and for designing a monetary policy oriented to inflation control. The main problem we have to deal with when estimating the growth cycle is its instability at the end of sample, due to data revisions on the one hand and to specific characteristics of most detrending filters on the other. During the past year we have compared alternative univariate techniques for the estimation of the growth cycle and we now regularly publish three alternative growth cycle estimates based on the Hodrick-Prescott filter, the Christiano-Fitzgerald filter and the Unobserved Components models filter in the Eurostatistics publication. We are accompanying these estimates with appropriate meta-information describing the characteristics of alternative procedures. Furthermore, we are investigating several multivariate detrending techniques based on structural VAR and multivariate unobserved component models which are presented in Mazzi, Mitchell and Moauro (2008) and in Lemoine, Mazzi, Monperrus-Veroni and Reynes (forthcoming).

[FIGURE 4 OMITTED]

Figure 4 shows the latest EA GDP trend-cycle estimates from 1995Q1 to 2008Q4 using the Hodrick-Prescott filter, obtained in May 2009. As is shown clearly, we are in a negative phase of the growth cycle, which confirms the results from the growth cycle coincident indicator shown in figure 2. Due to the endpoint estimation problems characterising detrending filters, some inconsistencies between the signals delivered by the GCCI and the growth cycle estimates cannot be excluded. Users should be advised of that possibility and the producers of business cycle indicators should appreciate the unreliability of signals coming from different methods.

5. Conclusions

The paper has synthetically presented several current Eurostat projects aimed at building up a system of rapid estimates giving a clear picture of the short-term economic situation at the Euro Area level. The results presented here are preliminary and further investigation is needed before taking a final decision on the communication strategy for this kind of information. Nevertheless some results are very encouraging and the approaches proposed are considered to be methodologically sound, easily understood, as well as replicable and suitable for communication in a clear and transparent way. In order to improve the overall quality of the estimates presented in this paper, we are working on the following lines: 1) incorporating as much national information as possible into Euro Area models, 2) investigating, especially in the field of flash estimates, the possibility of constructing estimates using an indirect approach, working at Member States level rather than at the Euro Area level, 3) investigating the usefulness of introducing additional data sources into our models, especially in the case of Euro-MIND, to increase the reliability of some components estimates, 4) analysing more sophisticated data and model selection techniques, 5) testing alternative specifications for our turning points coincident indicator to reduce their lagging characteristics especially for the BCCI, 6) constructing a chronology and a turning point indicator for the acceleration cycle, 7) studying alternative solutions, recently proposed in the literature, to increase the reliability of endpoint estimates of detrending filters. Finally, it is important to note that there are several synergies among ongoing projects which still need further investigation. For example, the coincident indicators of GDP growth could be used to improve the performance of the EuroMIND for the current quarter and in principle EuroMIND itself can replace the Industrial Production Index in the specification of both GCCI and BCCI.

REFERENCES

Anas, J., Billio, M., Ferrara, L. and Mazzi, G.-L. (2008), 'A system for dating and detecting turning points in the Euro Area', Manchester school, 76(5), pp. 549-77.

Frale, C., Marcellino, M., Mazzi, G.L. and Proietti, T. (2008), 'A monthly indicator of the Euro Area GDP', CEPR Discussion Paper 7007, www.cepr.org/pubs/dps/DP7007.asp.asp.asp.

Mazzi, G.-L., Mitchell, J. and Moauro, F. (2008), 'Structural VAR based estimates of the Euro Area output gap: theoretical considerations and empirical evidence', paper presented at the 28th Symposium of Forecasters, Nice.

Lemoine, M., Mazzi, G.-L., Monperrus-Veroni, P. and Reynes, F. (forthcoming),'Real time estimation of potential output and output gap for the euro-area: comparing production function with unobserved components and SVAR approaches', Journal of Forecasting.

Gian Luigi Mazzi and Gaetana Montana *

* Europe Commission Eurostat. e-mail: Gianluigi.Mazzi@ec.europa.eu; Gaetana.Montana@ec.europa.u.

