Cargando…
The Weighted‐Average Lagged Ensemble
A lagged ensemble is an ensemble of forecasts from the same model initialized at different times but verifying at the same time. The skill of a lagged ensemble mean can be improved by assigning weights to different forecasts in such a way as to maximize skill. If the forecasts are bias corrected, th...
Autores principales: | , , |
---|---|
Formato: | Online Artículo Texto |
Lenguaje: | English |
Publicado: |
John Wiley and Sons Inc.
2017
|
Materias: | |
Acceso en línea: | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5784419/ https://www.ncbi.nlm.nih.gov/pubmed/29399270 http://dx.doi.org/10.1002/2017MS001128 |
_version_ | 1783295446309928960 |
---|---|
author | DelSole, T. Trenary, L. Tippett, M. K. |
author_facet | DelSole, T. Trenary, L. Tippett, M. K. |
author_sort | DelSole, T. |
collection | PubMed |
description | A lagged ensemble is an ensemble of forecasts from the same model initialized at different times but verifying at the same time. The skill of a lagged ensemble mean can be improved by assigning weights to different forecasts in such a way as to maximize skill. If the forecasts are bias corrected, then an unbiased weighted lagged ensemble requires the weights to sum to one. Such a scheme is called a weighted‐average lagged ensemble. In the limit of uncorrelated errors, the optimal weights are positive and decay monotonically with lead time, so that the least skillful forecasts have the least weight. In more realistic applications, the optimal weights do not always behave this way. This paper presents a series of analytic examples designed to illuminate conditions under which the weights of an optimal weighted‐average lagged ensemble become negative or depend nonmonotonically on lead time. It is shown that negative weights are most likely to occur when the errors grow rapidly and are highly correlated across lead time. The weights are most likely to behave nonmonotonically when the mean square error is approximately constant over the range forecasts included in the lagged ensemble. An extreme example of the latter behavior is presented in which the optimal weights vanish everywhere except at the shortest and longest lead times. |
format | Online Article Text |
id | pubmed-5784419 |
institution | National Center for Biotechnology Information |
language | English |
publishDate | 2017 |
publisher | John Wiley and Sons Inc. |
record_format | MEDLINE/PubMed |
spelling | pubmed-57844192018-02-02 The Weighted‐Average Lagged Ensemble DelSole, T. Trenary, L. Tippett, M. K. J Adv Model Earth Syst Research Articles A lagged ensemble is an ensemble of forecasts from the same model initialized at different times but verifying at the same time. The skill of a lagged ensemble mean can be improved by assigning weights to different forecasts in such a way as to maximize skill. If the forecasts are bias corrected, then an unbiased weighted lagged ensemble requires the weights to sum to one. Such a scheme is called a weighted‐average lagged ensemble. In the limit of uncorrelated errors, the optimal weights are positive and decay monotonically with lead time, so that the least skillful forecasts have the least weight. In more realistic applications, the optimal weights do not always behave this way. This paper presents a series of analytic examples designed to illuminate conditions under which the weights of an optimal weighted‐average lagged ensemble become negative or depend nonmonotonically on lead time. It is shown that negative weights are most likely to occur when the errors grow rapidly and are highly correlated across lead time. The weights are most likely to behave nonmonotonically when the mean square error is approximately constant over the range forecasts included in the lagged ensemble. An extreme example of the latter behavior is presented in which the optimal weights vanish everywhere except at the shortest and longest lead times. John Wiley and Sons Inc. 2017-11-29 2017-11 /pmc/articles/PMC5784419/ /pubmed/29399270 http://dx.doi.org/10.1002/2017MS001128 Text en © 2017. The Authors. This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs (http://creativecommons.org/licenses/by-nc-nd/4.0/) License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made. |
spellingShingle | Research Articles DelSole, T. Trenary, L. Tippett, M. K. The Weighted‐Average Lagged Ensemble |
title | The Weighted‐Average Lagged Ensemble |
title_full | The Weighted‐Average Lagged Ensemble |
title_fullStr | The Weighted‐Average Lagged Ensemble |
title_full_unstemmed | The Weighted‐Average Lagged Ensemble |
title_short | The Weighted‐Average Lagged Ensemble |
title_sort | weighted‐average lagged ensemble |
topic | Research Articles |
url | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5784419/ https://www.ncbi.nlm.nih.gov/pubmed/29399270 http://dx.doi.org/10.1002/2017MS001128 |
work_keys_str_mv | AT delsolet theweightedaveragelaggedensemble AT trenaryl theweightedaveragelaggedensemble AT tippettmk theweightedaveragelaggedensemble AT delsolet weightedaveragelaggedensemble AT trenaryl weightedaveragelaggedensemble AT tippettmk weightedaveragelaggedensemble |