ESS Guidelines on Temporal Disaggregation, Benchmarking and Reconciliation

ESS guidelines on temporal disaggregation, benchmarking and reconciliation

From annual to quarterly to monthly data

Version 29, June 2017

ESS guidelines on temporal disaggregation: benchmarking and reconciliation

Forewords

Temporal Disaggregation methods are used to disaggregate low frequency (LF) time series to higher frequency (HF) series, where some temporal constraints (such as either the sum, or the average, or the first or the last value of the resulting HF series) are consistent with the LF series and or some accounting constraints must be satisfied by the temporal disaggregated series. In literature, the Temporal Disaggregation procedures closely related to the temporal distribution (or interpolation), which occurs when annual data are available and sub-annual data are required, are known as Benchmarking procedures, while the term Reconciliation generally indicates methods applied to satisfy the accounting (or cross sectional or contemporaneous) constraints. In particular, Benchmarking are defined as methods that are used to derive high frequency series consistent with their corresponding low frequency benchmarks and, at the same time, preserve the short-term movements of high frequency economic indicators. Reconciliation refers to methods that may be necessary for deriving HF series and that comply with both low and high frequency aggregation constraints, generally accounting constraints. Temporal Disaggregation can be performed with or without one or more high frequency indicator series.

Standard methods for Temporal Disaggregation are: Denton ( 1971), Denton-Cholette, Dagum and Cholette (1994) Chow-Lin (1971) and its variants Fernandez (1981) and Litterman (1983). Among these, Denton and Dagum-Cholette and its variants (see. Dagum and Cholette, 2006) are primarily concerned with movement preservation, generating a series that is similar to the indicator, whether or not the indicator is correlated with the low frequency series. That is, considering the most common case from Annual to Quarterly, the benchmarked series should maximally preserve the quarterly growth rates of the indicators given the annual benchmark constraints. When there isn’t a related indicator, or when the indicator has a poor quality, some mathematical techniques to distribute LF data in HF series are recommended. A standard method for Temporal Disaggregation without an indicator is the Boot, Feibes and Lisman (1967).

The establishment of common Guidelines for Temporal Disaggregation within the European Statistical System (ESS) is an essential step towards a better harmonisation and comparability of infra-annual statistics, especially Principal European Economic Indicators (PEEIs). The ESS Guidelines on Temporal Disaggregation address the need for harmonisation expressed by many users such as European Commission services: The definition of best practices in the field of Temporal Disaggregation has been postponed for a long time as it was necessary that they reached at the end the harmonization procedures of the variable on which Temporal Disaggregation are mainly used.

The ESS Guidelines on Temporal Disaggregation present both theoretical aspects and practical implementation issues in a friendly and easy to read framework. They meet the requirement of principle 7 (Sound Methodology) of the European Statistics Code of Practice and their implementation will also be in line with principles 14 (Coherence and Comparability) and 15 (Accessibility and Clarity). The Guidelines also foster the transparency of Temporal Disaggregation practices by encouraging documentation and dissemination of practices

The Guidelines includes sections with a policy for seasonal adjustment, quarterly national account, labour market and revisions consistent with the Guidelines on these topics. The specification of alternatives take into account the operational issues.

Acknowledgments

The European Commission expresses its gratitude and appreciation for the work carried out by the authors of this Guidelines: XXXXXXX

The European Commission would like to thank the chairs of the ESS Working Group Methodology an and the Working Group National Accounts, XXXX, together with all members for their useful comments and support.

The European Commission would also like to thank all the members of the Task Force on Temporal Disaggregation, from Eurostat, from National Statistical Institutes, from the ECB and from National Central Banks who have contributed with their comments to improve the Guidelines.