Keywords: Coincident indicators; cyclical estimates; flash estimates; monthly GDP; turning points detection JEL Classifications: C53; E32; E37

I. Introduction

Regular monitoring of the economic situation requires, especially under uncertain conditions, a complete, timely and reliable set of statistics and statistical indicators giving a clear picture of economic movements.

The Euro Area system of short-term statistics was built up in the 1990s with the adoption of a set of new legal acts. Since the beginning of the new century, when a set of Principal European Economic Indicators (PEEls) were defined, Eurostat has launched an improvement strategy focusing mainly on the reduction of publication delays and on the progressive harmonisation of data production processes at national and European level.

In addition, action has been taken to fill the availability gap, particularly in the areas of services and labour-market statistics. Nowadays, almost all PEEls are available and tangible results have been achieved concerning timeliness, especially due to the use of flash estimation techniques (for GDP and HICP) or to earlier estimates based on an incomplete set of information (for Retail Trade).

Nevertheless the situation is not yet satisfactory from the user's point of view: 1) further improvements to timeliness are required without a significant decrease in accuracy; 2) the length of Euro Area series is not as long as needed for economic modelling and business cycle analysis purposes; 3) the frequency at which data are available is, in some cases, not appropriate to ensure a regular follow-up of the cyclical situation and an effective implementation of monetary policy actions. This last drawback is particularly relevant and requires some additional thought. In the past, the Industrial Production Index was considered to be the reference variable for short-term analyses. With the continuous growth of the services sector and the associated decrease of the industrial sector, the role of the Industrial Production Index has also been substantially re-thought. In addition, the empirical evidence of cyclical movements in the services sector has convinced business cycle analysts of the need to find a new and more comprehensive measure of economic activity. This belief has been reinforced by the fact that in the recent years, some industrial cycles have not generated fluctuations for the whole economy. Finally, the high degree of volatility of the Industrial Production Index, which prevents a clear identification of signals, has also played a role in diminishing the relevance of this variable. For these reasons, ongoing studies aim to produce a more significant monthly indicator for business cycle purposes.

Another issue is that statistics themselves cannot answer the needs of all users because data give a general measure of a phenomenon. Analysts often want to extract specific signals from data in order to emphasise specific cyclical features. Gaps in statistics can be filled by specific indicators to complete the information set to be put at the disposal of analysts.

Thus, statisticians are required, starting with official statistics, to: a) extract from statistics some special signals to help improve our understanding of the economic situation; b) combine official statistics to derive new indicators filling specific gaps in data availability.

Improvements in statistics using statistical and econometric techniques, as well as compiling new indicators using existing statistics, are not traditionally considered part of official statistics. In fact, in recent years, more and more statistical agencies have been involved in such activities. It is also clear that statisticians, having comprehensive knowledge of data and of production systems, are in a privileged position to produce reliable and statistically sound estimates and indicators. Users could benefit from the availability of such estimates and indicators, using them in their analyses as well as in the decision-making process.

This paper briefly presents some current activities going on in Eurostat related to the improvements of PEEIs and to the construction of new indicators in order to help analysts and policymakers interpret correctly the economic situation.

In section 2 we discuss alternative ways to increase data timeliness and we present some preliminary results. In section 3 we describe our approach for the construction of a Euro Area monthly indicator of economic activity, which can be viewed as a very good proxy of GDP at the monthly frequency. Section 4 presents the statistical framework for business cycle analysis developed at

Eurostat. This framework consists mainly of two parts: a turning points dating and detection system and a set of growth cycle estimates. The most recent results are also presented. Section 5 presents some conclusions and lines of enquiry for further investigation and improvement.

2. Increasing data timeliness

Users, particularly institutional ones, consider the timeliness of statistics as one of their most important characteristics. In theory they should welcome real-time or quasi real-time information on main economic indicators. However, data producers are often reluctant to anticipate the release date because of the loss of precision that such anticipation could afford.

