Using R for Actuarial Science

1 Introduction 2

2 Some Features of R 2

3 SOA Tables Database 4

3.1 Binary Files of SOA DB 4
3.2 Implementation 6

4 Vectorization 9

5 Conclusion 11
Introduction 2

1 Introduction
The prestigious 1998 Association for Computing Machinery Award for software systems was
awarded to John Chambers of Bell Labs for the S system. The citation reads, For the S system,
which has forever altered how people analyze, visualize, and manipulate data. R is an open
source implementation of S and S-PLUS its commercial implementation.

R is an extensible, well documented language and environment with a core group of develop-
ers spread out over many countries. The size of its global user group and the diversity in its
applications are some of its many strengths. Aimed at actuaries and especially actuarial science
students who are pondering over the choice of a computing platform to complement MS Excel,
the insurance industry defacto standard, this article makes a case for R.

Designed for statistical computing and embraced by scores of statisticians, most if not all statis-
tical needs of an actuary should be ably served. Hence discussion of its statistical prowess will
be conspicuously absent. The next section will highlight some features of potential interest to
actuaries. Life contingent computations use mortality tables and most of the important tables
are found in the SOA tables database.The third section discusses an implementation of access to
this binary database with the intention of providing the readers with a good starting point for
computing on the life side. Vectorization, the subject of the penultimate section, is an important
concept for computing on R, like on APL and other interpreted vector languages . There we
discuss a vectorized solution for an important class of actuarial algorithms.

2 Some Features of R
First, R has an effective programming language with a simple syntax. The R Language Definition,
describes the syntax as having "a superficial similarity with C, but the semantics are of the FPL (func-
tional programming language) variety with stronger affinities with Lisp and APL.". A Brief History of S
by Richard A. Becker is a worth a read to know of the influences of other languages on the de-
velopment of S. For example, the following is taken from it; "... the basic interactive operation of
S, the parse/eval/print loop, was a well-explored concept, occurring in APL, Troll, LISP, and PPL, among
others. From APL we borrowed the concept of the multi-way array (although we did not make it our ba-
sic data structure), and the overall consistency of operations. The notion of a typeless language (with no
declarations) was also present in APL and PPL.".

Second, R is highly extensible and seamless extensions are carried out using packages. Impor-
tantly, C and FORTRAN code can be linked and called at run time from R not only enabling

N. D. Shyamal Kumar ASA, an assistant professor of Actuarial Science at the University of Iowa, is a member of
the computer science section council. He can be contacted at [email protected].
Some Features of R 3

use of existing code but also providing avenues for accelerating computationally intensive tasks.
And in the other direction, R objects can be manipulated in C. Talking about speed, support for
BLAS (Basic Linear Algebra Subroutines) is provided in R. Also, there is a package called snow
which implements a simple mechanism for computing on workstation clusters. For more details
consult Writing R Extensions.

Third, its reputed ability to produce high quality graphics with a minimal amount of code. The
base graphics system contains both high and low level commands. For a demo of its capabilities,
type demo(graphics) on the R command prompt. Especially important is its ability to generate
graphics for publication in many formats including postscript, pdf, jpeg and png. Also, there is
a separate graphics sub-system called grid which is considered to be more powerful and another
package called gridBase for combining grid and base graphics output. Below is an example
of the use of gridBase, similar to one in Murrell (2003), to combine the ease of a highlevel base
graphics system command to draw a barplot with the power of grid used here to rotate the x-axis
labels.

