Proposal Description

We would like to present new statistical frameworks and corresponding cutting-edge tools for solving common time series analysis problems in official statistics, such as imputing missing values, optimizing seasonal adjustment, filtering, benchmarking and temporal disaggregation.
Most of these tools are related to the JDemetra+ software (www.jdemetra.org), the reference software for time series analysis that is specifically designed to address the statistical problems encountered in producing official statistics. It is recommended by Eurostat and widely used around the world.

We would like to include the following four talks, leaving one slot open for a potential local contribution. This could lead to new collaborations, for example with African countries, some of which are already JDemetra+ users.

The rjdverse R package Suite for Time Series Econometrics in Official Statistics, Anna Smyk, Insee, France

JDemetra+ provides algorithms for seasonal adjustment, trend-cycle extraction, outlier detection, nowcasting, and revision analysis. All its algorithms, implemented in Java, are accessible through a graphical user interface and also through an ecosystem of R packages, the rjdverse (www.github.com/rjdverse). We would like to showcase some of its advanced forecasting (nowcasting) and seasonal adjustment capabilities, including high-frequency data. But also provide practical applications of its state-space modelling framework.

Multivariate techniques for Benchmarking and Temporal Disaggregation, Corentin Lemasson, National Bank of Belgium

Multivariate techniques for Benchmarking and Temporal Disaggregation handle systems of time series variables rather than treating each series independently. Unlike traditional univariate methods, multivariate methods can enforce contemporaneous or accounting constraints alongside the temporal ones and have the advantage of using more information in the estimation process.

The R package rjd3bench provides access to the highly efficient JDemetra+ algorithms for temporal disaggregation, benchmarking and reconciliation. This presentation introduces the package through practical examples, with a particular focus on reconciliation and multivariate temporal disaggregation.

Cross-validation in X-13 and STL, Daniel Ollech, Bundesbank, Germany

Until recently, users of the seasonal adjustment methods in X-13 and STL had to rely either on rigid ad hoc rules or on subjective visual inspection to select appropriate seasonal filters. To address this limitation, we introduce a cross-validation framework for these methods that provides a data-driven, objective mechanism for filter selection. In this study, we evaluate its performance in extensive simulation studies, systematically assessing their accuracy, stability, and adaptability under varying seasonal patterns. Additionally, we demonstrate their practical utility through real-world applications, encompassing both conventional monthly and quarterly time series as well as higher-frequency data.

Time Series Imputation Using a Neural Network Model, Luis Sanguiao Sande, INE, Spain

Sensor data from traffic loops and cameras is valuable for estimating road tourism, but it often has gaps due to sensor malfunctions. These gaps require imputation to restore the time series. To address this issue, a novel deep learning model was developed using Python, combining layers for temporal features and autoencoder-like layers for spatial features. This architecture allows for additional regressors and employs a sequential training process in which the components are pre-trained independently and subsequently fine-tuned collectively. This flexible implementation effectively handles time-varying gaps and can be easily adapted to other high-frequency, multivariate time series.