Overview of GDP flash estimation methods
Overview of GDP flash estimation methods
(preliminary draft)
6May2015
EUROSTAT task force ‘GDP Flash at T+30 days’
Contribution of the working group Methods and estimation technique
Contributors[1]:
Filippo Moauro (ISTAT, Italy with the role of supervisor), Claudia Cicconi (ISTAT, Italy) SamuHakala(Statistics Finland), Kristina Kiriliauskaite (Statistics Lithuania), Stefan Leist and KornelMahlstein(State Secretariat for Economic Affairs SECO, Swiss Confederation), and Michal Široký (Czech Statistical Office).
Executive summary.The system of European GDP flash estimate is currently released at t+45 days under the coordination of Eurostat. In order to investigate the possibilities to reduce its delay at t+30 days and share the experiences among member states, Eurostat launched a task force at the end of 2012. The works have involved 16 member states plus Switzerland in 2013 and 2014. A particular focus has been devoted to estimation methods and this document represent the main result of a dedicated sub-group of the task force. The document provides a general overview of estimation methods in use among member states also, accompanied by guidelines on best practices. After the introductory chapter, the document discusses preliminary analysis of available data, model strategies, methods for forecasting missing data including the extensions to multivariate setups and the analysis of results.
Two main estimation strategies have been identified: according to the ‘direct’ approach last quarter of a given GDP component is directly estimated, whereas the ‘indirect’ method requires extrapolating the last quarter by means of a temporal disaggregation procedure. The difference is given by the model used for extrapolation: under the former approach the set of related data is directly used for extrapolating the last quarter of the GDP component; under the latter extrapolation is carried out in two steps: first extrapolating the indicator in the last quarter and then running the temporal disaggregation for the GDP component of interest.Following the same logic, when no indicator is available a one- or two-steps time series method applies, respectively,for the direct or the indirect approach.
A summary view of the answers to a questionnaire on estimation methods are also shortly presented in an annex of this document. The questionnaire has helped at understanding the experience of member states, their approach in terms of accounting method, detail of compilation, price evaluation, type of adjustment on data, if a preliminary analysis is effectively carried out, more used indicators and, of course, more technical aspects related to methods in use.
Therefore the document is an attempt to merge the experiences among member states in a unified view and allows the reader to follow a step by step approach to flash estimate, with explanations and analysis to efficiently carry out the exercise. The approach is pragmatic, with link to available software, references to the literature and a particular stress on the elements that are common between advanced users and beginners coming or not from the same statistical agency. The scope is to facilitate the communication of technical aspects concerning estimation methods in the community of people working in the area of quarterly accounts. Therefore the best practices presented here are a way to harmonize the process of flash estimation at national level and, indirectly, to contribute at increasing the quality of European aggregates released at short delay.
Index
- Introduction
- Background
- Purpose
- Preliminary analysis
2.1.Objectives
2.2.Analysis of available data
2.3.Detail of compilation
2.4.Graphical analysis
2.5.Outliers’ detection and their removal
- Forecasting missing data
- Introduction
- Model strategies
- Direct approach
- ADL models
- Model selection
- Estimation and diagnostics
- Available software
- Dynamic factor models
- Model selection
- Estimation and diagnostics
- Available software
- Indirect approach
- Model selection
- Estimation and diagnostics
- Available software
- Pure forecasting methods
- ARIMA models
- Model selection, estimation, diagnostics and the forecasting stage
- Available software
- Structural time series models
- 3.5.2.1.Model selection, estimation, diagnostics and the forecasting stage
- Available software
- Multivariate extensions
- Introduction
- VAR modeling approach
- Modeling and statistical treatment
- Available software
- Multivariate STS models
- Modeling and statistical treatment
- Available software
- Analysis of results
- Comparative analysis of results
- Measures of dispersion
- Summary error statistics
- Forecast encompassing
- Forecast combination
References
Annex –Questionnaire on estimation methods
- Introduction
With the intention to assess the feasibility to produce a flash estimate of the euro area and the European Union quarterly GDP at T+30 days, Eurostat settled down in 2013 a task force involving a wide number of member states. Main purpose is to share the practice in this area and the need to provide the group with a set of guidelines to assist both experts and beginners in the domain of flash estimate of quarterly national accounts.
From a technical standpoint estimation of quarterly GDP at T+30 days poses a problem of missing data for the last observation, as well as the need to provide modeling solutions to efficiently combine scarce available information with forecasts. More common situations are either the lack of last- or last two months of the quarter Twhen monthly short term business statisticsare available, or one of observations completely missing in last quarter.
This document introduces the technical framework for flash estimation of GDP by means of a pragmatic approach. The focus is on more used methods by member states participating the works of the task force, providing a summary overview but without standing on their relative merits.The overview do not limit to forecasting methods but also covers other important practical aspects like preliminary analysis of available data and comparative analysis of results relative to alternative methods. For this reason the document tracks also a strategy for flash estimating GDP, taking into account the debate on technical aspects of estimation within the task force, as well as the answers from a specific survey conducted on member states.