ESS guidelines on temporal disaggregation: benchmarking and reconciliation

Introduction

Table of contents

Forewords 2

Acknowledgments 3

1. Introduction 6

1.1 Motivation for Guidelines 6

1.2 Scope of Guidelines 6

1.3 Costs and risks 6

1.4 Background to Guidelines and basic definitions 7

1.5 Methods 7

1.6 Principles for benchmarking and reconciliation 8

2. A policy for benchmarking and reconciliation 10

2.1 General Policy for Benchmarking and Reconciliation 10

2.2 The need for domain specific Benchmarking and Reconciliation policies 11

2.3 Consistency across domain specific Benchmarking and Reconciliation policies 12

2.4 Stability of Benchmarking and Reconciliation policies 13

3. General methodological aspects 14

3.1 Mathematical vs Statistical Methods 14

3.2 Target variables 16

4. Temporal disaggregation 17

4.1 Temporal disaggregation 17

4.2 Selection on one or more proxies 19

5. Benchmarking 21

5.1 Choice of benchmarking method 21

6. Reconciliation 23

6.1 Choice of reconciliation method 23

7. Specific Issues 25

7.1 Choice of the software 25

7.2 End of Series 26

7.3 Benchmarking and Reconciliation in Seasonal Adjustment 27

7.4 Benchmarking and reconciliation in Flash estimates 29

7.5 Outliers identification and treatment 30

7.6 Treatment of short series 31

7.7 Benchmarking and reconciliation in chain linked series 32

7.8 Benchmarking and reconciliation in National Account 34

7.9 Benchmarking and reconciliation in Labour Market 35

7.10 Advanced and recent method 36

8. Presentation 37

8.1 General revision policy and release calendar 37

8.2 Accuracy of Benchmarking and Reconciliation 38

8.3 Metadata (?) 39

References 39

1. Introduction

1.1 Motivation for Guidelines

The European Statistical System (ESS) developed these Guidelines to help data producers with deriving high frequency data (e.g. Quarterly/Monthly) from low frequency data (e.g. annual) respecting the related temporal and accounting constraints. With this aim, it identifies the best practice among Benchmarking and Reconciliation techniques in order to:

· achieve harmonisation across national processes

· enhance comparability between results

· increase the robustness of European aggregates.

The Guidelines are aimed at all ESS indicators that are politically and/or economically important, for example the Europe 2020 indicators, the Macroeconomic Imbalance Procedure indicators and the Key European Indicators. The Guidelines provide a consistent framework for temporal disaggregation, benchmarking and reconciliation, taking advantage of similarities in the process to define a common vocabulary to facilitate communication and comparison between practitioners.

1.2 Scope of Guidelines

The Guidelines cover all issues related to benchmarking and reconciliation from annual to quarterly to monthly data, from the choice of the methods, to revisions and documentation. References are provided at the end. Whether experts or non-experts, the framework for benchmarking and reconciliation remains the same, only the level of detail in the analysis varies. Each stage of the benchmarking and reconciliation process is explained, and options described. Out of these options three alternative courses of action are highlighted: (A) Best alternative (B) Acceptable (C) To be avoided.

(A) The best alternative should always be the target for producers. It can always be achieved with enough effort.

(B) The acceptable alternative should only be tolerated if time or resource issues prevent alternative (A).

The objective of the Guidelines is help producers move to alternative (A).

1.3 Costs and risks

The costs of applying alternative (A) are significant as benchmarking and reconciliation is time consuming in terms of human resources and requires a common and well defined IT structure.

The risks of not applying to alternative (A) are that inappropriate or low-quality benchmarking and reconciliation can generate misleading results, for example including over-smoothing, increasing the probability of false signals leading to misinterpretation of dynamic of the data. This will negatively affect credibility and hence ultimately trust in statistics.

1.4 Background to Guidelines and basic definitions

Temporal disaggregation is a fundamental process in data production and dissemination for coherency and the interpretation of aggregates and time series. Benchmarking are defined as methods that are used to derive high frequency series consistent with their corresponding low frequency benchmarks and, at the same time, preserve the short-term movements of high frequency economic indicators. Reconciliation refers to methods that may be necessary for deriving HF series and that comply with both annual and HF aggregation constraints, generally accounting or contemporaneous constraints.

Relevant basic definitions are the following. Balancing: adjusting preliminary values of stock or flow cross-section data to given contemporaneous constraints; Interpolation: the adjusting initial values of a high frequency stock time series to low frequency observations of higher reliability; Temporal distribution/Benchmarking: distributing initial values of a high frequency flow time series to low frequency totals (or averages) of higher reliability; Extrapolation: the estimation of more recent data when low frequency totals are not yet available; Multivariate temporal disaggregation: the contemporaneous statistical treatment of both balancing and distribution/interpolation; Indicator or proxy: a time series available at least at the same frequency as the target variable that is very closely related to the target variable (in economically/definitionally/statistically sense) and available on a regular basis, for a sufficiently long time series. Despite the conceptual differences between benchmarking, temporal distribution, interpolation and extrapolation, they are considered similar in the context of regression-based estimation.