The trade-off between quality and timeliness is a key consideration when deciding on the timing of macroeconomic statistics releases. For the Euro Area the timeliness is not only relevant in absolute terms but also in comparative terms with respect to US release dates. Despite significant improvements in Euro Area release dates observed over the past 5-6 years, the gap between Euro Area and US releases is still very relevant. The only exception to this is in the case of Consumer Prices where we anticipate the US release by means of our flash estimates.

An increase in timeliness can be achieved using various strategies:

1) speeding up of the data production process, using advanced survey techniques, simplied questionnaires, use of administrative information;

2) the use of statistical and econometric techniques to extrapolate indicators by using an incomplete information set;

3) the use of so called EU sampling techniques, which consist in the construction of European samples which do not necessarily generate nationally significant ones;

4) the construction of coincident indicators for the main economic variables, which give an idea of their latest trends.

The first strategy is clearly a structural one which can be feasible in the medium-to-long term and which can be considered, from the producers' point of view, to be the most suitable approach to this issue.

The third strategy implies defining a European sampling scheme applicable to the most relevant surveys on which key economic variables are based (e.g. the industrial production index, quarterly level cross surveys, etc). This solution might be difficult when applied to National Accounts which are mainly a synthesis of various macroeconomic information. Until now, Eurostat has obtained interesting results based, at least partially, on this approach in the field of Retail Trade Turnover. In this paper we concentrate mainly on current studies related to the second and fourth strategies.

2. I Use of statistical and econometric techniques

The use of statistical and econometric techniques can contribute significantly to the increase of timeliness in the short and medium term. In this context, the key tool is represented by a set of forecasting techniques adapted to estimate the recent past or the present instead of the near future. Since 2006, we have been working on a simulation study aimed at producing flash estimates for the Producer Price Index at t+16, Employment at t+30 and, more recently, GDP at t+0, t+15 and t+30. The flash estimates, which we are working on, are based on the following main principles agreed with Eurostat Production Units:

* whenever partial information on the target variable, either at geographic or sectoral level is available, this has to be included in the flash estimation model;

* soft data (e.g. business and consumer surveys) may be integrated into the model on condition that a minimum amount of hard data is available;

* in order to increase forecasting accuracy, statistically related indicators {e.g. conventional earnings in the case of flash estimates of the Labour Cost Index) may be used in the model either in the case of the unavailability of significant partial information on the target variable or to complement such partial information;

* any hypothesis based on economic theories has to be avoided in the model specifications;

* purely univariate models should not be taken into account;

* the selected model should be as simple as possible, statistically sound, easy to use in the regular production process and characterised by a very simple dynamic specification where needed.

In the simulation exercise we pay particular attention to the identification of the most appropriate variable selection strategies and to the comparison of alternative model specifications to achieve the best performance of our estimates. Forecast combination techniques are also tested as a means of increasing the reliability of estimates.

Table 1 shows the real-time simulation for the period April 2007-March 2009 of the flash estimates for the Producer Price Index against the Eurostat first and final estimates for the Euro Area. The selected model is based on the German Producer Price Index and on oil price data according to the BIC selection criterion. Results appear quite encouraging; despite the presence of two cases of sign discordance, often the flash estimates are quite close to Eurostat first estimates. The model reacts fairly well to the change of regime even if a little slowly. Our opinion is that this constitutes a very good starting point, which could be easily improved either by adding more national information or by including Euro Area Import Prices which were not available at the start of the simulation period.

2. 2 Coincident indicators

Coincident indicators aim to forecast the evolution of economic variables during the reference period or just after it. For this purpose, they are based on the same principles as leading indicators. The main advantage of coincident indicators is that they are subject to fewer constraints than flash estimates (see section 2.1 above), even if, once again, the use/imposition of economic relationships is not recommended. In recent years, we have investigated alternative model specifications for GDP, IPI and more recently Employment. The results for the IPI are not very satisfactory mainly because of the high degree of volatility of this indicator. In this paper we briefly present our approach to constructing a coincident indicator of GDP. We have compared four different models: 1) a first bridge model containing only Business and Consumer Surveys data and financial variables, 2) a second bridge model which differs from the first by the inclusion of the IPI, 3) a bridge model based only on financial variables 4) a factor based model, where data are selected by the LARS (Least Angle Regression) algorithm. In the real-time simulation we have used, for the first part, an estimate based on combining forecasting techniques with equal weights of the three bridge models and, for the latest part, the factor model which seems to perform slightly better than the others.

