Data on fire losses in Hotels and Motels collected by the National Fire Incident Reporting System in the USA between 1983-1995 are statistically analysed. It is shown, using as illustration the year 1988 data, that the non-zero fire losses closely follow a lognormal distribution. For the years 1983-1995 the deflated mean and coefficient of variation, as well as the percentage of zero-loss fires, have remained practically constant. A new technique for estimating high quantiles of the distribution, based on recent work in the Theory of Extreme Value Distributions, is presented and used to obtain the 99% quantile of the fire loss for the year 1988. The result matches closely the estimate obtained by linear interpolation, but an estimate based on the lognormal distribution parameters derived from the global data seriously underestimates the percentile. A table of the coefficients required to calculate high quantiles for the whole period (1983-1995) is given.
[1]
F. Mosteller,et al.
Understanding robust and exploratory data analysis
,
1985
.
[2]
Dennis D. Boos,et al.
Using extreme value theory to estimate large percentiles
,
1984
.
[3]
G. Ramachandran,et al.
Extreme Value Theory and Large Fire Losses
,
1974,
ASTIN Bulletin.
[4]
Heinz Luck.
Economics of fire protection with fire detectors
,
1973
.
[5]
I. Weissman.
Estimation of Parameters and Large Quantiles Based on the k Largest Observations
,
1978
.
[6]
Milton C. Chew.
Distributions in Statistics: Continuous Univariate Distributions-1 and 2
,
1971
.
[7]
Theory & Methods: Estimation for Type II domain of attraction based on the W statistic
,
1999
.
[8]
A. M. Hasofer,et al.
A Test for Extreme Value Domain of Attraction
,
1992
.