75ρ>0.75), which emphasizes the (expected) assymmetric contribution to waves at a certain location by different phases of a certain atmospheric pattern. This is also associated with larger PSS values, especially for coastal locations like Barcelona Erastin datasheet or Valencia (see Fig. 8), lower model biases ( Fig. 9), and smaller absolute RE values along the Catalan coast. However, the improvement in model performance from Setting 2 to Setting 3 is much smaller than that from Setting 1 to Setting

2, which is reflected in all skill measures. The next significant improvement is achieved by the inclusion of the lag-1 dependent variable Hs(t-1)Hs(t-1), i.e., the term ΔtΔt in Eq. (2), as a predictor to predict Hs(t)Hs(t) (Setting 4). The average ρ   score is now around 0.85, with values around 0.9 being seen at many locations ( Fig. 7). This is also associated with greater model skill (larger PSS values) and lower biases (FBI values to closer to unity; see

solid red curves in Fig. 8 and Fig. 9). The average RE (in absolute value) along the Catalan coast is 4.3% for the median HsHs, 14% for the 95th percentile, and 16% for the 99th Dabrafenib percentile, which is reasonably good in the context of HsHs prediction. Being the most complex model among the first group of model settings, Setting 5 includes the term ΔswΔsw as defined in Section 4.2 to further improve representation of swell waves. As summarized earlier, Setting 5 performs the best among Settings 1–5, although the improvement over Setting 4 is small in general. In fact, the small difference between the results of Settings 4 and 5 might be explained by the relatively Uroporphyrinogen III synthase short fetches of the study area and, consequently, the small impact of assuming no time lag δδ between the

origin of swell waves and their propagation to the point of interest as in Settings 3 and 4. In the open ocean where fetches are considerably larger, the difference might be more remarkable. Along the Catalan coast, the improvement of Setting 5 over Setting 4 is more noticeable. As shown in Fig. 8 and Fig. 9, Setting 5 is more skillful than Setting 4 in predicting smaller waves, although it is comparable to Setting 4 for predicting higher waves. Compared to Setting 4, the average absolute RE decreases by 4%, 55% and 50% for, respectively, the 50th, 90th, and 99th (see the dashed red curves in Fig. 10). Thus, we choose to focus on Setting 5 in the subsequent analysis. The second group of model settings (Settings 6–8) are compared in Fig. 11, Fig. 12, Fig. 13 and Fig. 14. They involve the same set of potential predictors as does Setting 5, but with a transformation applied to the wave heights HsHs and/or the squared SLP gradients G0G0, to explore the effect of transforming the non-negative data on model performance, as explained in Section 4.4.