1. Seasonal Forecasting in Canada
G.J. Boer
Canadian Centre for Climate Modelling and Analysis
Environment Canada

2. People involved
Many people involved to a greater or lesser extent over the years…including
- McGill
- CCCma
- RPN
- CMC
- others

3. Evolution of SP in Canada
Long slow rather low priority evolution…
Winter of 82/83 was a “wake up” call
Traditional (subjective) forecasts were for cold winter
Some (even in AES) pointed to a volcano and predicted cold winter

5. Early studies initial studies of “potential predictability”
observed tropical SST anomaly imposed as boundary condition provides forcing
international “response” experiments
in ACRN we analyzed AGCM1 (T20) response
served as test bed for seasonal forecast approach

6. Observed T850

7. Modelled T850 with specified SSTA

8. Early 2-tier forecast experiment
Objective, operational approach
2-tier with no information from the future
- 1st tier: SSTA forecast (by persistence)
- 2nd tier: AGCM forecast with reanalysis initial conditions
Ensembles of forecasts
Retrospective forecasts for January from 79-86
Systematic error correction
Measures of skill

10. Geographical distribution of skill and spread/skill relationships
(Boer,1993: MWR)

11. History of SP/HFP
CMC had need for operational seasonal forecasts but limited resources
Based on previous work we proposed 1-season 2-tier forecasts with AGCM
- subjective approach costly and experience lacking
- statistical approach would need development
- prefer physically-based approach with evolution path
Externally funded loose (part time) “network”
- University and Government researchers
- CMC computers
- Forum of researchers, meteorologists, management
Result was the HFP and ultimately operational seasonal forecasts at CMC

12. Sacred Principles of SP/HFP
Forecasts to be based on clearly described and justified objective methods
Forecasts must provide quantitative results that may be objectively verified
Forecasts must be accompanied by measures of historical skill
Changes in forecast procedures require objective evidence of improvement

13. Historical forecasting project (HFP1)
Stable system produces historical seasonal forecasts (i.e. hindcasts)
Operational context; no information from the future
Multi-model; 1 weather forecasting model (RPN) and 1 climate model (CCCma)
Basis for CMC operational seasonal forecasts
Connection to international programs
- SMIP2/HFP
- APCN/APCC

14. SMIP2
SMIP1 and initial SMIP2 were WGSIP projects to investigate “potential predictability” with imposed observed SSTs
SMIP2/HFP extended to actual skill, either
- 2 tier with forecast SST
- 1 tier coupled
SMIP2 analysis and results
- this workshop
- lessons for TFSP 1st experiment (CHFP)

15. CMC forecasts based on HFP2
Multi-model approach - 4 models
GCM2 T32L10
GCM3 T63L32
SEF T95L27
GEM 1.8oL50
HFP2
35 years (from 1969) … extend to 2006
10 member ensembles for 4 models = 40 member ensemble
4 month forecast made every month
12 rolling seasons

16. Some HFP2 results Deterministic and probabilistic forecasts
Multi-model and geographically distributed skill
Bias correction, combination, calibration, verification
Post-processing and added value

17. Deterministic quantitative forecasts
Bias removed – deal with anomalies of observations X and forecasts Y
How best to combine/scale results from 4 models?
Try in 5 ways for ensemble mean forecasts {Y}:
- YU => unweighted – grand mean of {Y}
- YM => scale each {Y} with its variance
- YT => scale each {Y} with total variance
- YL => scale Yu linearly to minimize MSE
- YR => multi-model regression on YM to minimize MSE

18. Deterministic forecasts
Skill measures
- correlation
- MSSS = 1- MSE/MSEo
Concentrate on DJF but available for all seasons
Based mainly on cross-validated results from S. Kahrin

19. Temp: Correlation/MSSS – globe
little difference in methods of combining model results … but
linear (1 parameter) scaling degrades correlation while increasing MSSS
multi-model (4 parameter) scaling degrades both
Corr

20. T Correlation/MSSS - Canada
similar result except scaling to reduce MSE
- degrades correlation - also degrades MSSS
- multi-model scaling is the worst
reiterates the difficulty in scaling when sample size small and skill modest
Corr

21. DJF Temperature Correlation MSSS MSSS- multi-model scaling
MSSS-linear scaling

22. Summary Reasonable skill for first season forecasts
Little difference based on methods of combining multi-model results
Differences in scaling multi-model results
- no scaling is OK
- linear scaling (1 constant) improves global MSSS (i.e. over ocean) but degrades correlation
- multi-model scaling (4 constants) degrades both MSSS and correlation
For HFP, multi-model information apparently adds skill:
- by averaging out noise (increased ensemble size)
- by simple averaging out of error in the signal

23. Deterministic 3-category forecast
next 3-month season (HFP1)
short lead (or 1 month)
disseminated via web
historical skill (percent correct) map welded to forecast map
Canada only (but results available for globe)
May-June-July

24. Temperature - Percent Correct
DJF
Little sensitivity to MM combination method
DJF

25. Probabilistic forecasts
next 3-month season
probabilities for each of 3 categories given in map form
reliability diagram attached