For each quarter we produce three estimates of GDP: the first at the end of the second month (t-30), the second at the end of the quarter (t+0) and the last at the end of the first month of the current quarter (t+30).

Table 2 shows the real-time results obtained by using the factor model for data selection over the period 2007Q1-2009Q1, complemented by the LARS algorithm from 2009. We compare the three estimates of the coincident indicator to the Eurostat flash estimate. The results appear very encouraging. We observe only one case of sign discordance in 2008Q2, where the growth is around zero. Furthermore, the coincident indicator anticipates GDP growth correctly in most cases. In the fourth quarter of 2008, we note an unusually sized error between the estimates of the coincident indicator and the Eurostat flash one. Obviously, the simulation period presented in the table is quite short and further simulations are required, but the coincident indicator presented in this section appears to be robust and reliable enough.

3. Monthly indicator of economic activity

GDP is obviously the ideal candidate as the reference variable for short-term and business-cycle analysis but, unfortunately, it is only available on a quarterly basis and the production of monthly GDP, completely based on National Accounts standards, still appears infeasible. For this reason, several studies have been conducted recently investigating alternative ways to construct monthly proxies of GDP. Examples of such indicators are available in Sweden, Finland, Estonia and the UK, as well as in Canada. From 2006 onwards, we have investigated the possibility of constructing a Euro Area monthly indicator of economic activity as much as possible consistent with GDP.

3.1 Euro-MIND: a Euro Area monthly indicator of economic activity

The methodology proposed here is presented in more detail in Frale, Marcellino, Mazzi and Proietti (2008). It may be described synthetically as follows:

1) We base the construction of the monthly indicator of economic activity on a disaggregate approach represented by the output and expenditure breakdowns of GDP at the quarterly base.

2) For each disaggregate GDP component, a set of monthly indicators is carefully selected, including both macroeconomic variables and survey answers.

3) The indicator is based on information at both the monthly and quarterly level, rather than the monthly only, modelled with a dynamic factor specification cast in state-space form.

4) The state space methodology has the flexibility of handling data with different frequencies of observation. This is achieved by suitably defining the states of the system so as to convert temporal aggregation into a systematic sampling problem.

5) Since estimation of the multivariate dynamic factor model can be numerically complex, computational efficiency is achieved by implementing univariate filtering and smoothing procedures.

6) Special attention is paid to chain-linking and its implications for the construction of a monthly indicator of economic activity, via a multistep procedure that exploits the additivity of the volume measures expressed at previous year prices.

7) The estimate of the Euro Area monthly indicator of economic activity is obtained by combining the estimates from the output and expenditure sides, with optimal weights reflecting their relative precision.

8) The resulting pooled estimator is more precise than each of its two components, paralleling the results on the usefulness of pooling in the forecasting literature. The resulting estimates are benchmarked to quarterly national accounts produced by Eurostat so that the full consistency between monthly and quarterly estimates is achieved.

9) We provide an explicit measure of uncertainty around the indicator, which is particularly relevant in a decision-making context and for evaluation purposes.

[FIGURE 1 OMITTED]

A generalisation of this model using two factors, where the second one contains business and consumer surveys data is presented in Frale, Marcellino, Mazzi and Proietti (forthcoming). This extension increases the forecasting ability of the model at one-two-three steps ahead. A real-time simulation of the one factor-based model has been carried out since 2006 with very encouraging results. In this simulation we are producing estimates at t+45 each month, so that, in month t, we produce the estimate for the month t-2. At the same point in time, estimates for months t-1, t, t+1 can be obtained by using the two-factor version of the model.