%age Distbn. of the PL's taken at Schools

− CompAct, Nov 99.
midpts <- barplot(c(39,30,20,16,13,12), axes = FALSE,

40
col=c("#7AF1FE","#FCCBFF","#F8DD9E","#9FC9FF",
"#00BFDD","#EB78E4"),ylim=c(0,40));
30

axis(2); axis(1, at = midpts, labels = FALSE);

vps<-baseViewports();
20

pushViewport(vps$inner, vps$figure,vps$plot);
grid.text(c("C/C++","Basic/VB","Java","SAS/S-Plus",
10

"Pascal","FORTRAN"),x=unit(midpts,"native"),
y=unit(-1,"lines"),just="right",rot=60);
popViewport(3);
0

N
va
+

/VB

lus

ca
C+

RA
Ja

s
−P
sic

Pa
C/

RT
S/S
Ba

FO
SA

And last but not the least, R has many utilities for accessing data. Access to data resident in a
Relational Database Management Systems is provided by several packages. Any system provid-
ing an ODBC interface can be accessed using RODBC. This list not only includes most important
RDBMSs but also databases like MS Access. Also, on Windows, ODBC drivers for text files, Excel
files and Dbase files are available. System specific packages are available for Oracle and MySQL.
R provides tools for accessing data in a binary format which is used in the implementation of
the next section. There is a package to parse XML files using either the DOM or the SAX mech-
anism. Of course, there is a rich support for reading text files. Also, there are packages which
make R a DCOM client (and server too) which help in accessing for example, live (financial) data.
R Data Import/Export is the relevant manual for details.

Since R is well documented, the above description has been kept brief. Moreover, the above
list is just a small subset of its wide ranging capabilities. Observe that some of the described
SOA Tables Database 4

features are not part of the core of R but of one of the many packages which seamlessly extend
it. Access to such features requires installation of relevant package(s), a matter of just a few
clicks. At this point, a tour of its official web site at http://www.r-project.org and a browse of
An Introduction to R would be a good idea. And if already convinced to try it out, go ahead
and install your free download!

3 SOA Tables Database

Here we discuss an implementation of access to the SOA tables database. To keep the discus-
sion self contained, the first sub-section describes the SOA tables database files (binary format)
tables.dat and tables.ndx. The next sub-section discusses the implementation along with
some examples of its use.

3.1 Binary Files of SOA DB

The database owes its existence to a joint project of the Technology Section of the SOA, the Inter-
national Section of the SOA and other volunteers. It consists of two binary files - tables.dat,
the data file, and tables.ndx, the index file. The description of the storage format for these files
can be found in the help of TableManager, a software copyrighted by Steve Strommen, FSA,
MAAA, and distributed freely by the SOA. The description is repeated below to keep the arti-
cle self contained. The Table Manager web page also has a MS Excel addin available for free
download for those interested in importing table values directly into Excel. A visit to the page is
highly recommended.

tables.ndx
The primary purpose of this file, as any index file, is to facilitate fast access to a table from the
tables.dat file. This is done by providing the offset, i.e. the number of bytes from the start
of the file, at which the data for a table begins. This is a binary file with a sequence of fixed
size records with an initial offset of 58 bytes (due to descriptive information). Each record has
five fields as depicted in the figure below. The character strings in this file, unlike those of the
tables.dat, are NULL byte terminated (as in C). The usage code is the same as that of the data
file.

4 Bytes 50 Bytes 4 Bytes 31 Bytes 1 Byte

32-bit Integer Character String 32-bit Integer Character String 8-bit Integer
Table Number Table Name Offset Country Usage Code
4 54 58 89 90
Figure 1 An Index File Record
SOA Tables Database 5

Record Types for tables.dat

Type Content Storage Format Possible Values
1 Table Name Character String
2 Table Number 32-bit Integer
3 Table Type Single Character • S - Select
• A - Aggregate by Age
• D - Aggregate by Duration
4 Contributor Character String
5 Source Character String
6 Volume Character String
7 Observation Peri- Character String
od
8 Unit of Observa- Character String
tion
9 Construction Character String
Method
10 Reference Character String
11 Comments Character String
12 Minimum Age 16-bit Integer
13 Maximum Age 16-bit Integer
14 Select Period 16-bit Integer
15 Maximum Select Age 16-bit Integer
16 No. of Decimal 16-bit Integer
Places
17 Table Values Sequence of 8-byte IEEE
floating types
18 Hash Value 32-bit Unsigned Integer
19 Country Character String
20 Usage 8-bit Integer • 0 - No Data
• 1 - Insured Mortality
• 2 - Annuitant Mortality
• 3 - Population Mortality
• 4 - Selection Factors
• 5 - Projection Scale
• 6 - Lapse Rates
• 99 - Miscellaneous

tables.dat
This binary file is a sequence of generic records with a single consecutive sub-sequence of such
records (terminating with a record of type 9999 with missing length and data fields) pertaining
to a single table. The storage format of a generic record is shown below.

