Historical perspective: earlier ensembles and forecasting forecast skill. (2nd September 2019)

• Record Type: Journal Article
• Authors: Kalnay, Eugenia
• Other Names: Buizza Roberto guestEditor.; Weisheimer Antje guestEditor.
• Is Part Of: Quarterly journal of the Royal Meteorological Society. Volume 145(2019)Supplement 1
• Page Start: 25
• Page End: 34
• Publication Date: 2019-09-02
• Subjects: Combining multiple forecasting systems (Poor person ensembles) -- Early ensemble forecasting -- Forecasting forecast skill; Meteorology -- Periodicals; 551.5
• DOI: 10.1002/qj.3595
• Languages: English
• ISSNs: 0035-9009

Abstract

Abstract: The history of ensemble forecasting is briefly reviewed, including the pioneering approaches of stochastic dynamical forecasting (Epstein, 1969, Tellus, 21, 739–759), Monte‐Carlo forecasting (Leith, 1974b, Monthly Weather Review, 102, 409–418), Lagged Average Forecasting (Hoffman and Kalnay, 1983, Tellus, 35A, 100–118) and Scaled LAF (Ebisuzaki and Kalnay, 1991, pp. 6.31–6.32), as well as early predictions of forecast skill (McCalla and Kalnay, 1988, pp. 634–640; Palmer and Tibaldi, 1988, Monthly Weather Review, 116, 2453–2480; Kalnay and Ham, 1989, pp. 24–27; Molteni and Palmer, 1990, Monthly Weather Review, 119, 1088–1097), and their operational results (Wobus and Kalnay, 1995, Monthly Weather Review, 123, 2132–2148). Ensembles of multiple models (aka Poor person Ensemble Prediction Systems or PEPS) are generally quite successful compared to stochastically perturbed ensembles probably because the centers that created them each try to be one of the best, if not the best, so their uncertainties in the models and the analyses are more realistic than just stochastic perturbations. Abstract : Schematic of the first two practical ensemble‐forecasting methods: Monte Carlo (MC, Leith, 1974b) forecasts (left) and Lagged Average Forecasts (LAF, Hoffman and Kalnay, 1983, right). The crosses represent the analyses at the initial time, and the dots are Monte Carlo perturbations. The MC ensemble average with m members reduces the long‐term forecast error covariance compared to the error covariance of a climatological forecast from 2 to (1 + 1/ m ). LAF takes advantage of operational available forecasts, and thus the LAF initial perturbations include "errors of the day", which improves forecast of the skill compared to MC initial perturbations.