Wednesday, Aug 5: 8:30 AM - 10:20 AM
1746
Topic-Contributed Paper Session
Thomas M. Menino Convention & Exhibition Center
Room: CC-252A
David Findley (1940-2025) was a fellow of ASA, and a pioneer in the field of time series and seasonal adjustment. This session reviews his impact on model selection, seasonal adjustment, seasonality diagnostics, and other topics, in terms of both methodological development and software implementation. The speakers are all statistical experts who had substantial interaction or collaboration with David Findley over the course of his 35-year career at the U.S. Census Bureau. The session should be of interest to persons who knew David Findley, or were familiar with his work, as well as those who have an interest in the topics of seasonal adjustment, model selection, and seasonality. The session format consists of five speakers: William Bell (Senior Statistician for Small Area Estimation, US Census Bureau), John Aston (Harding Professor of Statistics in Public Life, University of Cambridge), Brian Monsell (NSR Consulting and formerly Time Series Principal Researcher at US Census Bureau), Donald Martin (Professor, North Carolina State University), and Tucker McElroy (Senior Statistician for Time Series, US Census Bureau). The chair is Tommy Wright (Director of the Center for Statistical Research and Methodology, US Census Bureau).
Applied
No
Main Sponsor
Business and Economic Statistics Section
Co Sponsors
Government Statistics Section
Presentations
David Findley (1940-2025) was a fellow of the ASA and a pioneer in the field of time series and seasonal adjustment, with a 35-year career at the U.S. Census Bureau. This talk reviews his methodological impact on model selection and seasonality diagnostics, and discusses recent advances that are motivated by his work. First we trace the development of ideas stemming from comparison of forecast error sample paths, and develop a comparison test for non-stationary time series models. Second, we review both frequency and time domain developments in seasonality diagnostics, and describe how these are assessments of persistency associated with seasonality. We present the theory and empirical performance of the new diagnostics for unit root specification and seasonality.
In this talk, we first consider David Findley's excellent work on seasonal adjustment of time series. The focus will be on a 2006 paper that considered frequency domain properties of finite (length 49 and 109) concurrent and symmetric SEATS and X-11/12-ARIMA filters for monthly time series. Two main conclusions of the paper were that the squared gains of infinite ARIMA model-based filters are not reliable diagnostics for series of the lengths given above, and the squared gains and phase delays of the concurrent seasonal adjustment filters provide information that is different from (and more valuable for real-time analysis than) that provided by the squared gains of symmetric filters. Beyond that, the work illustrates Dr. Findley's commitment to excellence in sharing informative and useful research.
In the latter part of the talk, we look at data filtering in a very different sense: that of efficiently filtering through Markovian time series to compute distributions of pattern statistics using an auxiliary Markov chain with minimal state spaces.
This talk will start by showing how seasonal adjustment software developed at the Census Bureau's Time Series Research Staff evolved under the leadership of David Findley, starting with PC versions of X-11 and X-11 ARIMA to the software used currently, X-13ARIMA-SEATS. The software added capabilities related to seasonal adjustment diagnostics, regARIMA modeling capabilities, and automated model selection (including outlier detection and AICC related testing).
The last half of the talk will show more recent developments in seasonal adjustment software at the US Census Bureau and innovations in methodology at the US Bureau of Labor Statistics and other agencies. These innovations are seen in adopting a more recent software architecture (Java for JDemetra+ and Python for SeasCen), as well as methodological changes in utilizing unobserved component models at the Bureau of Labor Statistics.
Keywords
Seasonal Adjustment
Signal Extraction
Statistical Software
High Frequency Time Series
Unobserved Component Models
Python
David Findley arrived at the Census Bureau in 1980 to build a staff for research and consulting on seasonal adjustment methods and other time series topics. This, in turn, led to a resurrection of the development of seasonal adjustment software at Census and to the development of new software for time series modeling. I will start with a brief review of time series modeling software developments up through the X-12 and X-13ARIMA-SEATS programs, focusing on some methodological aspects. I will then continue this theme with a discussion and illustration of the modeling capabilities of the RegComponent software which is included in the new SeasCen software platform.
Keywords
RegARIMA model
ARIMA component model
seasonal adjustment
structural model
David was an accomplished mathematician who went from being a functional analyst to being a statistician who worked on official statistical problems. In this talk, we will take a combined approach, celebrating both aspects of David's life, and looking at how functional data analysis methods can be used in official statistical problems. We will do this in relation to analyses of temporally and spatially related curves (dependence being another of David's interests). We will show that many public policy questions can be framed using functional data and functional concepts of smoothness and continuity used to help address these, gaining insight into problems with considerable policy implications. [Joint work with Luke Barratt, Humboldt-Universität zu Berlin].
Keywords
Functional Data Analysis
House Price Statistics
COVID-19
Spatial Statistics
Time Series