Freakonometrics

A few days ago, a former student of mine, David, contacted me about Box-Cox tests in linear models. It made me look more carefully at the test, and I do not understand what is computed, to be honest. Let us start with something simple, like a linear simple regression, i.e.

Let us introduced - as suggested in Box & Cox (1964) - the following family of (power) transformations

on the variable of interest. Then assume that

As mentioned in Chapter 14 of Davidson & MacKinnon (1993) - in French - the log-likelihood of this model (assuming that observations are independent, with distribution ) can be written

We can then use profile-likelihood techniques (see here) to derive the optimal transformation.

This can be done in R extremely simply,

> library(MASS)
> boxcox(lm(dist~speed,data=cars),lambda=seq(0,1,by=.1))

we then get the following graph,

If we look at the code of the function, it is based on the QR decomposition of the matrix (since we assume that is a full-rank matrix). More precisely, where is a matrix, is a orthonornal matrix, and is a upper triangle matrix. It might be convenient to use this matrix since, for instance, .  Thus, we do have an upper triangle system of equations.

> X=lm(dist~speed,data=cars)$qr

The code used to get the previous graph is (more or less) the following,

> g=function(x,lambda){
+ y=NA
+ if(lambda!=0){y=(x^lambda-1)/lambda}
+ if(lambda==0){y=log(x)}
+ return(y)} 
> n=nrow(cars)
> X=lm(dist~speed,data=cars)$qr
> Y=cars$dist
> logv=function(lambda){
+ -n/2*log(sum(qr.resid(X, g(Y,lambda)/
+ exp(mean(log(Y)))^(lambda-1))^2))}
> L=seq(0,1,by=.05)
> LV=Vectorize(logv)(L)
> points(L,LV,pch=19,cex=.85,col="red")

As we can see (with those red dots) we can reproduce the R graph. But it might not be consistent with other techniques (and functions described above). For instance, we can plot the profile likelihood function,

> logv=function(lambda){
+ s=summary(lm(g(dist,lambda)~speed,
+ data=cars))$sigma
+ e=lm(g(dist,lambda)~speed,data=cars)$residuals
+ -n/2*log(2 * pi)-n*log(s)-.5/s^2*(sum(e^2))+
+ (lambda-1)*sum(log(Y))
+ }
> L=seq(0,1,by=.01)
> LV=Vectorize(logv)(L)
> plot(L,LV,type="l",ylab="")
> (maxf=optimize(logv,0:1,maximum=TRUE))
$maximum
[1] 0.430591
 
$objective
[1] -197.6966
 
> abline(v=maxf$maximum,lty=2)

The good point is that the optimal value of is the same as the one we got before. The only problem is that the -axis has a different scale. And using profile likelihood techniques to derive a confidence interval will give us different results (with a larger confidence interval than the one given by the standard function),

> ic=maxf$objective-qchisq(.95,1)
> #install.packages("rootSolve")
> library(rootSolve)
> f=function(x)(logv(x)-ic)
> (lower=uniroot(f, c(0,maxf$maximum))$root)
[1] 0.1383507
> (upper=uniroot(f, c(maxf$maximum,1))$root)
[1] 0.780573
> segments(lower,ic,upper,ic,lwd=2,col="red")

Actually, it possible to rewrite the log-likelihood as

(let us just get rid of the constant), where

Here, it becomes

> logv=function(lambda){
+ e=lm(g(dist,lambda)~speed,data=cars)$residuals
+ elY=(exp(mean(log(Y))))
+ -n/2*log(sum((e/elY^lambda)^2))
+ }
>
> L=seq(0,1,by=.01)
> LV=Vectorize(logv)(L)
> plot(L,LV,type="l",ylab="")
> optimize(logv,0:1,maximum=TRUE)
$maximum
[1] 0.430591
 
$objective
[1] -47.73436

with again the same optimal value for , and the same confidence interval, since the function is the same, up to some additive constant.

So we have been able to derive the optimal transformation according to Box-Cox transformation, but so far, the confidence interval is not the same (it might come from the fact that here we substituted an estimator to the unknown parameter .

