Modelling hydrological data with and without long memory

The paper focuses on the problem of long-range dependence when analysing time series of hydrological data. Three time series are analysed: the monthly rainfall in the town of Florence, Italy; the daily minimum temperatures in the same town; and, finally, the daily water inflow to Lake Maggiore, Italy. Heuristic methods and maximum likelihood estimation of a parametric model are used to investigate the Hurst phenomena and to detect whether long-range dependence is present in any of the time series. We found that long-range dependence is not present in the first series but it is present in the last two. The daily water inflow to Lake Maggiore was modelled by a fractionally differenced ARIMA model (FARIMA) which contains a long-range dependence component. It is shown that the fit is much better than the one provided by more traditional ARIMA models that do not have such a component.