Figure 1 presents the growth rate of Euro-MIND from January 2005 to February 2009 as estimated in April 2009, together with confidence intervals at 95 per cent. Looking at the graph, it is important to note that the evolution of the indicator is quite regular and it follows cyclical movements very well. The estimates appear very stable and not volatile, which is also confirmed by analysing subsequent vintages for the same period. The main point on which the indicator still needs some improvement is represented by the way it estimates the months of the current quarter, especially in the recession phase. Our indicator delivers negative growth rates (e.g. in January and February 2009), which appear too optimistic in comparison with the expected results. A more accurate specification of the model for the financial services sector and for the demand side component will probably improve the ability of the mode[ to estimate the most recent months.

4. A statistical framework for business cycle analysis

The set of macroeconomic statistics regularly compiled by a statistical office represents a very useful instrument available to all users and analysts. Nevertheless, we have to recognise that not all the information needed by analysts is explicitly available from an investigation of statistics. Some signals need to be extracted in order to have a clearer picture of the cyclical evolution of the economy, complementing the information supplied by statistics. In this context, we have decided to launch several activities aiming at defining a coherent statistical framework for business cycle analysis. They include the construction of statistical turning point chronologies, the development of turning point coincident indicators and estimates of the growth cycle (i.e. the output gap in the case of GDP), which can support economic monitoring and decision-making processes.

4.1 Euro Area turning point coincident indicators The methodology for the construction of a Euro Area turning point chronology and a system of coincident turning point indicators is presented in Anas, Billio, Ferrara and Mazzi (2008). The methodology can be synthetically expressed as follows:

1) simultaneous analysis of the classical business cycle and the growth cycle in the so-called ABCD framework;

2) statistical dating of Euro Area turning points by means of a simple non-parametric dating rule;

3) comparison of Euro Area and Member States dating to achieve a final statistical chronology ensuring the maximum degree of consistency between the two approaches;

4) preliminary investigation of alternative models for the construction of coincident turning point indicators for the classical business cycle and the growth cycle, including the identification of appropriate number of regimes and thresholds;

5) variables selection for the growth cycle coincident indicators (Employment expectations, Construction confidence indicator, Financial situation of the last 12 months, IPI, Imports of intermediate goods);

6) construction of the growth cycle coincident indicators (GCCI) as a weighted mean of the transition probability returned by five univariate two regime Markov Switching models fitted on each variable. An equal averaging weighting scheme is used.

7) variable selection for the business cycle coincident indicators (IPI, New cars registration and Unemployment rate);

8) construction of the business cycle coincident indicators (BCCI) as a weighted mean of the transition probability returned by three univariate regime Markov Switching models fitted on each variable. The following weighting scheme is used, IPI = 0.34, Unemployment rate = 0.46, New car registrations = 0.20.

[FIGURE 2 OMITTED]

[FIGURE 3 OMITTED]

Filtered probabilities may be viewed as the probabilities of being in a recession phase delivered by each component of the indicators. Indicators deliver the joint recession probabities. A higher value of the indicator corresponds to a high probability of being in a recession. The threshold (set at 0.5) corresponds to a decision rule; values exceeding the threshold indicate a recession phase, values below the threshold correspond to an expansionary phase.

A real-time simulation of the two indicators against respectively the business cycle and the growth cycle chronology has been carried out to check the reliability of the models as well as to discover possible false signals. The main results are that the two indicators do not show any significant evidence of false signals and that they are slightly lagging with respect to the corresponding chronologies. Each month we produce estimates of business cycle coincident indicators and growth cycle coincident indicators for the month t-2, based on filtered probabilities. Estimates for the month t-1 and t are based on forecast probabilities. Figures 2 and 3 show the behaviour of the two indicators GCCI and BCCI as estimated in April 2009. In both graphs the black bold line is the constant threshold equal to 0.5. When the indicators deliver values higher than 0.5 we are respectively in a growth cycle or business cycle recession phase. On the other hand, when the indicators deliver values below 0.5, we are in an expansion phase for both cycles. The black lines show the values of the two indicators obtained by averaging the filtered probabilities of the components. The red dashed line at the end corresponds to the value obtained by averaging forecasting probabilities instead of filtered ones. Looking at the indicators, the negative phase for the growth cycle started in April 2007 and still continues (see figure 2). Concerning the business cycle, the recession started in October 2008 and still persists (see figure 3). As already mentioned, both indicators appear to be slightly lagging and this is particularly true for the BCCI. In fact, we now think that the business cycle recession started in the first half of 2008. From this point of view it is obvious that BCCI still needs some improvement. Nevertheless, it has to be noted that it is preferable to have indicators detecting later turning points than ones delivering false signals or anticipating too many turning points.