A few years ago, I went to listen to Roger Nelsen who was giving a talk about copulas with fractal support. Roger is amazing when he gives a talk (I am also a huge fan of his books, and articles), and I really wanted to play with that concept (that he did publish later on, with Gregory Fredricks and José Antonio Rodriguez-Lallena). I did mention that idea in a paper, writen with Alessandro Juri, just to mention some cases where deriving fixed point theorems is not that simple (since the limit may not exist).

The idea in the initial article was to start with something quite simple, a the so-called transformation matrix, e.g.

Then the idea, then, is to consider , where  is the tensor product (also called Kronecker product) of with itself. Here, the support of is

> k=4
> M=matrix(c(1,0,0,1,
+            0,1,1,0,
+            0,1,1,0,
+            1,0,0,1),k,k)
> M
[,1] [,2] [,3] [,4]
[1,]    1    0    0    1
[2,]    0    1    1    0
[3,]    0    1    1    0
[4,]    1    0    0    1

> N=kronecker(M,M)
> N[,1:4]
[,1]  [,2] [,3] [,4]
[1,]     1    0    0    1
[2,]     0    1    1    0
[3,]     0    1    1    0
[4,]     1    0    0    1
[5,]     0    0    0    0
[6,]     0    0    0    0
[7,]     0    0    0    0
[8,]     0    0    0    0
[9,]     0    0    0    0
[10,]    0    0    0    0
[11,]    0    0    0    0
[12,]    0    0    0    0
[13,]    1    0    0    1
[14,]    0    1    1    0
[15,]    0    1    1    0
[16,]    1    0    0    1

> for(s in 1:3){N=kronecker(N,M)}

> image(N,col=c("white","blue"))

Consider our loss-ALAE dataset, and - as in Frees & Valdez (1998) - let us fit a parametric model, in order to price a reinsurance treaty. The dataset is the following,

> library(evd)
> data(lossalae)
> Z=lossalae
> X=Z[,1];Y=Z[,2]

The first step can be to estimate marginal distributions, independently. Here, we consider lognormal distributions for both components,

> Fempx=function(x) mean(X<=x)
> Fx=Vectorize(Fempx)
> u=exp(seq(2,15,by=.05))
> plot(u,Fx(u),log="x",type="l",
+ xlab="loss (log scale)")
> Lx=function(px) -sum(log(Vectorize(dlnorm)(
+ X,px[1],px[2])))
> opx=optim(c(1,5),fn=Lx)
> opx$par
[1] 9.373679 1.637499
> lines(u,Vectorize(plnorm)(u,opx$par[1],
+ opx$par[2]),col="red")

The fit here is quite good,

For the second component, we do the same,

> Fempy=function(x) mean(Y<=x)
> Fy=Vectorize(Fempy)
> u=exp(seq(2,15,by=.05))
> plot(u,Fy(u),log="x",type="l",
+ xlab="ALAE (log scale)")
> Ly=function(px) -sum(log(Vectorize(dlnorm)(
+ Y,px[1],px[2])))
> opy=optim(c(1.5,10),fn=Ly)
> opy$par
[1] 8.522452 1.429645
> lines(u,Vectorize(plnorm)(u,opy$par[1],
+ opy$par[2]),col="blue")

It is not as good as the fit obtained on losses, but it is not that bad,

Now, consider a multivariate model, with Gumbel copula. We've seen before that it worked well. But this time, consider the maximum likelihood estimator globally.

> Cop=function(u,v,a) exp(-((-log(u))^a+
+ (-log(v))^a)^(1/a))
> phi=function(t,a) (-log(t))^a
> cop=function(u,v,a) Cop(u,v,a)*(phi(u,a)+
+ phi(v,a))^(1/a-2)*(
+ a-1+(phi(u,a)+phi(v,a))^(1/a))*(phi(u,a-1)*
+ phi(v,a-1))/(u*v)
> L=function(p) {-sum(log(Vectorize(dlnorm)(
+ X,p[1],p[2])))-
+ sum(log(Vectorize(dlnorm)(Y,p[3],p[4])))-
+ sum(log(Vectorize(cop)(plnorm(X,p[1],p[2]),
+ plnorm(Y,p[3],p[4]),p[5])))}
> opz=optim(c(1.5,10,1.5,10,1.5),fn=L)
> opz$par
[1] 9.377219 1.671410 8.524221 1.428552 1.468238

Marginal parameters are (slightly) different from the one obtained independently,

> c(opx$par,opy$par)
[1] 9.373679 1.637499 8.522452 1.429645
> opz$par[1:4]
[1] 9.377219 1.671410 8.524221 1.428552

And the parameter of Gumbel copula is close to the one obtained with heuristic methods in class.

