Fig. 2

The evolution of correlation skill and density drivers in the hindcasts. The correlation skill in the Labrador Sea top 500 m density (ρ500) between the hindcast systems and the reanalyses ORAS4 (a) and DP3-assim (b), as a function of lead time. A lead-time dependent bias correction (assessed against ORAS4 and DP3-assim separately) is applied to the hindcasts before calculating the skill. The regression slope between ρ500 and ρ _T 500 (c), or ρ500 and ρ _S 500 (d) in the hindcast systems as a function of lead time. For comparison, the same regression slope in ORAS4 (grey star) and DP3-assim (black cross) is displayed to the left of the axes with 90% confidence intervals (boxes), estimated via a bootstrap analysis. In addition, the same regression slope calculated from the control simulations (see Table 1) is shown to the right of the axes (further investigation of the stationarity of the control regression coefficients is shown in Supplementary Fig. 4). Finally, scatter plots of control simulation regression slope (x-axis, constant) against hindcast regression slope (y-axis) are inlaid as a function of lead time to highlight how the hindcasts revert to the nature described by the control simulations within a few years. The reanalyses are also shown with symbols coloured as before and can be seen to span a large portion of the hindcast behaviour. The one-to-one line is shown in green