# Using bootstrap to estimate the 95th percentile and confidence interval for skewed data

The problem:
I have data of sales per day during a certain period (n=7939). The data is rather skewed (see the first image below). I would like to propose the number of items to resupply every day such that for 95% of the days there is enough stock. It is given that the stock is resupplied every day and a big surplus of stock is considered waste. We can disregard seasonal influences.

Picture of the data: Approach:
Since I want to consider the number of items that would satisfiy 95% of the days, I simply took the 95th percentile of the sales. However, taking this percentile gives us no idea of the confidence interval. I felt there must be a more solid approach.

To resolve this, I applied bootstrap resampling and computed the 95th percentile of each sample. I performed this operation a thousand times. The idea is to make use of the central limit theorem. With normally distributed sample percentiles I can provide the mean and the 0.025 and 0.975 percentile of the sample percentiles as confidence interval.

However, the problem is that my data is very "binned" and I don't know where to go from here.

Basically I have two questions:

• Given my problem, is this a valid approach? Is there a different approach that might be more suitable for this problem?
• Is the binned distribution a problem for assuming normality (and thus provide the confidence around the mean 95th percentile). • There are bootstrap correction methods that take into account higher order moments in order to deal with skewness etc Sep 4 '18 at 6:44
• I noticed that my histogram is binned because of the wrong bin size. I used the approach KDG mentioned below to achieve the result. Sep 4 '18 at 10:54

Maybe you don't have to adopt normal distribution. Why don't you just use the 2.5% percentile and 97.5% percentile of boot strap sample percentiles as the confidence interval? I simulated usual bootstrap method and it seems work when comparing to the method using binomial distribution.

I don't have your data so I made some data from gamma distribution which is skewed.

#making data
set.seed(1)
x<-rgamma(7000,5,0.3)
hist(x)

x_sorted<-sort(x)

x_sorted[round(7000*0.95)] # estimate of 95% x


This is the bootstrap code I ran.

#method1 bootstrap
bootx95p(x_sorted,1000,0.05)

bootx95p<-function(x,b,alpha){
# b is the number of bootstraps.
# alpha is your type I error

n<-length(x)
p<-round(n*0.95)
xp<-rep(0,b)

for(i in 1:b){
x_boot<-sample(x,n,replace=TRUE)
x_boot<-sort(x_boot)
xp[i]<-x_boot[p]
}

xp<-sort(xp)
a<-round(alpha/2*(b+1))
CIup<-xp[b-a+1]
CIlow<-xp[a]

cat(' CI (',CIlow,', ',CIup,')','\n')
hist(xp)

}


The estimate would be 30.56664 and this is the result of the bootstrap method : CI ( 30.0623 , 31.08694 ) The below is the histogram of the distribution of 95th percentile of sample percentiles acquired from bootstrap method. And this is the method you also suggested using binomial distribution.

 #2 using bionomial
up=qbinom(0.975,7000,0.95)
low=qbinom(0.025,7000,0.95)
x_sorted[up]
x_sorted[low]


The result is quite similar :

> x_sorted[up]
 31.08901
> x_sorted[low]
 30.04189


As someone may have noticed from my English, I am not a native English speaker and even learning English. So It would be appreciated if someone correct my grammar.

• Hi KDG, thank you for your extensive reply. Actually the first part that you mention "Why don't you just use the 2.5% percentile and 97.5% percentile of boot strap sample percentiles as the confidence interval?" is exactly what I was doing. Your code helped me to figure out the binomial approach though, thanks for that! Sep 4 '18 at 10:52
• You're welcome. I also had a good time reviewing these methods. Thank you for having put this question.
– KDG
Sep 4 '18 at 12:04