1.5 Methods

A variety of mathematical and statistical methods have been developed and applied by researchers to solve problems of temporal disaggregation, the process of estimating unobserved sub-annual series from observed annual values.

Temporal disaggregation methods are used to disaggregate low frequency (LF) time series to higher frequency (HF) series, where some temporal constraints (such as either the sum, or the average, or the first or the last value) of the resulting HF series is consistent with the LF series and/or some accounting constraints must be satisfied by the temporal disaggregated series.

Benchmarking are defined as methods that are used to derive HF series consistent with their corresponding LF benchmarks and, at the same time, preserve the short-term movements of high frequency economic indicators. In particular Benchmarking distributes (or interpolates) LF data to construct time series of benchmarked estimates (“back series”) and extrapolates HF estimates for a period for which benchmarks are not yet available (“forward series”).

Reconciliation refers to methods may be necessary for deriving HF series that comply with both LF and HF aggregation constraints and with contemporaneous (accounting) constraints.

The target variables can be flow, stock, and index series. Typically flows are associated with the speed of a phenomenon (e.g: sales per month, births per year) then the LF values of flows correspond to the sums of HF values, (e.g. the annual values correspond to the sums of the sub-annual values). Stock variables refer to the level of a phenomenon at a very specific date, that is the annual values of stocks series (e.g. inventories) pertain to one single sub-annual value, usually the last value of the LF period. For the purpose of benchmarking and interpolation, index series are such that their annual benchmarks pertain to the annual averages of a sub-annual series. Thus, the annual values of index series correspond to the average of the sub-annual values, whether the underlying variable is a flow or a stock.

1.6 Principles for benchmarking and reconciliation

1 The goal of a Benchmarking procedure is to derive series of high-frequency data (HF) (e.g., quarterly data) with series of less frequent data (LF) (e.g., annual data) for a certain variable or a given set of variables into a consistent time series. HF estimates must be coherent with LF estimates: e.g. sum/average of quarterly estimates in a year equals the annual estimate (benchmark) and their HF dynamic must be as close as possible to that of the short-term indicator. Benchmarking via temporal distribution/interpolation of LF values (estimates) among HF periods satisfy these requirements, especially when the LF estimate is available, or via extrapolation, for most recent high frequency periods for which a LF estimate is not yet available..

2 The reconciliation procedure is required to obtain consistency in high frequency data that are subject to low frequency aggregation constraints. These can be temporal or/and accounting (contemporaneous) constraints. Temporal constraints refers to coherency between the HF values of a series and the corresponding LF values (e.g. the sum of the quarterly data in a year must be equal to the annual value), while the accounting constraints (or cross sectional constrain) refers to the coherency of sub-component to a margin total at a given time period (e.g. the sum of the regional values must be equal to the national values for a given period of time, or the sum of a values according to a classification breakdown must be equal to the general total). The main difference with benchmarking is that the reconciled estimates have to satisfy all possible level of temporal constraints. As an example, quarterly values added by the institutional sector may be required to be in line with ANA estimates by the institutional sector and independently derived quarterly values added for the total economy.

3 The target variables can be flow, stock, and index series. Typically Flows are associated with the speed of a phenomenon (e.g: sales per month, births per year) then, the LF values of flows correspond to the sums of HF values, (e.g. the annual values correspond to the sums of the sub-annual values); Stock variables, on the other hand, pertains to the level of a phenomenon at a very specific date, that is the annual values of stocks series (e.g. inventories) pertain to one single sub-annual value, usually the last value of the LF period. For the purpose of benchmarking and interpolation, index series are such that their annual benchmarks pertain to the annual averages of a sub-annual series. Thus, the annual values of index series correspond to the average of the sub-annual values, whether the underlying variable is a flow or a stock.

4 Temporal Disaggregation methods can be performed with or without related indicators. When methods with indicators are applied, the HF series is estimated on the basis of external higher frequency data linked to the relevant variable of interest. If no indicator is available the HF estimates are calculated using some purely mathematical criterion, or according to time series methods