Statistical models are an alternative to numerical models for reconstructing storm surges at a low computational cost. These models directly link surges to metocean variables, i.e., predictors such as atmospheric pressure, wind and waves. Such reconstructions usually underestimate extreme surges. Here, we explore how to reduce biases on extremes using two methods—multiple linear regressions and neural networks—for surge reconstructions. Models with different configurations are tested at 14 long-term tide gauges in the North-East Atlantic. We found that (1) using the wind stress rather than the wind speed as predictor reduces the bias on extremes. (2) Adding the significant wave height as a predictor can reduce biases on extremes at a few locations tested. (3) Building on these statistical models, we show that atmospheric reanalyses likely underestimate extremes over the 19th century. Finally, it is demonstrated that neural networks can effectively predict extreme surges without wind information, but considering the atmospheric pressure input extracted over a sufficiently large area around a given station. This last point may offer new insights into air-sea interaction studies and wind stress parametrization.