Each paragraph presenting a forecasting approach also describes 3 sets of alternatives in a logic already experimented with success by Eurostat in guidelines and handbookscovering other statistical domains. The first alternative, denoted as A) represents the best approach to be aimed at; the alternative (B) is for acceptable methods,representing a good alternative in presence of certain characteristic of available data or constraints in dedicated resources; the last alternative (C) shows approaches that should be avoided.When possible the methods presented here are complemented by a list of references and information about available software.
Therefore thegoal of the document is to provide colleagues involved in flash estimates with a summary description of the methods used across member states and the instruments to converge their practices towards acceptable (B methods) or even best practices (A methods). The document also allows the reader to follow a step by step approach to flash estimate, with explanations and analysis to efficiently carry out the exercise.
This guidelinesdo not pretend to be exhaustive and the authors acknowledge that the literature on forecasting is vast and subject to continuous development. Therefore this material is a first collection of suggestions and reflections on flash estimates coming from the data production domain, with the certainty that the work is subject to a continuous improvement.
The document is organized as follow: section 1 presents background information on GDP flash estimate at T+30 days (section 1.1) and main purposes of the document (section 1.2). Section 2 is devoted to preliminary analysis, defining objectives (section 2.1), strategies for available data (section 2.1), detail of compilation (section 2.3), graphical analysis (section 2.4) and outliers’ detection and their removal (section 2.5). Forecasting missing data is the subject of section 3: after an introduction (section 3.1)and a track of model strategies (section 3.2), we present more used methods split in direct (section 3.3) and indirect approaches (section 3.4), pure forecasting methods (section 3.5) and multivariate extensions (section 3.6).Section 4 is on the analysis of results, presenting more common measures for a comparative analysis of results (section 4.1) and a short discussion on both forecast encompassing (section 4.2) and forecast combination (section 4.3). Finally an annex presents the questionnaire on estimation methods with a summary view of received answers from member states participating the task force and a short discussion on main evidences.
1.1.Background
Quarterly GDP data are probably the most relevant economic short term statistics produced by the European statistical system. Their release is coordinated by Eurostat and from September 2014, in occasion of entering in force of the new ESA 2010 data transmission program, it has become effective a release calendar at 45, 60 and 90 days after the end of the reference quarter for data dissemination.
In recent discussions at national and international level involving statistical institutes, users and practitioners emerged a clear need to anticipate the GDP flash from a delay of 45 to 30 days. Among European countriesUnited Kingdom, Belgium, Lithuania and Spain are already publishing a GDP flash estimate at t+30 days and other countries are testing, or have tested in the recent past, the feasibility to reduce this delay.
There are several reasonsfor the alignment of GDP flash estimate to a deadline at t+30 days: timely information is requested by policy makers and stakeholders and the need of an efficient early warning system have been exacerbated by the recent economic crisis; moreover a flash estimate at T+30 days would align the European calendar to that of US, where GDP is released according to a 30-60-90 days timetable.
Therefore the idea by Eurostat to launch in 2013 a task force on the feasibility to produce a flash estimate of the euro area and European Union quarterly GDP at t+30 days.
From the technical standpoint some objectives were immediately defined at the beginning of the works of the task force:
- the interest of the estimates limited to seasonal adjusted quarterly growth rates of volume measures;
- the methodological framework under which to move for estimation, following source and methods sketched in a series of handbooks released by Eurostat, from the ESA 2010 manual, to those on price and volume figures, until the recent edition of the handbook on quarterly rational accounts;
- the limits of these references considering that GDP flash at t+30 days poses both a problem of producing data under time restrictions and a problem of missing data in more recent periods;
- the Eurostat’s purpose to compile euro area and European Union GDP flash following the so called ‘direct method’, i.e. using member states estimations and imputing missing countries in a second step;
- and finally to establish a deadline in 2016q1 for a first t+30 flash estimate for the euro area and European Union quarterly GDP in case of positive response by the task force on feasibility of estimates.
Moreover it is clear by all institutions involved in the project that if the final target is European aggregates, their compilation require data at member state level. Therefore a decision by Eurostat to move GDP flash estimate to t+30 days could represent a stimulus for member states of aligning their corresponding releases at the same deadline.
Statistical agencies, central banks and other stakeholders around the world are day by day engaged in forecasting economic indicators and, of course, real GDP is a variable of primary interest. Therefore modeling solutions and tools to efficiently combine scarce available information with forecasts is matter of perpetual debate.
1.2.Purpose
This document aims at providing help to national accountants involved in flash estimates on methods and technics for a correct practice. The stress is on the elements that are common between advanced users and beginners coming or not from the same statistical agency and that facilitate the communication in the community of people working in the area of quarterly accounts. Therefore the best practices presented here are a way to harmonize the process of flash estimation at national level and, indirectly, to contribute to produce robust European aggregates.
