One of the assumptions of flood frequency analysis is that annual maximum flood peaks are independently and identically distributed. Recent work has shown there exist persistent climate modes that modulate regional climates over multiyear timescales around the globe. Such persistence raises the question whether annual maximum floods are indeed independently and identically distributed. This study revisits this assumption. Noting that a significant shift in Pacific and Indian Ocean sea surface temperatures and other atmospheric variables occurred in the mid-1940s, 41 flood records in New South Wales, Australia, were stratified into pre-1945 and post-1945 records. It was found that the two-parameter lognormal distribution adequately fitted the stratified samples, and in many cases the stratified distributions were significantly different. In fact, the ratio of the post-1945 to the pre-1945 20-year flood exceeded one for 37 of the 41 sites. The evidence that the flood probability model is climate dependent for the case study region is strong. This has implications for flood risk assessment requiring inter alia the need to distinguish between short- and long-term flood risk. In the presence of long-term climate persistence, traditional flood frequency analysis can at best only provide estimates of long-term or unconditional flood risk. The estimation of short-term flood risks will require improved mechanistic understanding of multidecadal climate variability and the development of stochastic models that explicitly recognize such secular variations.