Now that we have a model, let us play with it, to price a reinsurance treaty. But first, let us see how to generate Gumbel copula... One idea can be to use the frailty approach, based on a stable frailty. And we can use Chambers et al (1976) to generate a stable distribution. So here is the algorithm to generate samples from Gumbel copula

> alpha=opz$par[5]
> invphi=function(t,a) exp(-t^(1/a))
> n=500
> x=matrix(rexp(2*n),n,2)
> angle=runif(n,0,pi)
> E=rexp(n)
> beta=1/alpha
> stable=sin((1-beta)*angle)^((1-beta)/beta)*
+ (sin(beta*angle))/(sin(angle))^(1/beta)/
+ (E^(alpha-1))
> U=invphi(x/stable,alpha)
> plot(U)

Here, we consider only 500 simulations,

Based on that copula simulation, we can then use marginal transformations to generate a pair, losses and allocated expenses,

> Xloss=qlnorm(U[,1],opz$par[1],opz$par[2])
> Xalae=qlnorm(U[,2],opz$par[3],opz$par[4])

In standard reinsurance treaties - see e.g. Clarke (1996) - allocated expenses are splited prorata capita between the insurance company, and the reinsurer. If  denotes losses, and  the allocated expenses, a standard excess treaty can be has payoff

where  denotes the (upper) limit, and  the insurer's retention. Using monte carlo simulation, it is then possible to estimate the pure premium of such a reinsurance treaty.

> L=100000
> R=50000
> Z=((Xloss-R)+(Xloss-R)/Xloss*Xalae)*
+ (R<=Xloss)*(Xloss<L)+
+ ((L-R)+(L-R)/R*Xalae)*(L<=Xloss)
> mean(Z)
[1] 12596.45

Now, play with it... it is possible to find a better fit, I guess...

Today, we will go further on the inference of copula functions. Some codes (and references) can be found on a previous post, on nonparametric estimators of copula densities (among other related things).  Consider (as before) the loss-ALAE dataset (since we've been working a lot on that dataset)

> library(MASS)
> library(evd)
> X=lossalae
> U=cbind(rank(X[,1])/(nrow(X)+1),rank(X[,2])/(nrow(X)+1))

The standard tool to plot nonparametric estimators of densities is to use multivariate kernels. We can look at the density using

> mat1=kde2d(U[,1],U[,2],n=35)
> persp(mat1$x,mat1$y,mat1$z,col="green",
+ shade=TRUE,theta=s*5,
+ xlab="",ylab="",zlab="",zlim=c(0,7))

or level curves (isodensity curves) with more detailed estimators (on grids with shorter steps)

> mat1=kde2d(U[,1],U[,2],n=101)
> image(mat1$x,mat1$y,mat1$z,col=
+ rev(heat.colors(100)),xlab="",ylab="")
> contour(mat1$x,mat1$y,mat1$z,add=
+ TRUE,levels = pretty(c(0,4), 11))

Kernels are nice, but we clearly observe some border bias, extremely strong in corners (the estimator is 1/4th of what it should be, see another post for more details). Instead of working on sample on the unit square, consider some transformed sample , where is a given function. E.g. a quantile function of an unbounded distribution, for instance the quantile function of the distribution. Then, we can estimate the density of the transformed sample, and using the inversion technique, derive an estimator of the density of the initial sample. Since the inverse of a (general) function is not that simple to compute, the code might be a bit slow. But it does work,