Record Type Code - 16 bit Integer Length of the Data - 16 bit Integer Data - Variable Length
Figure 2 A Generic Record
SOA Tables Database 6

There are twenty record types as listed in the following inset. A caveat is that not all the record
types are relevant to a table and moreover not all the relevant types are necessarily present. The
latter becomes important as the data packaged as list may contain some empty components. The
character strings neither have a terminating NULL byte (like C) nor a leading byte containing its
length (like Pascal) - but note that the length can be read off the third and fourth bytes of the
generic record. R as a platform is convenient to read binary files as it has support for reading
IEEE floating point numbers.

The table values are stored in the ascending order of the index (Age or Duration) for types "A"
and "D". For type "S", the order is issue age wise in the ascending order of the select duration
until we hit the maximum issue age and thereafter in the ascending order of the age.

3.2 Implementation
RCode for TblSearch
"TblSearch" <-
function (Na = "", C = "", U = "", No = "")
{
rec <- function(...) {
c(readBin(z, integer(), size = 4), sub("", "", readChar(z,
50)), readBin(z, integer(), size = 4), sub("", "",
readChar(z, 31)), readBin(z, integer(), size = 1))
}
readChar(z <- file("tables.ndx", "rb"), 58)
x <- sapply(1:((file.info("tables.ndx")$size - 58)/90), rec)
close(z)
apply(matrix(c(Na, C, U, No, 2, 4, 5, 1), 4, 2), 1, function(y) {
x <<- x[, grep(y[1], x[as.integer(y[2]), ], perl = TRUE),
drop = FALSE]
})
data.frame(No = as.integer(x[1, ]), Name = x[2, ], Country = x[4,
], Usage = as.integer(x[5, ]), Offset = as.integer(x[3,
]))
}

The first function, TblSearch, provides a query facility for the file tables.ndx. The query has
to be in the form of a PERL regular expression and can be over each of the fields excepting offset.
The result is the set of records satisfying all of the individual constraints (intersection) packaged
as a dataframe object. This function not only helps a user search for tables in the database but
also is used by the function Tbl. The following are some comments on the code.
SOA Tables Database 7

i. A user may leave unspecified the trailing arguments which would then assume the default
null value..
ii. The function rec reads the five fields for each record and is used by the sapply function.
sapply, used to avoid an explicit loop, returns an array containing the fields for all of the
records in the file.
iii. Once the records are read, successively the regular expressions are used to filter the records
using the apply function.
iv. The last statement packages the array as a dataframe object.

Below are some examples of its use.

1. To list the US insured mortality age last tables for male smokers, one could use
TblSearch("(?i)[^e]male.*[^n]smoker.*last","US")

2. For US insured mortality age last basic tables of 1980, one could use
TblSearch("(?i)1980.*basic.*nearest","US","1")

3. To list just the names of all US annuitant mortality tables, one could use
TblSearch("","US","2")$Name

The second function, Tbl, given a vector of either table numbers or offsets returns a list of lists
containing all the fields for each table requested. If the input is a vector of table numbers then
TblSearch is used to get the offsets. The code could have been a lot shorter but for the need
to have this function vectorized - this way there is just a single read of the file for all the tables
combined. It does not return a dataframe like the earlier function because the mortality rates
data is of varying sizes. To facilitate recursive indexing of the list of lists, the following function
has proved useful. It is left uncommented as the logic is rather straightforward.

