Update to seasonal adjustment methodology for the Retail Trade Survey

This page provides information on the changes in methodology used to calculate seasonal adjustment for retail trade.

These updates are being made to better account for the impact of COVID-19, and to address changes in the seasonality of each series since the Retail Trade Survey redesign in 2017. These enhancements will take effect from the publication of Retail trade survey: December 2023 quarter on 23 February 2024.

On this page

• Overview
• Background to seasonal adjustment
• Additive outliers for seasonal adjustment and trend
• Changes to trading day effects
• Changes to specific series
• Impact on the data

Overview

We are implementing multiple enhancements to the seasonal adjustment methodology for the Retail Trade Survey.

We used these updates to:

• improve our approach to extreme outliers using the new Stats NZ standards for additive outliers in seasonal adjustment
• resume publishing trend series, which had been suppressed during the period significantly affected by COVID-19
• review our seasonal adjustment practices, now that we have more than five years of data since our 2017 redesign, including:
  • identifying series for which seasonal adjustment is no longer appropriate
  • removing trading day factors.

While seasonally adjusted data is revised with each new quarter of data, these updates will cause larger than normal revisions. Details on the impact of these changes and the reasons behind them are outlined in the following sections. We have included two files to help understand them: a CSV file with the data before and after implementation of the new methodology up to the September 2023 quarter, and an Excel workbook listing seasonal adjustment treatments before and after implementation .

This new methodology will be used from our next publication Retail trade survey: December 2023 quarter, on 23 February.

On 24 November 2023 we published a 0.0 percent change in the total retail trade sales volume for the September 2023 quarter compared with the June 2023 quarter. Using the updated seasonal adjustment methodology, the September quarter volume of total retail sales would have shown a modest fall (-0.6 percent).

This methodology change has no impact on our actual data (not seasonally adjusted), on our deflators, nor on any other Stats NZ collection. While our data is used by other collections (for example, in quarterly gross domestic product estimation), we independently calculate seasonal adjustment across collections using comparable methodology. This maintains consistency where possible while recognising the unique nature of each series.

Background to seasonal adjustment

Stats NZ performs seasonal adjustment on our time series. When COVID-19 hit in 2020, we made decisions on how to handle the impact on our time series given limited international guidance or best practice on such unprecedented extreme outliers.

There were ongoing challenges in recognising and specially treating factors that should be allowed to influence seasonal adjustment calculations from those that are extraordinary. COVID-19 lockdowns, restrictions, and the removal of these had varying impacts on all our retail series (for example, industries, regions, sales, and stocks) over time. Given those challenges we could not publish trend levels to our usual quality standards and they have been suppressed for the last few years.

Now that the COVID-19 impacts have largely diminished from our time series we are updating our approach to seasonal adjustment and introducing the new Stats NZ standard for treating extreme outliers using additive outliers. These enhancements also allow us to start publishing trend levels again.

Seasonal adjustment and automatic outliers in time series after COVID-19 has more information on the new Stats NZ approach to additive outliers.

While improving our approach to extreme outliers and reintroducing trends, we took the opportunity to review other aspects of our seasonal adjustment methodology. We last redesigned the Retail Trade Survey in 2017, and we generally need a five-year time series to assess seasonality, so now was a good time to look at a variety of enhancements to our seasonal adjustment methodology.

Additive outliers for seasonal adjustment and trend

Seasonal adjustment and automatic outliers in time series after COVID-19 describes how we previously managed extreme outliers using additive outliers at Stats NZ and the methodology changes we are now implementing for retail trade.

We have applied methodology consistent with other Stats NZ collections wherever possible. However, the retail trade industries have been thoroughly analysed to ensure the most appropriate treatment is used for each unique series. We have continued to make use of manual additive outliers during the most significantly affected COVID-19 period (March 2020 to December 2022 quarters, inclusive) for retail trade and implemented automatic detection from the March 2023 quarter onwards. Automatic outlier detection can occasionally change which quarters it assigns as an outlier when data from a new quarter is added to a time series. Manually assigning outliers during the COVID-19 period ensures the selected outliers will not change as more data is added. This reduces the magnitude of revisions to seasonally adjusted data.

Revision of the manual adjustments and use of some other complex treatments to manage impacts through COVID-19-affected periods, mean we have improved the quality of trend series to the point they can be published again.

A list of outliers and other complex treatments applied to each retail trade series before and after our updates is available from Download documents.

A number of other Stats NZ collections have already adopted our new internationally aligned standards for additive outlier detection, including the Consumers price index, Household labour force survey, Quarterly employment survey, and Value of building work put in place. All our collections will eventually align with the new standards, after we have analysed the methodology in the context of each unique series.

Changes to trading day effects

We regularly review the quality of our seasonal adjustment. While implementing enhancements to our methodology to manage extreme outliers from COVID, we took the opportunity to review other aspects of our seasonal adjustment methodology. We generally need a five-year time series to assess seasonality and our last redesign of the Retail Trade Survey was in 2017.

Trading day adjustment tries to estimate the different contribution of each day of the week to the total, based on past data. It then adjusts the value of each quarter based on the frequency of each day of the week within that period.

