Home: 3915 Fulton Street, N.W. Washington, DC 20007-1376 |
Consultant: U.S. Census Bureau Statistical Research Division 4600 Silver Hill Road Washington, DC 20233 |
|||
Telephone: (202) 337-7101 Fax: (202) 337-7039 david.findl ey@ieee.org |
Telephone: (301) 763-8773 Fax: (301) 763-8399 david.f.fin dley@census.g ov |
University of Cincinnati 1958-62 B.S. in Mathematics
University of
Cincinnati 1962-63 M.A. in Mathematics
University of Maryland 1963-64
(Research Assistant)
University of Heidelberg 1964-65 (Hilfsassistent)
University of Frankfurt 1966-67 Dr. Phil. Nat. in Applied Mathematics
(Magna cum laude)
From June 2009, Consulting and training related to seasonal adjustment
July 1980-June 2009 U.S. BUREAU OF THE CENSUS Washington, D.C.
From March 2002, Senior Mathematical Statistician for Time Series Methods. Before this, Principal Researcher for Time Series Analysis. My duties included investigating and approving proposed changes in Census Bureau time series analysis procedures, providing research leadership for the Time Series Staff of the Statistical Research Division, and carrying out independent research to improve the Census Bureau's seasonal adjustment and related procedures. I also led the organization of five international research conferences.
1996-2001 GEORGE WASHINGTON UNIVERSITY
Research Consultant in the Department of Statistics. I supervised the PhD Dissertation research of James L. Cantor (PhD, 2001). This was described as the best doctoral dissertation done in the Department in the decade up to its completion.
2001 INSTITUTES OF STATISTICAL SCIENCE AND ECONOMICS Academia Sinica Taipei
Visiting Research Fellow (6 weeks). Working with Ching-Zong Wei, I completed a paper on model selection. I also presented a short course on time series modeling and seasonal adjustment with X-12-ARIMA.
1994 VICTORIA UNIVERSITY Wellington, New Zealand
Visiting Research Fellow sponsored by the New Zealand Foundation for Research, Science and Technology (two weeks). I commented on a variety of seasonal adjustment research projects underway at the Institute of Statistics and Operations Research and worked with Thomas Mikosch to extend the theory supporting my graphical model-selection procedure to cover more situations, including models for time series with infinite variance.
1993 INSTITUTE OF STATISTICAL SCIENCE Academia Sinica Taipei
Visiting Research Fellow (6 weeks). Working with Ching-Zong Wei, I completed a paper on AIC and developed the supporting large sample theory for some earlier work of mine on the use of recursive estimation procedures for model selection.
1992 UNIVERSITY OF LANCASTER United Kingdom
Senior Research Fellow of the UK Science and Engineering Research Council (one month). I developed a theory of model overfitting to data due to parameter estimation and gave lectures on my model selection research.
1991 INSTITUTE OF STATISTICAL SCIENCE Academia Sinica Taipei
Visiting Research Fellow (five months). I completed a research program with Ching-Zong Wei on moment bounds useful for model selection theory supporting Akaike's Information Criterion and gave lectures on model selection theory.
1987 INSTITUTE OF STATISTICAL MATHEMATICS Tokyo
Visiting Professor (four months). I carried out research on model selection procedures and (with Genshiro Kitagawa) their application to ship autopilot design. I was the first person to hold a Visiting Professorship after a reorganization of the Institute into a PhD granting institution that created such positions.
1986 UNIVERSITY OF WISCONSIN Madison, Wisconsin
University Lecturer (one week). I lectured on the use of bootstrap methods in population forecasting.
1985 UNIVERSITY OF HEIDELBERG Federal Republic of Germany
Visiting Professor of the Institute for Applied Mathematics (3 months). I carried out research on minimax procedures for determining the decomposition components of model-based seasonal adjustment as well as research on the uniqueness of time series representations.
1982-1984 BABCOCK AND WILCOX COMPANY Lynchburg, Virginia
Consultant for the design of statistical controllers for thermoelectric power plants utilizing TIMSAC programs and methods.
