Nearest-neighbour resampling is a non-parametric method for generating scalar or vector-valued time series (U. Lall and A. Sharma, 1996: Water Resources Research, 32,679-693). It is used here for the joint simulation of daily rainfall and temperature at 25 stations in the German part of the Rhine basin. The main advantage of a non-parametric resampling technique is that it preserves the spatial association of daily rainfall over the drainage basin and the dependence between daily rainfall and temperature. To reproduce the autocorrelation structure of the data, a new day is resampled from the historical data by conditioning on the generated values for the previous day. Only the k nearest neighbours of the latter are considered for resampling.  Summary statistics of the daily precipitation and temperature fields have been used to find the nearest neighbours in the historical data. The choice of k turns out to be rather crucial for the reproduction of  properties of extreme multi-day rainfall. With regard to extreme river discharges in the Netherlands, large rainfall over periods of 4 to 20 days during the winter half of the year (October-March) is of interest. Besides for unconditional simulation of weather variables, nearest-neighbour resampling can also be used for conditional resampling of daily precipitation and temperature, given the atmospheric circulation. Daily circulation indices are available for the period 1881-1995. There is considerable low-frequency variability in the  atmospheric circulation during that period. Conditional simulation is used to explore the influence of this variability on extreme multi-day precipitation amounts. This work is supported by the Institute for Inland Water Management and Waste Water Treatment RIZA as part of a larger study to get a better insight into the likelihood of extreme river discharges in the Netherlands.

ADRI BUISHAND
Royal Netherlands Meteorological Institute
Postbus 201
3730 AE De Bilt, The Netherlands
buishand@knmi.nl
 
 

The last decade has seen major developments in the statistical methodology for extreme values of data. I will apply some of the most commonly known methods to a variety of climate data from three different continents. I will produce evidence to support both spatial and temporal change in climate extremes.

SARALEES NADARAJAH
School of Mathematical Sciences
University of Nottingham
University Park
Nottingham NG7 2RD, England
snh@pmn1.maths.nott.ac.uk
 
 

Much concern has been expressed (especially within the insurance industry) about increases in economic damage associated with extreme weather events. Questions remain about the extent to which these trends are attributable to changes in climate, as opposed to increased societal vulnerability. A case in point is the damage caused by North Atlantic tropical cyclones (primarily hurricanes) making landfall in the U.S. In an attempt to control for changes in vulnerability, this damage data set has been adjusted for inflation, wealth, and population at risk (Pielke and Landsea, 1998).
The basic stochastic model is a compound Poisson process, so that total annual damage is represented as a random sum. The total number of damaging cyclones per year is fit reasonably well by a Poisson distribution, and the monetary damage for individual storms by the lognormal. However, the extreme right-hand tail of the storm damage distribution appears to be heavy, with the estimated shape parameter of a generalized Pareto indicating that its variance is not necessarily finite.
Exploiting the random sum representation with a lognormal distribution of storm damage, only a relatively small fraction of the variation in annual damage totals is associated with fluctuations in the number of cyclones. Weak evidence is present of an increasing trend in the intensity parameter of the Poisson process governing the occurrence of cyclones, as well as somewhat stronger evidence of a decreasing trend in the mean of the lognormal distribution for storm damage (adjusted for societal vulnerability). Quite a bit stronger evidence exists of a dependence of these parameters on covariates such as the El Niño phenomenon. Implications of the apparently heavy-tailed distribution of storm damage are also explored, making use of its relationship to the distribution of maximum annual damage.

*The National Center for Atmospheric Research is sponsored by the National Science Foundation.

RICHARD W. KATZ
National Center for Atmospheric Research
P O Box 3000
Boulder, Colorado 80307, USA
rwk@ucar.edu
 
 

Wind storms have enormous economic and social importance. On one hand, expected extreme wind speeds are a main ingredient in building codes and hence have an important influence on cost of construction.  On the other hand, wind storms which exceed what has been taken into account at the building stage regularly lead to catastrophic losses, a recent example being the very severe damage caused by Hurricane Mitch in Central America.

This talk will survey a number of papers on statistical modelling and analysis of extreme winds. Topics include (i) site-by-site modelling of winds at 12 locations in Sweden, (ii) regression, Bayesian, and space
deformation techniques to utilize spatial dependencies for improved modelling, with application to extreme winds measured in Sweden and generated from a climatological model for the Gulf and Atlantic costs of the USA, and (iii) prediction of economic damage, and the relation between wind speeds and insurance loss, based on a 12 year data base from Sweden.

HOLGER ROOTZEN
Department of Mathematical Statistics
Chalmers University of Technology
S-412 96 Gothenburg, Sweden
rootzen@math.chalmers.se