23-27 Oct. 2006NOAA 31st Annual Climate Diagnostics and Prediction Workshop Predictability & Prediction of Seasonal Climate over North America Lisa Goddard, Simon Mason, Ben Kirtman, Kelly Redmond, Randy Koster, Wayne Higgins, Marty Hoerling, Alex Hall, Jerry Meehl, Tom Delworth, Nate Mantua, Gavin Schmidt (US CLIVAR PPAI Panel)
Time Series of Prediction Skill operational Potential predictability Research forecasts (1) Understand the limit of predictability (2) Identify conditional predictability (e.g. state of ENSO or Indian Ocean) (3) Document the expected skill to judge potential utility of the information for decision support (4) Set a baseline for testing improvements to prediction tools and methodologies (5) Set a target for real-time predictions. (Courtesy of Arun Kumar & Ants Leetmaa)
Real-time prediction skill… North America, 1-month lead, seasonal terrestrial climate Provide a template for verification - What are the best metrics? Best for who? - Pros & cons of current metrics - Can we capture important aspects of variability (e.g. trends, drought periods)? Estimate skill of real-time forecasts - How predictable is N. America climate? - Benefit of multi-model ensembling? Provide baseline against which we can judge future advances - How best to archive/document for future comparison? - Are we missing something? (i.e. statistical models)
Forecast Data Dynamical models (single): CCCma – Canadian Centre for Climate Modeling and Analysis KMA – Korean Meteorological Agency MGO- Main Geophysical Observatory, Russia NASA/GMAO-National Aeronautics and Space Administration, USA RPN – Canadian Meteorological Centre ECHAM4.5 – MPI (run at IRI) CCM3.6 – NCAR (run at IRI) ECMWF –European Center for Medium Range Weather Forecasts Meteo-France – Meteorological Service, France LODYC- Laboratoire d'Océanographie Dynamique et de Climatologie, France Met Office – UK Meteorological Office MPI – Max Planc Institute for Meteorology, Germany CERFACS – European Centre for Research and Advanced Training in Scientific Computing, France INGV-Instituto Nazionale di Geofisica e Vulcanolgia, Italy NOAA-CFS – National Oceanic Atmospheric Administration, USA Multi-Model of dynamical models (simple average) Statistical models (from CPC): CCA, OCN (others?) Multi-Model of dynamical + statistical models
ModelNXNYNMLS CCCma-GCM Mar1969-Dec2003 by 3 CCCma-GCM Mar1969-Dec2003 by 3 KMA Jan1979-Dec2002 MGO Nov1978-Nov2000 by 3 NASA-GMAO Feb1993-Nov2002 by 3 RPM Mar1969-Dec2000 by 3 ECHAM Jan1958-Dec2002 CCM Jan1958-Dec2002 ECMWF Feb1958-Nov2001 by 3 Meteo-France Feb1958-Nov2001 by 3 LODYC Feb1974-Nov2001 by 3 MetOffice Feb1959-Nov2001 by 3 MPI Feb1969-Nov2001 by 3 CERFACS Feb1980-Nov2001 by 3 INGV Feb1973-Nov2001 by 3 CFS Jan1981-Dec2003 Forecast Data
ModelNXNYNMLS CCCma-GCM Mar1969-Dec2003 by 3 CCCma-GCM Mar1969-Dec2003 by 3 KMA Jan1979-Dec2002 MGO Nov1978-Nov2000 by 3 NASA-GMAO Feb1993-Nov2002 by 3 RPM Mar1969-Dec2000 by 3 ECHAM Jan1958-Dec2002 CCM Jan1958-Dec2002 ECMWF Feb1958-Nov2001 by 3 Meteo-France Feb1958-Nov2001 by 3 LODYC Feb1974-Nov2001 by 3 MetOffice Feb1959-Nov2001 by 3 MPI Feb1969-Nov2001 by 3 CERFACS Feb1980-Nov2001 by 3 INGV Feb1973-Nov2001 by 3 CFS Jan1981-Dec2003 Forecast Data: JJA & DJF ( )
Verification Data & Metrics OBSERVATIONAL DATA: 2.5x2.5 deg 2m T: CRU-TSv2.0 ( ) Precipitation: CMAP ( ) VERIFICATION MEASURES Metrics consistent with WMO - SVS for LRF (Standardised Verification System for Long Range Forecasts) Deterministic information : - MSE & its decomposition - correlation, mean bias, & variance ratio Probabilistic information: - Reliability diagrams, regionally accumulated - ROC areas for individual grid boxes
Mean Squared Error
Pro: * Gives some estimate of uncertainty in forecast (i.e. RMSE). Con: * Can not infer frequency of large errors unless precise distributional assumptions are met. Recommendation: * Perhaps simple graph or table showing frequency of errors of different magnitudes would be appropriate.
Correlation : Temperature DJF
Correlation : Temperature JJA
Correlation : Precipitation DJF
Correlation : Precipitation JJA
Correlation Pros: * Commonly used; familiar * Gives simple overview of where models are likely to have skill or not Con: * Merely measure of association, not of forecast accuracy Recommendation: * Avoid deterministic metrics
Example Ensemble forecasts of above-median March – May rainfall over north-eastern Brazil
ROC Areas : DJF Temperature BELOW-NORMAL
ROC Areas : DJF Temperature ABOVE-NORMAL
ROC Areas : JJA Temperature ABOVE-NORMAL
ROC Areas : JJA Temperature BELOW-NORMAL
ROC Areas : DJF Precipitation ABOVE-NORMAL
ROC Areas : DJF Precipitation BELOW-NORMAL
ROC Areas : JJA Precipitation ABOVE-NORMAL
ROC Areas : JJA Precipitation BELOW-NORMAL
ROC Areas Pros: * Can treat probabilistic forecasts * Can be provided point-wise * Can distinguish ‘asymmetric’ skill Cons: * Fails to address reliability
RELIABILITY
Reliability Pros: * Treats probabilistic forecasts * Relatively easy to interpret * Provides most relevant information on usability of forecast information over time Cons: * Difficult to provide for individual grid points, especially for short time samples
Temperature Trends over North America %-Area Covered by “Above-Normal”
Temperature Trends over North America %-Area Covered by “Above-Normal”
Observed Precipitation over North America Anomalies relative to Percent difference relative to JJA DJF Frequency (# years out of 4) for precipitation in BN category
Frequency of Below-Normal Precipitation JJA in 4 2 in 4 3 in 4 4 in 4 OBSERVATIONS
Frequency of Below-Normal Precipitation DJF in 4 2 in 4 3 in 4 4 in 4 OBSERVATIONS
Summary What’s an appropriate template? - Skill metrics should be flexible (i.e. user defined “events”, categories, thresholds) - Probabilistic forecasts must be treated probabilistically!!! How are we doing? - Could be better. Encouraging performance estimates by some measures, but inadequate performance on important aspects of climate variability. - Missing elements necessary for seasonal prediction? Baseline??