We seek to improve estimation of extreme sea state severities offshore Japan. In this tropical cyclone-dominated region, magnitudes of large values of storm severity (significant wave height, HS) observed at a location of interest are highly dependent on the trajectories of tropical cyclones relative to the location. As a result, a naive estimate for a return value of storm severity at a location of interest corresponding to a long return period, made using a relatively short period of observational or hindcast data, shows unrealistically large spatial variation. To address this issue, we propose a pragmatic statistical representation for cyclone sea state severity in space and time, consisting of (1) an extreme value model for the maximum of storm severity per cyclone (over space and time), and (2) a model for the “exposure” of a location to a random cyclone event. For a particular location, exposure quantifies the maximum storm severity observed during a cyclone event, expressed as a fraction of the storm peak severity (over space and time). Importantly, exposure is quantified per location on an absolute spatial lattice, independent of cyclone path relative to the location. Numerous statistical diagnostic tests are performed to justify that modelling assumptions made are consistent with data. Resulting return value estimates have plausible magnitudes and show plausible spatial variation, and in particular reflect bathymetric and shielding effects of coastlines and islands. © 2018 Elsevier Ltd