Flood frequency analysis in urban watersheds is complicated by nonstationarities of annual peak records associated with land use change and evolving urban stormwater infrastructure. In this study, a framework for flood frequency analysis is developed based on the Generalized Additive Models for Location, Scale and Shape parameters (GAMLSS), a tool for modeling time series under nonstationary conditions. GAMLSS is applied to annual maximum peak discharge records for Little Sugar Creek, a highly urbanized watershed which drains the urban core of Charlotte, North Carolina. It is shown that GAMLSS is able to describe the variability in the mean and variance of the annual maximum peak discharge by modeling the parameters of the selected parametric distribution as a smooth function of time via cubic splines. Flood frequency analyses for Little Sugar Creek (at a drainage area of 110 km²) show that the maximum flow with a 0.01-annual probability (corresponding to 100-year flood peak under stationary conditions) over the 83-year record has ranged from a minimum unit discharge of 2.1 m³ s^{- 1} km^{- 2} to a maximum of 5.1 m³ s^{- 1} km^{- 2}. An alternative characterization can be made by examining the estimated return interval of the peak discharge that would have an annual exceedance probability of 0.01 under the assumption of stationarity (3.2 m³ s^{- 1} km^{- 2}). Under nonstationary conditions, alternative definitions of return period should be adapted. Under the GAMLSS model, the return interval of an annual peak discharge of 3.2 m³ s^{- 1} km^{- 2} ranges from a maximum value of more than 5000 years in 1957 to a minimum value of almost 8 years for the present time (2007). The GAMLSS framework is also used to examine the links between population trends and flood frequency, as well as trends in annual maximum rainfall. These analyses are used to examine evolving flood frequency over future decades. ?? 2009 Elsevier Ltd.