4.2 Growth cycle estimates

The accurate and timely identification of turning points is a very important source of information for policymakers and analysts. Unfortunately, even the most accurate system of detecting turning points fails to supply details on the shape and the main features of the cycle. This is especially true for the growth cycle, which needs to be extracted for the seasonally adjusted version of indicators by means of &trending techniques. From the policymakers' and analysts' point of view, an accurate estimate of the growth cycle is of crucial importance, particularly for monitoring inflationary pressures and for designing a monetary policy oriented to inflation control. The main problem we have to deal with when estimating the growth cycle is its instability at the end of sample, due to data revisions on the one hand and to specific characteristics of most detrending filters on the other. During the past year we have compared alternative univariate techniques for the estimation of the growth cycle and we now regularly publish three alternative growth cycle estimates based on the Hodrick-Prescott filter, the Christiano-Fitzgerald filter and the Unobserved Components models filter in the Eurostatistics publication. We are accompanying these estimates with appropriate meta-information describing the characteristics of alternative procedures. Furthermore, we are investigating several multivariate detrending techniques based on structural VAR and multivariate unobserved component models which are presented in Mazzi, Mitchell and Moauro (2008) and in Lemoine, Mazzi, Monperrus-Veroni and Reynes (forthcoming).

[FIGURE 4 OMITTED]

Figure 4 shows the latest EA GDP trend-cycle estimates from 1995Q1 to 2008Q4 using the Hodrick-Prescott filter, obtained in May 2009. As is shown clearly, we are in a negative phase of the growth cycle, which confirms the results from the growth cycle coincident indicator shown in figure 2. Due to the endpoint estimation problems characterising detrending filters, some inconsistencies between the signals delivered by the GCCI and the growth cycle estimates cannot be excluded. Users should be advised of that possibility and the producers of business cycle indicators should appreciate the unreliability of signals coming from different methods.

5. Conclusions

The paper has synthetically presented several current Eurostat projects aimed at building up a system of rapid estimates giving a clear picture of the short-term economic situation at the Euro Area level. The results presented here are preliminary and further investigation is needed before taking a final decision on the communication strategy for this kind of information. Nevertheless some results are very encouraging and the approaches proposed are considered to be methodologically sound, easily understood, as well as replicable and suitable for communication in a clear and transparent way. In order to improve the overall quality of the estimates presented in this paper, we are working on the following lines: 1) incorporating as much national information as possible into Euro Area models, 2) investigating, especially in the field of flash estimates, the possibility of constructing estimates using an indirect approach, working at Member States level rather than at the Euro Area level, 3) investigating the usefulness of introducing additional data sources into our models, especially in the case of Euro-MIND, to increase the reliability of some components estimates, 4) analysing more sophisticated data and model selection techniques, 5) testing alternative specifications for our turning points coincident indicator to reduce their lagging characteristics especially for the BCCI, 6) constructing a chronology and a turning point indicator for the acceleration cycle, 7) studying alternative solutions, recently proposed in the literature, to increase the reliability of endpoint estimates of detrending filters. Finally, it is important to note that there are several synergies among ongoing projects which still need further investigation. For example, the coincident indicators of GDP growth could be used to improve the performance of the EuroMIND for the current quarter and in principle EuroMIND itself can replace the Industrial Production Index in the specification of both GCCI and BCCI.

