The procedure for estimating probable maximum loss (PML) for natural catastrophes has evolved over the past few decades from a rather simplistic deterministic basis to a more sophisticated methodology based on loss exceedance probability curves, generated using catastrophe modelling software. This development process is reviewed, with an emphasis on the earthquake peril, which, because of its widespread threat to critical industrial installations, has been at the forefront of most PML advances. The coherent risk definition of PML is advocated as an improvement over standard quantile methods, which can give rise to anomalous aggregation results failing to satisfy the fundamental axiom of subadditivity, and so discouraging the pooling of risks.