26. May-June-July

27. Brier Skill Score (BSS)
Where
● Three methods to estimate Pf (Kharin and Zwiers 2003):

28. Briar Skill Score (BSS)
BN=below
NN=normal
AN=above
C =count
G =gaussian
GA=gaussian
“adjusted”

29. BSS
DJF Temperature
Above
Below
Normal

30. Reliability
Above Normal Below
Globe
Canada

31. Summary for Probability Forecast
Simple gaussian fit better than counting
modest skill over Canada/Globe
reasonably reliable forecasts – but not sharp
“optimum” is gaussian fit over Canada(land) and adjusted over ocean
(see J.-S. Fontecilla poster for more HFP2 results)

32. Prediction/predictability studies
Stratospheric influence (QBO with K. Hamilton)
Bayesian applications
Diagnostic studies (see H. Lin’s poster)
GOAPP: coupled analysis/forecasting
- coupled processes
- predictability studies (from days to decades)
- ocean analysis
- coupled prediction (see Wm Merryfield’s poster)
- CHFP (=> TFSP project)

33. Diagnostics Other investigations based on HFP, HFP2 data
e.g. SST and PNA prediction and connection with tropical SST
Improved precipitation forecasts
See Hai Lin poster
Lin, Derome and Brunet

34. Initial coupled model NINO3
Initial coupled model NINO3.4 SST prediction attempts using the CCCma CGCM
El Nino
El Nino
Obs.
Obs.
Ensemble
Avg.
SST ANOM
Ensemble
Avg.
La Nina
La Nina
The pattern of the solar radiation of the AGCM3 is similar to that of the Era40. The maximum of the solar radiation is over the east Pacific because the convection over the west Pacific “warm pool” brings cloudiness over the area. On the equatorial east Pacific, there is a sinking motion and hence clear sky with more solar radiation.
The difference field shows model overestimates radiation over the east Pacific (~30 W/m^2) and ITCZ and over the south central Pacific and underestimates off the equator.
Obs.
Ensemble
Avg.
SST ANOM
Obs.
Ensemble
Avg.
Forecast range (months)
Forecast range (months)

35. Stratospheric aspects: the QBO
More or less annual reversal of winds in tropical stratosphere
One of the few long timescale processes in the atmosphere
QBO may affect extratropics by affecting propagation of large scale waves
Can a knowledge of the QBO add skill to extratropical seasonal forecasts?

36. QBO and QBO index
U(p,q)
Hamilton and Hsieh(2002) analyze equatorial zonal wind between 70-10hPa
obtain “circular non-linear principle component” representation of QBO
u(p,t) = U(p,q(t)) + u’
U(p, q) is a structure periodic in q which “optimally” represents the QBO
m=2
m=4
phase q(t)

37. QBO index q(t) q(t) is QBO index from average U(p, q) over 30-50hPa
filtered version of QBO
intended to capture QBO “essential structure”
long timescale make it easy to forecast

38. Observations and forecasts
Can knowledge of QBO provide additional skill in N winter?
For 4-model HFP2 ensemble mean DJF forecast Y and a single parameter adjustment
X = aY + W , a = XY/sY2
Nominal skill is
SHFP=rxy2

39. Skill SHFP of scaled HFP2 forecast
500
850 hPa Temperature
500 hPa geopotential

40. Additional skill
We ask if information in some index I can improve the forecast beyond HFP2
For X = aY + W, attempt to account for remaining variance W using index I
W = bI+ e, b=WI/sn2
For this approach skill is additive
S = SHFP + SI
skill from HFP2 + from index I
Try two indices: n = NINO3.4, q = QBO

41. Additional skill Sn from Nino3.4 Index
500
500 hPa geopotential
850 hPa Temperature

42. Additional skill Sq from QBO index q
500
500 hPa geopotential
850 hPa Temperature
850 hPa Temperature
500 hPa geopotential
Sq > 0.12 is nominally significant

43. Baldwin et al. 2001
Since the QBO is driven by “sub-grid” tropical waves it is missing in AGCMs
HFP2 models rapidly loose their initial conditions and their representation of the QBO

44. QBO and skill
HFP2 captures available skill from SSTs in extratropics (NINO3.4 doesn’t help)
Apparently some modest extratropical signal from the QBO lacking in HFP2 results
This additional skill, such as it is, is
- in N winter
- in North Atlantic region
- might interact with NAO
QBO skill could potentially be added statistically or by improving QBO in models

45. Coupled forecasting Follow the Sacred Principles GOAPP
- coupled processes
- predictability studies (from days to decades)
- ocean analysis
- coupled prediction
- CHFP (=> TFSP project)
Initial experiments (see Bill Merryfield’s poster)

46. GOAPP Theme I Theme II Analysis and mechanisms Ocean data assimilation
Predictability of the coupled system
Diagnostic
Prognostic
Coupled forecast initialization
Coupled historical forecasting project
Verification
Ocean data assimilation
Ocean Analysis and forecasting
Regional
Global
Role of eddies
Applications
Data assimilation
Coupling
Analysis methods
Modes of variability
Limits to predictability
Value of forecasts

47. Summary
Reasonably sophisticated objective 2-tier multi-model forecast system developed for Canada
Based on Sacred Principles we are undertaking:
- comprehensive analysis of HFP2 results
- optimum realistic correction, calibration, implementation
Developments via GOAPP of CHFP including decadal prediction and predictability

48. End of presentation