REFERENCES

Anas, J., Billio, M., Ferrara, L. and Mazzi, G.-L. (2008), 'A system for dating and detecting turning points in the Euro Area', Manchester school, 76(5), pp. 549-77.

Frale, C., Marcellino, M., Mazzi, G.L. and Proietti, T. (2008), 'A monthly indicator of the Euro Area GDP', CEPR Discussion Paper 7007, www.cepr.org/pubs/dps/DP7007.asp.asp.asp.

Mazzi, G.-L., Mitchell, J. and Moauro, F. (2008), 'Structural VAR based estimates of the Euro Area output gap: theoretical considerations and empirical evidence', paper presented at the 28th Symposium of Forecasters, Nice.

Lemoine, M., Mazzi, G.-L., Monperrus-Veroni, P. and Reynes, F. (forthcoming),'Real time estimation of potential output and output gap for the euro-area: comparing production function with unobserved components and SVAR approaches', Journal of Forecasting.

Gian Luigi Mazzi and Gaetana Montana *

* Europe Commission Eurostat. e-mail: Gianluigi.Mazzi@ec.europa.eu; Gaetana.Montana@ec.europa.u.

Table 1. Euro Area flash estimates of monthly producer price inflation (t/t-I) Eurostat Eurostat Error Error Flash First Final First Final t+ 16 estimate estimate estimate estimate 2007m4 0.31 0.45 0.38 0.14 0.07 2007m5 0.28 0.29 0.38 0.01 0.10 2007m6 0.31 0.13 0.14 -0.18 -0.17 2007m7 0.25 0.26 0.30 0.01 0.05 2007m8 0.31 0.06 0.16 -0.25 -0.15 2007m9 0.24 0.36 0.41 0.12 0.17 2007m10 0.33 0.64 0.70 0.31 0.37 2007m11 0.78 0.90 0.91 0.12 0.13 2007m12 0.11 0.10 0.18 -0.01 0.07 2008m01 0.43 0.85 0.83 0.42 0.40 2008m02 0.74 0.66 0.57 -0.08 -0.17 2008m03 0.65 0.70 0.60 0.05 -0.05 2008m04 0.97 0.79 0.78 -0.18 -0.19 2008m05 0.93 1.21 1.19 0.28 0.26 2008m06 0.90 0.96 1.11 0.06 0.21 2008m07 1.73 1.23 1.40 -0.50 -0.33 2008m08 -0.25 -0.50 -0.39 -0.25 -0.14 2008m09 0.23 -0.20 -0.15 -0.43 -0.38 2008m10 0.01 -0.80 -0.89 -0.81 -0.90 2008m11 -1.26 -1.96 -2.07 -0.70 -0.81 2008m12 -1.19 -1.58 -1.49 -0.39 -0.30 2009m01 -0.79 -0.88 -1.13 -0.09 -0.34 2009m02 -0.57 -0.43 -0.43 0.14 0.14 2009m03 -0.69 -0.70 -0.70 -0.01 -0.01 Table 2. Coincident indicator of quarterly GDP growth, per cent Estimates Estimates Estimates Eurostat t-30 t+0 t+30 Flash t+45 2007Q1 0.5 0.5 0.5 0.57 2007Q2 0.4 0.4 0.4 0.34 2007Q3 0.7 0.7 0.9 0.71 2007Q4 0.3 0.3 0.4 0.41 2008Q1 0.25 0.3 0.45 0.8 2008Q2 0.2 0.2 0.0 -0.2 2008Q3 -0.3 -0.3 -0.2 -0.19 2008Q4 -0.5 -0.8 -1.0 -1.5 2009Q1 -1.0 -1.6 -1.6 -2.5

Printer friendly Cite/link Email Feedback | |

Author: | Mazzi, Gian Luigi; Montana, Gaetana |
---|---|

Publication: | National Institute Economic Review |

Geographic Code: | 4E |

Date: | Oct 1, 2009 |

Words: | 4888 |

Previous Article: | Confidence and leading indicators: introduction. |

Next Article: | Nowcasting is not just contemporaneous forecasting. |

Topics: |