1975-1982 UNIVERSITY OF TULSA Tulsa, Oklahoma
Associate Professor of Mathematical Sciences (Tenured September 1, 1979). My activities included teaching a two-semester graduate statistical time series course each year and the organization of two international research conferences, the Applied Time Series Symposium of May 14-15, 1976 and the Second Applied time Series Symposium of March 3-5, 1980. I arranged for the University of Tulsa to be the distributor of the TIMSAC computer program packages developed at the Institute of Statistical Mathematics in Tokyo. These were the most technically advanced statistical time series packages available at the time.
1977-1980 CITIES SERVICE OIL COMPANY Tulsa, Oklahoma
Consultant in Time Series Analysis to the Geophysical Research Group in the Laboratory of Cities Service Oil Company (full-time in the summers, one day a week during the academic year). My work included the application of Kalman filtering and time-varying autoregressive models to seismic signal processing. Several of the techniques which I developed and programmed were incorporated into the production program library.
1968-1975 UNIVERSITY OF CINCINNATI Cincinnati, Ohio
Assistant Professor of Mathematics. My duties included teaching the beginning graduate courses in functional analysis, mathematical statistics, and probability as well as undergraduate courses in time series analysis and design of experiments.
1973 CAMBRIDGE UNIVERSITY United Kingdom
Visiting Scholar of St. John's College (four months). I carried out research on extensions of calculus for function spaces.
1967-1968 UNIVERSITY OF FRANKFURT Germany
Wissenschaftlicher Assistent. My activities included supporting masters (Diplom) and doctoral students and leading an advanced seminar in Functional Analysis.
1965-1966 UNIVERSITY OF CINCINNATI Cincinnati, Ohio
Instructor of Mathematics. I taught first and second year undergraduate courses.
Julius Shiskin Award Committee (1999-present)
Committee on Fellows of
the American Statistical Association (1992-96)
Associate Editor of the
Annals of Statistics (1983-92)
Chair (2001),
Program Chair (1988) and Secretary-Treasurer (1984) of the Business and
Economics Section of the American Statistical Association
Council of the
American Statistical Association (1978-80)
American Statistical Association (Life Member)
Institute of Electrical and Electronic Engineers (Life Member)
Institute
of Mathematical Statistics (Life Member)
Royal Statistical Society (Life
Member)
Sigma Xi
U.S. Department of Commerce Gold Medal (1997)
1996 Julius Shiskin Award
(Sponsored by the Washington Statistical Society, the Business and Economic
Statistics Section of the American Statistical Association, and the National
Association of Business Economists)
Fellow of the American Statistical
Association (1987)
U.S. Department of Commerce Silver Medal (1986)
U.S. Department of Commerce Bronze Medal (1983)
Fluency in German
Limited speaking ability in Japanese and Spanish
Technical reading knowledge of French, German, Italian, Russian and Spanish
Since 1980, my research has chiefly been concerned with time series modeling and model selection
procedures and with seasonal adjustment and diagnostics for seasonal adjustment. Most of the resulting
articles and papers are co-authored, and it has been my good fortune to have
many exceptionally able collaborators. Some of the model selection work is
broadly motivated, but most of the research has been connected to efforts to
strengthen the capabilities and diagnostics of the U.S. Census Bureau's
seasonal adjustment software. This software, currently X-12-ARIMA version 0.3,
is used worldwide to estimate seasonal and other calendar effects in monthly
and quarterly economic data. It is freely downloadable from
http://www.census.gov/srd/www/x12a/x12downv03_pc.html
A Windows interface program WinX12 that greatly facilitates the use of
X-12-ARIMA can be found at http://www.census.gov/srd/www/winx12/
X-12-ARIMA will be replaced later in 2010 with X-13-ARIMA-SEATS; see [11]
below for more information about the latter and about the state of seasonal
adjustment methodology around 2005. For update announcements regarding
X-12-ARIMA and related software, register with
http://lists.census.gov/mailman/listinfo/x12a-announce.
