S2D Prediction Activities at CCCma(p)  HFP2 (2-tier seasonal) basis of CMC operational forecasts Analyses Trend in SIP forecasts  CHFP1 and 2 (1-tier seasonal/interannual) initial approach and results analyses and initialization development path  DHFP (decadal) potential predictability results WGCM/WGSIP/CMIP5 and decadal prediction

 Retrospective forecasts are crucial for - establishing forecast skill - providing forecast climatology for bias correction - guiding forecast calibration and post-processing Why retrospective forecasts?  Current EC 2-tier operational system: - 4 AGCMs x 10-ensemble - validated by 2 nd Historical Forecast Project (HFP2) - 4-month retrospective forecasts initialized each month  Validate also 1-tier coupled forecasts in CHFP

HFP2  CMC deterministic and probabilistic seasonal forecasts are based on HFP2  HFP2 (2 nd Historical Forecasting Project) retrospective 1-season forecast experiment operational context – no information from the future 2-tier - SST forecast provides boundary conditions for AGCM forecasts multi-model, multi-realization – each of 4 models produces 10-member ensemble forecasts we use results from 33-year period

Seasonal forecasts Jan-Feb-Mar 2009 Deterministic Probabilistic

Skill assessment of HFP2 forecasts  analysis by S. Kharin et al. (2008)  deterministic forecasts continuous valued categorical combining multi-model ensemble mean forecast information  unweighted  variance weighted (2 methods)  regression improved scale unweighted ensemble mean (single parameter) scale each model mean (four parameter) skill measures: correlation and MSSS  probabilistic categorical combining multi-model forecast information  count value  gaussian  adjusted gaussian skill measures: Brier skill score  benefits of Multi-model approach

Correlation and MSSS for Sfc air T methods of combining multi-model forecasts 1 parameter regression weighting 4 parameter regression weighting

Percent correct, 3-category forecasts Seasonal variation of skill No dependence on combination method

3-category probabilistic forecasts Count methods Gaussian methods Normal Below Adjusted gaussian Brier skill score Reliability

The virtues of the MME approach For same overall ensemble size more models give better scores

More models are better Diminishing returns Four models is a “practical” choice

Summary  Deterministic forecasts not critical how MME formed scaled MME result (1 parameter) increases MSSS but degrades correlation scaled MME result (4 parameters) degrades scores  Probabilistic forecasts gaussian method better than count adjusted gaussian only better over ocean (on account of persisted SSTAs)  Multi-model methods for same ensemble size more models the better because of diminishing returns, 4 models (as HFP2) is reasonable “practical” choice

Climate trends and seasonal forecasts  trend in HFP2 no GHG or other forcing in HFP2 implicit in persisted SSTAs and initial conditions characterization of trend statistical correction  trend in CHFP1 GHG forcing effect in coupled forecast model Boer (2008)

Background  Observational studies show long timescale trends thought to be due to GHG and other forcing - together with long timescale variability  Externally forced trends should provide an additional seasonal forecasting signal - if present and properly forecast  Do we capture these trends in the 33- years of retrospective forecast data of HFP2 (and does it matter)?

Linear trends in for DJF  For global means =  + - fit linear trends by least squares  forecast trend weaker than trend in NCEP reanalysis data  we also fit trends to each point Forecast NCEP years Units: o C -17  17

Trends in SAT from NCEP and HFP2 DJF NCEP DJF HFP2 JJA NCEP JJA HFP2 Units o C/decade

Trends in Z500 from NCEP and HFP2 DJF NCEP DJF HFP2 JJA NCEP JJA HFP2 Units m/decade

Trends  trends are larger at higher latitudes, over land and in winter – the signature of global warming  trends in forecasts are weaker than in reanalysis data  suggests that lack of GHG forcing may degrade the forecasts

 spatial average of MSE over globe  3 components error in trend error in non-trend component cross-product Time evolution of error

Dealing with anomalies minimizes MSE in DJF T850 Anomalies Trends set to zero at t=0 global meansMSE over globe trend total non-trend cross-product NCEP HFP2 Mean error Mean square error

Can we improve skill by adjusting trend?  try simple statistical adjustments based on scaling or trend adjustment  only cases where single parameter is estimated to avoid over-fitting  evaluate in cross-validation mode  use MSSS and Correlation scores

Statistical adjustments

Scaled Trend added Trend replaced Raw forecast

Trend replaced Raw forecast Scaled Trend added

Trend replaced Raw forecast Scaled Trend added

No radiative forcing trend in model Weak radiative forcing trend Stronger radiative forcing trend Temperature at 850hPa (T850) Anomaly Correlation 1 December Initialization Preliminary evidence of impact of radiative forcing trend on coupled model forecast skill Forcing can affect both land and ocean in this case

Summary  Observed/reanalysis trends have GHG signature  Trends in seasonal forecasts are weak  Statistically “correcting trends” gives some improvement in skill over Asia but not, unfortunately, over North America  Suggests seasonal forecasting models should include radiative forcing due to GHG and aerosols - physically realistic (if properly inserted) - some hope for improved forecasts  For coupled CHFP forecasts SSTs may also drift and we have some evidence GHG forcing is needed