After analysing the quality measures since the 2017 redesign, we decided to turn off trading day adjustments for all series in retail trade, to improve the accuracy of our seasonally adjusted series and align our methodology with international best practice.

Changes to specific series

Accommodation sales series

COVID-related New Zealand border restrictions meant no seasonality could be observed in the Accommodation sales series (RTTQ.SFU1C[A,S,T] and RTTQ.SFU1K[A,S,T]) from the June 2020 quarter through to the September 2022 quarter (inclusive). We have decided to turn seasonal adjustment off for Accommodation sales series through this period. Note, data will still be published under the seasonally adjusted series for this period, but it will be identical to the actual data.

The seasonal pattern of the series returned more clearly in the December 2022 quarter, so we resumed seasonally adjusting it from then.

To ensure a smooth transition back to seasonal adjustment, we have created a synthetic series for Accommodation, covering the affected period. This synthetic series will not be published but will be used to determine the seasonal factors.

Methodology for bridging seasonally adjusted international travel and migration series impacted by COVID-19 has more details on how a synthetic series can help us calculate the seasonal factors of a series that was temporarily disrupted.

Complex treatments applied to series affected by structural changes

While reviewing our automated outlier processes during the most COVID-affected periods, we identified some structural changes in specific industries which could be better managed with the introduction of complex treatments: level shifts or temporary changes. These structural changes are not necessarily related to COVID-19 disruptions but happened during the same period. As other structural changes are observed in the future, more level shifts or temporary changes may be applied.

Seasonal adjustment and automatic outliers in time series after COVID-19 has more information on level shifts and temporary changes.

In the September 2022 quarter the Electric and electronic goods retailing sales and stocks series (RTTQ.SFL1CS, RTTQ.SFL1KS, and RTTQ.SFL9CS) were affected by a level shift, with much of this industry’s value moving to be captured under the Department stores industry instead (RTTQ.SFF1CS, RTTQ.SFF1KS, and RTTQ.SFF9CS). We decided to apply level-shift treatments to the seasonal adjustment of both industries, to account for it.

In the March 2021 quarter, the Recreational goods retailing sales series (RTTQ.SFJ1CS, RTTQ.SFJ1KS) were impacted by a level shift, as some data from the Non-store and commission based retailing industry (RTTQ.SFE1CS, RTTQ.SFE1KS) changed to be captured in it. No level shift was implemented in the Non-store and commission based retailing series, as the change was not as significant to that industry. A level-shift treatment was introduced to the seasonal adjustment for Recreational goods retailing sales.

In the December 2022 quarter the Supermarket and grocery stores stocks series (RTTQ.SFA9CS) was impacted by a level shift, due to new activity in the industry. A level shift was introduced to seasonal adjustment for the stocks series in Supermarket and grocery stores retailing.

Other complex treatments are listed in the Additive outliers and complex treatments table, available from Download documents.

Series which no longer have seasonal patterns

It is good practice to regularly review our seasonal adjustment methodology and we generally need a minimum of five years data to identify new patterns. Introducing changes to additive outliers due to COVID impacts gave us the opportunity to look at seasonality changes that have emerged since our 2017 redesign.

We have identified four series in the Retail Trade Survey that no longer show a significant seasonal pattern, and we have therefore turned seasonal adjustment off:

• Motor vehicles and parts retailing sales, current values (RTTQ.SFP1CA)
• Motor vehicles and parts retailing sales, volumes (RTTQ.SFP1KA)
• Accommodation stocks (RTTQ.SFU9CA)
• Electrical and electronic goods retailing stocks (RTTQ.SFL9CA).

Note that data will still be published under these seasonally adjusted series, but it will be identical to the actual data starting from the September 2017 quarter. Trend series will still be produced and published.

Other changes in treatment to specific series not outlined above are listed in the Additive outliers and complex treatments table, available from Download documents.

Impact on the data

The collective impact of all these enhancements will cause revisions to the seasonally adjusted and trend series as far back as the September 2017 quarter (inclusive). This was when the last redesign of the Retail Trade Survey was done.

The biggest differences are observed in the COVID-19-affected periods, and in the most recent periods (2023 quarters) as the latest data points in a series are always the most volatile.

The total retail series reflect each of their constituent series (for example, the All industries total sales, volume series is seasonally adjusted indirectly, by summing the seasonally adjusted series of each industry), therefore the changes are a reflection of series where methodological improvements have occurred. For example, changes to address a period without seasonality in the Accommodation series have had a greater impact on total retail sales than most other methodology changes.

A CSV file is available from Download documents for comparison, which contains the complete dataset up to the September 2023 quarter, both under the previous and the new methodology. This will allow you to see the impact of the methodology change, without the additional revisions to seasonally adjusted data that are brought in by a new quarter’s data. These figures will be revised, as is normal seasonal adjustment practice, with publication of the December 2023 quarter results on 23 February 2024.

ISBN 978-1-99-104971-1

Enquiries

Melissa McKenzie
03 964 8439
[email protected]