"RI" <-
function (z, Name)
{
sapply(1:length(z), function(x) {
z[[c(x, grep(paste("^", Name, "$", sep = ""), c("Name",
"Number", "Type", "Contributor", "Source", "Volume",
"ObsnPeriod", "ObsnUnit", "Method", "Reference",
"Comments", "MinAge", "MaxAge", "SelPeriod", "MaxSelAge",
"NumDec", "Rates", "HashValue", "Country", "Usage"),
perl = TRUE))]]
})
}
SOA Tables Database 8

RCode for Tbl

"Tbl" <-
function (Offset, Num = TRUE)
{
FT <- c(1, 2, array(1, 9), array(2, 5), 3, 2, 1, 2)
Table <- vector(length(Offset), mode = "list")
Len <- 0
z <- file("tables.dat", "rb")
if (Num) {
Offset <- TblSearch("", "", "", paste("^", as.character(Offset),
"$", sep = "", collapse = "|"))$Offset
}
Offset <- sort(Offset)
Offset <- Offset - c(0, Offset[-length(Offset)])
for (i in 1:length(Offset)) {
readChar(z, Offset[i] - Len)
Len <- 0
F <- 0
T <- vector(20, mode = "list")
attributes(T) <- list(names = c("Name", "Number", "Type",
"Contributor", "Source", "Volume", "ObsnPeriod",
"ObsnUnit", "Method", "Reference", "Comments", "MinAge",
"MaxAge", "SelPeriod", "MaxSelAge", "NumDec", "Rates",
"HashValue", "Country", "Usage"))
while (F != 9999) {
F <- readBin(z, integer(), size = 2)
Len <- Len + 2
if (F != 9999) {
l <- readBin(z, integer(), size = 2)
Len <- Len + 2 + l
T[[F]] <- switch(FT[F], readChar(z, l), readBin(z,
integer(), size = l), as.array(readBin(z, double(),
n = l/8, size = 8)))
}
}
Table[[i]] <- T
}
close(z)
Table
}
Vectorization 9

Below are some queries, similar to the ones above, using the additional fields from tables.dat.
Examples of life contingency calculations will be part of the next section.

1. To find the names of all select and ultimate US insured mortality tables, one could use
RI((y<-Tbl(TblSearch("","US","1")$Offset,FALSE))[sapply(1:length(y),function(x){
y[[x]]$Type})=="S"],"Name")

2. The following lists the names of insured aggregate mortality tables whose rates do not satisfy
the monotonicity beyond the age 35.
x<-(y<-Tbl(sort(TblSearch("","","1")$Offset),FALSE))[sapply(1:length(y),function(x){
y[[x]]$Type})=="A"]
RI(x[as.logical(1-sapply(x,function(z) {min((z$Rates[-length(z$Rates)]<=z$Rates[-1])
[-(1:(35-z$MinAge))])}))],"Name")

4 Vectorization
Most actuarial quantities on the life side satisfy a linear difference equation of the first order. On
compiled languages, a simple loop is the way to go and the direction in most actuarial problems
happens to be backward, in some forward and in the rest one could choose either one. This is
the way one computes on any spread sheet platform too - the loop becoming relative references
to the preceding or succeeding cell.

When working with an interpreted vector language, explicit loops are inefficient in the presence
of an algorithm which is able to vectorize the problem. It is not that vectorization removes the loop
but rather that it makes the loop implicit i.e. it is executed internally. In the actuarial literature,
see Shiu (1987), Shiu & Seah (1987) and Giles (1993), it has been noted that the usual closed form
solution of the linear difference equation translates to elegant (concise) APL code. More than
the brevity of the APL code that the closed form yields to a vectorized solution is important. In
other words, as in Shiu (1987), that the closed form solution can be translated to a code which
avoids explicit loops is the key. Hence this solution will lead to not only concise but an efficient
solution to the problem on all vector languages where explicit loops are inefficient. Below are
the details.

Life contingencies abounds with difference equations of the first order. Some are listed below.