The Coupled Historical Forecast Project, version 1: Formulation, results, and progress towards CHFP2 Bill Merryfield, George Boer, Greg Flato, Slava Kharin, Woo-Sung Lee, Badal Pal, John Scinocca Canadian Centre for Climate Modelling and Analysis Environment Canada

Pilot Project: CHFP1  Based on CGCM3.1/T63 (IPCC AR4)  Simple SST nudging initialization after Keenlyside et al. (Tellus 2005): - Strongly relax SST to observed time series - Anomalous wind stress tends to set up correct equatorial thermocline configuration: Asia SouthAmerica warm cool Anomalous wind stress

 Construct 10 initial conditions for 1 Sep (e.g.) by combining atm and ocn states from preceding week:  Launch forecasts 1 Feb, 1June, 1 Sep, 1 Dec  (10 ensemble members) x (4 initializations yr -1 ) x 30 yrs  1200 years of coupled model integration CHFP1 Ensemble Generation

CHFP1 Persistence Damped persistence CHFP1 Results (Ensemble size=10) Anomaly correlation Mean square skill score NINO3.4 skills Not too bad for off-the shelf model, crude initialization Much room for improvement, improvement being realized…

Coupled Forecast System Development Path CHFP2 “Off the shelf” CGCM Simple SST nudging initialization Ocean data assimilation CHFP1 CGCM development 2D Var after Tang et al. (JGR 2004) Improved error covariances Atmospheric initialization Insertion of reanalyses Atm data assimilation Land initialization GHG forcing, trend correction, etc Off-line forcing by bias-corrected reanalysis S assim after Troccoli et al. (MWR 2002) AGCM, OGCM Analysis and Verification

Obs SSTA Nov 1982 Deterministic forecast SSTA Nov 1982 New AGCM, OGCM Lead=11 mo Much improved model ENSO with new AGCM, OGCM Exemplified by “hit” for 11-month lead prediction of 1982/83 El Nino:

NINO3 Power Spectrum CGCM4 CGCM3.8 CGCM3.1 OBS OBS PERIOD (Y)

Ocean Data Assimilation  Initially use approach of Tang et al. (JGR 2004): - input ocean reanalysis in lieu of observations - simple variational assimilation level-by-level (2D Var) - background error covariances of Derber & Rosati (JPO, 1989)  Assimilate multiple ocean analyses  Explore methods to improve error covariances

MULTI-ANALYSISEXP_ATMOSEXP_OCEAN Ensemble member Atmosphere Initial State 8/31 8/ 31 8/ 30 8/ 29 8/ 28 8/ 27 8/ 26 8/31 Ocean Initial state 8/31 8/ 31 8/ 30 8/ 29 8/ 28 8/ 27 8/ 26 Used Reanalysis Data for ocean assimilation GODASECMWFGFDLSODAINGVMETUKGODAS Ocean Initialization by multi-analysis assimilation  Experiment: compare NINO3.4 skill and ensemble spread for three ensemble initialization strategies: - Multi-analysis: off-line assimilation of 6 ocean analysis products (same atm) - Exp_atmos: 6 AGCM states from consecutive days prior to forecast start (same ocn) - Exp_ocean: 6 OGCM states from consecutive days prior to forecast start (same ocn)  : 22 years of Sep 1–initialized forecasts

Correlation multi-reanalysis exp_atmos exp_ocean Lead month Ensemble Spread  RMS Error NINO3.4 skill and ensemble spread Lead month Lines : Ensemble mean Symbols : Ensemble members  Improved skill at longer leads  Larger ensemble spread in first two months SST Forecast Skill Multi-analysis ocean initialization leads to

Atmospheric Initialization  SST nudging informs AGCM of boundary forcing, but not correct synoptic configuration, i.e. weather  major loss of skill in first month of forecast  Two approaches are being pursued: - Direct insertion of atmospheric analysis (cf. HFP2) - Simple assimilation of analysis into AGCM

Land Initialization CFCAS/GOAPP funded collaboration with A. Berg (Guelph) Force land surface model with bias-corrected reanalyses after Berg et al. (Int J Clim 2005) >.5 Berg et al., 2003: 2005 Correlation of NCEP monthly precip with gauge-based measurements in USA: before bias correction after bias correction

 Coupled forecasts offer means for seasonal forecasting at longer leads, where future evolution of SSTA is critical  Prototype CHFP1 competitive with 4-model HFP2 at 1-month lead, but has only simplest initialization  CHFP1 provides a benchmark against which model and initialization improvements leading to CHFP2 can be assessed  CCCma participation in international CHFP through CLIVAR/WGSIP Summary

Many scientific “opportunities”  improved models  analysis of variability and of modes of variability  improved analysis methods especially in the ocean for model initialization for verification for model development  ensembles ensemble generation multi-model ensembles  prediction studies and the “DHFP” WGSIP/WGCM/CMIP5 coordinated project

Coupled Forecast System Development Path CHFP2 DHFP1 “Off the shelf” CGCM Simple SST nudging initialization Ocean data assimilation CHFP1 CGCM development 2D Var after Tang et al. (JGR 2004) Improved error covariances Atmospheric initialization Insertion of reanalyses Atm data assimilation Land initialization GHG forcing, trend correction, etc Off-line forcing by bias-corrected reanalysis S assim after Troccoli et al. (MWR 2002) AGCM, OGCM Analysis and Verification

End of presentation