Regarding recent modeling research, the article [4] presents a successful new type of regression model for estimating day-of-week effects in stock time series, such as end-of-month inventories. The more recent paper [47] derives regression models for estimating the effects of moving holidays, like Easter, in such data. Further material and references on moving holiday effect modeling and detection can be found in [57] and [49]. The report [48] presents the current state of development of a new class of seasonal time series models for forecasting or seasonally adjusting time series with more heterogeneous seasonal components than classical seasonal ARIMA models can represent.
In the area of model selection, there are two veins of research:
(i) Forecast comparison tests and diagnosics. The article [3] presents two new statistical tests for deciding if one ARIMA model provides better multi-step-ahead forecast performance than another. Both improve upon the widely used Diebold-Mariano test. One replaces the latter test's inconsistent standard error approximation with a consistent estimator. Earlier alternatives to this new test were presented in [62], [22] and [23] without a complete large-sample theory. The other new test goes further by also accounting for the effects of model parameter estimation. The article [9], supplemented by [6], provides supporting theory for X-12-ARIMA's graphical diagnostic for comparing the out-of-sample forecast performance of two not necessarily correct regARIMA models. The article [6] develops asymptotic squared error and error autocovariance formulas for models with underspecified regressors and with ARIMA models for the regression error that might likewise be incorrect. Both for OLS and GLS regression coefficient estimation are covered.
(ii) Akaike's Information Criterion (AIC) and related criteria.The article [13] presents the first mathematically complete derivation of AIC for vector autoregressive models as well as a precise concept and measure of data overfitting by a model, together with theory connecting the value of the measure with a corresponding increase in mean square prediction error. The article also establishes the limit formulas (3.6) and (3.7) of [26]. In [26], these formulas are used to show that certain incorrect, non-nested autoregressive model pairs, the two models of which have differing numbers of coefficients whose estimates converge to zero, provide counterexamples to the principal of parameter parsimony and to the frequently encountered statement that consistent model selection criteria like Schwarz's BIC are generally to be preferred over AIC. (There has been confusion regarding statements in [26] about sequences being bounded in probability. The statements concern sequences that are known or shown, e.g. in (3.4), to converge in distribution: it seems not to be widely noted that such convergence implies boundedness in probability, even though this fact follows easily from the definitions.) The Encyclopedia of Statistical Science article [12] reviews the concepts underlying Akaike's minimum AIC criterion at a fairly elementary level and mentions some generalizations. The Statistical Science interview of Akaike [21] details some of his extensive experiences and broad perspectives regarding statistical modeling.
Regarding research on seasonal adjustment and
associated diagnostics, the innovations in methodology
and diagnostics of the original release of X-12-ARIMA are described in some
detail in [18]. More information about its seasonal adjustment stability
diagnostics and their connection with accuracy can be found in [28], [32] and
[56]. Background on the spectrum diagnostics and other frequency domain
diagnostics is given in [55], [58] and [8]. The article [16] illustrates the
application to eleven time series of a step-by-step procedure for (i)
producing an initial seasonal adjustment with the aid of automatic or default
options of X-12-ARIMA; (ii) using the software's diagnostics to determine if
its calculated seasonal adjustment (initial or otherwise) is acceptable and,
if it is not; (iii) identifying alternative software options to try in order
to remedy the problems identified in (ii). In the current version of
X-12-ARIMA, the automatic ARIMA modeling modeling command mentioned in [16]
now invokes a different automatic modeling procedure, one very closely based
on the procedure of the TRAMO-SEATS seasonal adjustment software. Also, the
more recently developed WinX12 interface, whose url was given above, can
produce most of the diagnostic graphs of the graphics program referred to in
[16], X-12-Graph. The latter program is now only available as a batch program,
which, in a single run, can produce all specified graphs from a multiple
series run of X-12-ARIMA. It can be downloaded from
http://www.census.gov/srd/www/x12graph/
U. S. Census Bureau's Seasonal Adjustment
Papers Site
http://www.census.gov/srd/www/sapaper/sapaper.html
Historical Papers Concerning X-11 Seasonal Adjustment Papers Site
http://www.census.gov/srd/www/sapaper/historicpapers.html
David Findley's Papers on the Statistical Research Division Research
Reports Site
http://www.census.gov/srd/www/byname.html#findleydavidf