Online Program

Toward Near Real-Time Economic Indicators Using Time Series Models: Statistics Canada’s Progress to Date (308005)

Keywords: nowcasting, time series, timeliness

The use of traditional methods for producing reliable estimates requires substantial time to compile, validate and analyze, resulting in significant delays between the end of the reference period and the time of publication. In order to improve the timeliness of its economic indicators, Statistics Canada is investigating the use of time series and now-casting methods to make estimates available in near-real time. These time series models include components that reflect long-term trends and seasonality from historical data, but can be significantly improved by including covariates in the model to make them more robust and efficient. The time series models that can be considered range from classical approaches such as regARIMA and State Space Models, to modern methods such as machine learning and neural networks. The potential covariates are many and are very diverse (including high-frequency data, other related indicators, partial information and big data sources). This presentation will give an overview of the progress made to date to identify preferred time series models and variable selection procedures. A proof-of-concept study conducted on Canadian retail trade data will be presented, illustrating the trade-off between aspects of quality such as timeliness and precision. The presentation will also include a discussion of practical considerations and a proposed set of criteria for publication of indicators in near real time in the context of official statistics.