Extreme sea-level events represent potentially devastating hazards and are expected to occur more frequently in the future as a consequence of rising mean sea levels (Simpson et al., 2015). Additionally, and irrespective of changes in long-term mean sea level, decadal variability modulates and will persist to modulate extreme sea-level characteristics like return heights and periods, therefore intermittently exacerbating or dampening long term changes. Predicting these variations is greatly relevant for coastal management but has, to our knowledge, not yet been attempted. Recently, Smith et al. (2020) showed that decadal variability in North Atlantic winter climate is highly predictable. We propose to capitalize on these results and develop a framework for predicting decadal changes in sea- level heights based on large-scale climate variables.

 We will assess the predictability of decadal variations in sea-level heights along the coast of Norway by following the steps outlined below.

1) Identify the impact of decadal variability on extreme sea-level return heights by using high-frequency observations from tide gauges along the Norwegian coast (results available from CHEX).

2) Develop a tailored sea-level indicator that represents the decadal variability. Potential candidates are large- scale climate indices like NAO, indicators derived from SST and/or SLP fields or combinations thereof. Those will form the input to a statistical model to reproduce the observed decadal variability. We will consider simple models such as multi-linear regression as well as more advanced approaches such as machine learning. The results of 2) will be used to complement the results already produced in 1) to lead to a manuscript on the impact of decadal variability on extreme sea-level events around Norway.

3) Assess the predictability of the sea-level indicator. Based on the results in 2) we will use initialized decadal hindcast simulations available from the CMIP5/CMIP6 repository - of which the in-house prediction model NorCPM is a member - and evaluate the predictability of the indicator. Following Smith et al. (2020), we will use the multi-model ensemble to maximize the predictive skill by signal-to-noise filtering.