> gaussian.kernel.copula.surface <- function (u,v,n) {
+   s=seq(1/(n+1), length=n, by=1/(n+1))
+   mat=matrix(NA,nrow = n, ncol = n)
+ sur=kde2d(qnorm(u),qnorm(v),n=1000,
+ lims = c(-4, 4, -4, 4))
+ su<-sur$z
+ for (i in 1:n) {
+     for (j in 1:n) {
+ 	Xi<-round((qnorm(s[i])+4)*1000/8)+1;
+ 	Yj<-round((qnorm(s[j])+4)*1000/8)+1
+ 	mat[i,j]<-su[Xi,Yj]/(dnorm(qnorm(s[i]))*
+ 	dnorm(qnorm(s[j])))
+     }
+ }
+ return(list(x=s,y=s,z=data.matrix(mat)))
+ }

Here, we get

Note that it is possible to consider another transformation, e.g. the quantile function of a Student-t distribution.

> student.kernel.copula.surface =
+  function (u,v,n,d=4) {
+  s <- seq(1/(n+1), length=n, by=1/(n+1))
+  mat <- matrix(NA,nrow = n, ncol = n)
+ sur<-kde2d(qt(u,df=d),qt(v,df=d),n=5000,
+ lims = c(-8, 8, -8, 8))
+ su<-sur$z
+ for (i in 1:n) {
+     for (j in 1:n) {
+ 	Xi<-round((qt(s[i],df=d)+8)*5000/16)+1;
+ 	Yj<-round((qt(s[j],df=d)+8)*5000/16)+1
+ 	mat[i,j]<-su[Xi,Yj]/(dt(qt(s[i],df=d),df=d)*
+ 	dt(qt(s[j],df=d),df=d))
+     }
+ }
+ return(list(x=s,y=s,z=data.matrix(mat)))
+ }

Another strategy is to consider kernel that have precisely the unit interval as support. The idea is here to consider the product of Beta kernels, where parameters depend on the location

> beta.kernel.copula.surface=
+  function (u,v,bx=.025,by=.025,n) {
+  s <- seq(1/(n+1), length=n, by=1/(n+1))
+  mat <- matrix(0,nrow = n, ncol = n)
+ for (i in 1:n) {
+     a <- s[i]
+     for (j in 1:n) {
+     b <- s[j]
+ 	mat[i,j] <- sum(dbeta(a,u/bx,(1-u)/bx) *
+     dbeta(b,v/by,(1-v)/by)) / length(u)
+     }
+ }
+ return(list(x=s,y=s,z=data.matrix(mat)))
+ }

On those two graphs, we can clearly observe strong tail dependence in the upper (right) corner, that cannot be intuited using a standard kernel estimator...

An alternative to describe tail dependence can be found in the Ledford & Tawn (1996) for instance. The intuition behind can be found in Fischer & Klein (2007)). Assume that and have the same distribution. Now, if we assume that those variables are (strictly) independent,

> library(evd); 
> data(lossalae)
> X=lossalae
> U=rank(X[,1])/(nrow(X)+1)
> V=rank(X[,2])/(nrow(X)+1
> fL2emp=function(z) 2*log(mean(U<=z))/
+ log(mean((U<=z)&(V<=z)))-1
> fR2emp=function(z) 2*log(mean(U>=1-z))/
+ log(mean((U>=1-z)&(V>=1-z)))-1
> u=seq(.001,.5,by=.001)
> L=Vectorize(fL2emp)(u)
> R=Vectorize(fR2emp)(rev(u))
> plot(c(u,u+.5-u[1]),c(L,R),type="l",ylim=0:1,
+ xlab="LOWER TAIL      UPPER TAIL")
> abline(v=.5,col="grey")

> tau=cor(lossalae,method="kendall")[1,2]
> library(copula)
> paramgauss=sin(tau*pi/2)
> copgauss=normalCopula(paramgauss)
> Lgaussian=function(z) 2*log(z)/log(pCopula(c(z,z),
+ copgauss))-1
> Rgaussian=function(z) 2*log(1-z)/log(1-2*z+
+ pCopula(c(z,z),copgauss))-1
> u=seq(.001,.5,by=.001)
> Lgs=Vectorize(Lgaussian)(u)
> Rgs=Vectorize(Rgaussian)(1-rev(u))
> lines(c(u,u+.5-u[1]),c(Lgs,Rgs),col="red")