The document also provides a design for flash estimation, from preliminary analysis of available data, to outlier detection and correction, model strategy and analysis of results.
The adoption of guidelines by statistical agencies is an element of transparency for data users who are strongly interested to know details of flash estimation concerning the production process, full model specification, reliability of estimates and elements for a comparative analysis of results.
- Preliminary analysis
2.1.Objective
Description:
Flash estimate of quarterly GDP components concerns estimation of recent quarters when the most suitable set of data for its computation is not available for reason of time.
The main objective of the preliminary analysis is to collect and analyze all data available at short delay and which could help and give robustness to the estimation of last quarters of GDP. These indicators should be selected among data measuring similar phenomenon of the GDP component. Related data should also show a similar pathto the levels of the GDP component or to the series after transformation (e.g. logarithmic transformation, difference operators, moving averages).
Available indicators should be collected taking into account the split at which quarterly accounts are currently compiled. If seasonal adjusted indicators are not available, seasonal adjustment should be applied to this set of data. Comparison at aggregated level should also be considered when the comparison between GDP components and related series is poor at detailed level.
Options:
-Running a detailed preliminary analysis
-Focus the analysis on currently used available indicators
-Not to do a preliminary analysis
Alternatives[2]:
A)Detailed preliminary analysis for all components
B)Analysis of indicators reduced to most relevant components
C)No preliminary analysis
2.2.Analysis of available data
Description:
Analysis of available data starts from the structure of GDP components provided by quarterly national accounts for volume measures. The purpose is to find a relationship of this data with relevant indicators available in time and able to provide useful information for estimation of GDP in the last quarter. Related indicators could be official short term statistics, other relevant economic data, business and consumer surveys. If indicators are not available with the same detail of GDP components the analysis could be performed at aggregated level (e.g. broad or aggregated NACE). Available indicators could be quarterly and/or monthly data. Among monthly data it should be considered all the indicators able to provide information at least on one month of the last quarter. Analysis of available data is limited to volume indicators. For data expression of nominal measures (like turnover indexes and retail sales) a suitable deflation should be preliminary performed.
Options:
-Detailed search of available indicators among a reasonable wide spectrum of data sources
-Limit the analysis to official statistics and/or to more relevant GDP components
-Limit the analysis to already used indicators for most relevant GDP components
Alternatives[3]:
A)Detailed search and analysis of related indicators at level of the available split of GDP by sub-component
B)Analysis of indicators reduced to official statistics and most relevant components
C)No search and analysis of alternative indicators
2.3.Detail of compilation
Description:
Flash estimate of GDP results from a multi-step exercise in which elementary estimates relative to sub-components are aggregated up following a pre-defined scheme. The elementary components could reflect the maximum detail of quarterly national accounts or a reasonable simplified split due to lack of available data. It could be implemented to sub-components from the production side, to those from the demand side or it can follow both the approaches. In this latter case the final GDP estimate could be benchmarked to ensure consistency among approaches.
In presence of timely indicators, the preliminary analysis could start from the split of GDP available in the current production of quarterly national accounts. In the following steps the analysis could develop towards a lower detail if there is evidence that indicators fit better at aggregated level.
Options:
-In presence of indicators from both the production and the demand side, compile the exercise according to both the approaches.
-At a preliminary stage limit the compilation only to one approach even if data from the other side are available.
-In presence of a detailed set of information, foresee a detailed compilation scheme. Such a detail should ensure the best use of the information given by the composition of quarterly national accounts. Complement the compilation with higher levels of aggregation (e.g. the split of value added into the 3 or 10 categories of the NACE classification). Two or more levels of aggregation allow a wider comparative analysis of results.
-Limit the compilation to a reasonable aggregation scheme including most relevant aggregates and taking into account the available split of indicators.
-Limit the compilation to macro levels of aggregation
Alternatives[4]:
A)Detailed compilation for components by the supply and the demand side. Complement the compilation with an analysis carried out at two or more levels of aggregation
B)Compile the exercise only from one side with e reasonable split into sub-components and at least two levels of aggregation
C)Limit the compilation to standard macro components.
2.4.Graphical analysis
Description:
A first graphical comparison in the time domain provides the analyst with useful information on the utility of available indicators in forecasting GDP components. Graphs should look at data in levelsor logarithms (better after standardization), first or seasonal differences or growth rates. The analysis could be also complemented by a very first run of the forecasting software using default methods; a first analysis of the residuals from this fit allows to form a preliminary idea of most suitable model specifications.
The analyst could collect information on:
- The structure of the GDP component in terms of trend-cycle-seasonal components;
- the presence of seasonality in the related indicator;
- the presence of outliers in the indicators;
- the fit of related series to the GDP component;
Even if time consuming especially in presence of a detailed breakdown of GDP into components by the demand and/or the supply side,a preliminary graphical analysis is relevant at least for most relevant series.