Fig. 4
From: Limits on determining the skill of North Atlantic Ocean decadal predictions
From Labrador Sea density drivers to large scale circulation. Correlation skill of years 2 to 5 volume averaged Labrador Sea top 500 m density (ρ500) in hindcasts assessed against ORAS4 (a) and DP3-assim (b) plotted against the regression slope between ρ500 and ρ T 500 (a) or ρ500 and ρ S 500 (b) in the control simulations for the same models. The hindcasts are separated into two groups based on a cut-off of r = 0.3 in year 2 to 5 ρ500 skill, when assessed against ORAS4. The symbol markers are as in Figs. 1 and 2 and the outlying hindcast system bcc-csm1-1 has been excluded from this analysis. The skill for just year 5 is also shown with the same but smaller symbols. The hindcast systems for which streamfunction data was provided are circled in panels a and b. The group average years 2 to 5 Atlantic overturning streamfunction skill assessed against ORAS4 (c, d), and DP3-assim (e, f), where the groups are as in panels a and b. Group 1: MIROC5, CNRM-CM5, FGOALS-g2, FGOALS-s2 and DePreSys3. Group 2: MPI-ESM-LR, GFDL-CM2p1, IPSL-CM5A-LR, MPI-ESM-MR and MRI-CGCM3. The time mean streamfunctions in the verifying reanalysis are contoured every three Sverdrups (1 × 106 m3 s−1) in red (positive, bold at 15 Sverdrups) and blue (negative). A total of 10 hindcast systems provided streamfunction data and every possible 5-member combination of these is computed to assess the significance of these group-average skill estimates. The hatching signifies skill at the 80% level, i.e., the group mean skill falls in either the top or bottom 10% of possible values