> paramgumbel=1/(1-tau)
> copgumbel=gumbelCopula(paramgumbel, dim = 2)
> Lgumbel=function(z) 2*log(z)/log(pCopula(c(z,z),
+ copgumbel))-1
> Rgumbel=function(z) 2*log(1-z)/log(1-2*z+
+ pCopula(c(z,z),copgumbel))-1
> Lgl=Vectorize(Lgumbel)(u)
> Rgl=Vectorize(Rgumbel)(1-rev(u))
> lines(c(u,u+.5-u[1]),c(Lgl,Rgl),col="blue")

Again, one should look more carefully at confidence bands, but is looks like Gumbel copula provides a good fit here.

As mentioned in the course on copulas, a nice tool to describe dependence it Kendall's cumulative function. Given a random pair with distribution  , define random variable . Then Kendall's cumulative function is

Genest and Rivest (1993) introduced that function to choose among Archimedean copulas (we'll get back to this point below).

From a computational point of view, computing such a function can be done as follows,

• for all , compute as the proportion of observation in the lower quadrant, with upper corner , i.e.

• then compute the cumulative distribution function of 's.

To visualize the construction of that cumulative distribution function, consider the following animation

Thus, here the code to compute simply that cumulative distribution function is

n=nrow(X)
i=rep(1:n,each=n)
j=rep(1:n,n)
S=((X[i,1]>X[j,1])&(X[i,2]>X[j,2]))
Z=tapply(S,i,sum)/(n-1)

The graph can be obtain either using

plot(ecdf(Z))

or

plot(sort(Z),(1:n)/n,type="s",col="red")

The interesting point is that for an Archimedean copula with generator , then Kendall's function is simply

If we're too lazy to do the maths, at least, it is possible to compute those functions numerically. For instance, for Clayton copula,

h=.001
phi=function(t){(t^(-alpha)-1)}
dphi=function(t){(phi(t+h)-phi(t-h))/2/h}
k=function(t){t-phi(t)/dphi(t)}
Kc=Vectorize(k)

Similarly, let us consider Gumbel copula,

phi=function(t){(-log(t))^(theta)}
dphi=function(t){(phi(t+h)-phi(t-h))/2/h}
k=function(t){t-phi(t)/dphi(t)}
Kg=Vectorize(k)

If we plot the empirical Kendall's function (obtained from the sample), with different theoretical ones, derived from Clayton copulas (on the left, in blue) or Gumbel copula (on the right, in purple), we have the following,

Note that the different curves were obtained when Clayton copula has Kendall's tau equal to 0, .1, .2, .3, ..., .9, 1, and similarly for Gumbel copula (so that Figures can be compared). The following table gives a correspondence, from Kendall's tau to the underlying parameter of a copula (for different families)

as well as Spearman's rho,

Consider a time series, generated using

set.seed(1)
E=rnorm(240)
X=rep(NA,240)
rho=0.8
X[1]=0
for(t in 2:240){X[t]=rho*X[t-1]+E[t]}

The idea is to assume that an autoregressive model can be considered, but we don't know the value of the parameter. More precisely, we can't choose if the parameter is either one (and the series is integrated), or some value strictly smaller than 1 (and the series is stationary). Based on past observations, the higher the autocorrelation, the lower the variance of the noise.

rhoest=0.9; H=260
u=241:(240+H)
P=X[240]*rhoest^(1:H)
s=sqrt(1/(sum((rhoest^(2*(1:300))))))*sd(X)

 
plot(1:240,X,xlab="",xlim=c(0,240+H),
ylim=c(-9.25,9),ylab="",type="l")
V=cumsum(rhoest^(2*(1:H)))*s
polygon(c(u,rev(u)),c(P+1.96*sqrt(V),
rev(P-1.96*sqrt(V))),col="yellow",border=NA)
polygon(c(u,rev(u)),c(P+1.64*sqrt(V),
rev(P-1.64*sqrt(V))),col="orange",border=NA)
lines(u,P,col="red")

Here, forecasts can be derived, with any kind of possible autoregressive coefficient, from 0.7 to 1. I.e. we can chose to model the time series either with a stationary, or an integrated series,

.