1. The curtate future lifetime of a life aged (x) satisfies,

ex = px + px ex+1

2. The actuarial present value of a whole life insurance with benefits payable at the end of the
year of death satisfies,
Vectorization 10

Ax = νqx + νpx Ax+1

3. Fackler’s formula for the reserves of a fully discrete annual whole life on (x) satisfies,

n Vx = νqx+n − Px + νpx+n n+1 Vx

4. Hattendorff’s formula for the variance of the loss random variable underlying a fully discrete
annual whole life on (x) satisfies,

Var (n L|K(x) ≥ n) = ν2 qx+n px+n (1 − n+1 Vx )2 + ν2 px+n Var (n+1 L|K(x) ≥ n + 1)

All of the above are backward in nature as the boundary value at the right end is known. More-
over, they are each just a particular case of the equation

xn = an + bn xn+1 , n = 1, 2, . . . , k with xk+1 known

It is easy to show that the solution is given by,

xk+1 k0 bi + kl=n al l−1

Q P Q
0 bi
xn = Qn−1 , where b0 = 1 and n = 1, 2, . . . , k
0 bi

The R code for the above, which for matter of style has not been compressed into a single line, is
encapsulated by the function BD.

"BD" <-
function (a, b, bv)
{
s <- rev(cumprod(c(1, b)));
(rev(cumsum(s[-1] * rev(a))) + s[1] * bv)/rev(s[-1])
}

As an example, below is an R code for a plot of the curve of curtate life expectation for each of
the CSO 1980 basic age nearest tables. Note that the curve for female smokers is higher than that
of male non-smokers.
y<-c("#7AF1FE","#FCCBFF","#F8DD9E","#9FC9FF","#00BFDD","#EB78E4")
x<-Tbl(TblSearch("(?i)1980.*basic.*nearest","US","1")$Offset,FALSE)
matplot(25:100,sapply(x,function(z) {BD(a,a<-(1-z$Rates[-c(1:(25-z$MinAge))]),0)}),type="l",
lwd=2,ylab="",xlab="x",lty=1,col=y,main="Curtate Expectation of Life\n 1980 US CSO Age
Nearest")
legend(x=70,y=50,gsub("(1980 US CSO Basic|Age nearest)","",RI(x,"Name")),fill=y,
horiz=FALSE)
Conclusion 11

5 Conclusion
As some of the other interpreted
Curtate Expectation of Life vector languages, R is a fantastic
1980 US CSO Age Nearest
tool for rapid prototyping; be-
sides, it offers a large coherent
set of utilities. This combina-
50

Female tion of a prototyping language

Female Nonsmoker
Female Smoker
with an environment could be
40

Male particularly desirable by those

Male Nonsmoker who compute as part of their
Male Smoker
job while not being responsi-
30

ble for development of software

systems. In fact, R allows one
20

to smoothly move from a proto-

type to a compiled code as it al-
10

lows for one to access such code

from within R.
0

40 60 80 100

For an actuarial science program, a powerful case can be made for adoption of R as the standard
software across courses.

Acknowledgement
I thank Elias Shiu and Luke Tierney for valuable discussions. All opinions and errors are mine
alone. This article has been written using mainly ConTEXt (a TEX macro-package), Metapost,
PERL and R. I am greatly indebted to all involved in their development.

References
GILES, T. L. 1993. Life Insurance Application of Recursive Formulas. Journal of Actuarial Practice 1,
No. 2:141-151..

SHIU, E. S. W. 1987. Discussion of "Life Insurance Transformations" by Douglas A. Eckley. Trans-

action of the Society of Actuaries XXXIX, 17-18.

SHIU, E. S. W. and SEAH, E. 1987. Discussion of "Financial Accounting Standards No. 87: Recursion
Formulas and other Related Matters" by Barnet N. Berin and Eric P. Lofgren Douglas A.
Eckley. Transaction of the Society of Actuaries XXXIX, 37-40.