## Product Launch: Actuarial Reserving Packages in R and Julia

Active Analytics Ltd: posted 26 August 2014 by Chibisi Chima-Okereke

### Introduction

Active Analytics is now completing the first release of an actuarial reserving package in R similar to our Reserving package in Julia. The development is done and we are testing, and writing the documentation. Both the R and Julia package are called Reserving.

This first version of the packages contain the Chain Ladder method. There are differences in the current versions of the packages in R and Julia. The Julia version is currently based on the 1993 Thomas Mack Chain Ladder paper and the R version is based on the 1999 version of the paper. We will update the Julia package in time. Please see our plan for future updates below.

#### Origin

We did not initially set out with the intention of creating an actuarial reserving package - we considered the chain ladder algorithm as something simple to play with programming languages. We've done this so much in this blog that we decided that we may as well write a package. Since we have created a package we intend to carry on adding functionality to increase its features in terms of calculation methods available.

Having learnt this lesson, we intend to carry on presenting other algorithms that will hopefully turn into packages in the future.

### Emphasis

The emphasis of the packages is on performance and content. All the components in the papers are calculated and the calculations in the R package are written in C++ using the Rcpp & RcppArmadillo packages. The calculations in the Julia package are largely written in C.

#### Performance

At Active Analytics the performance of our releases are important to us and the speed of the our chain ladder function rivals anything else out there commercial or otherwise. The documentation and benchmark of our Julia reserving package is given in our GitHub page. Here we benchmark our R Reserving chainladder function with the MackChainLadder function from the ChainLadder package:

# We only need the MackChainLadder function and a dataset from the ChainLadder package

# We use the microbenchmark function
mb = microbenchmark::microbenchmark

# We load our reserving package
require(Reserving)

Unit: microseconds
expr       min         lq     median         uq       max  neval
chainladder(dat)    37.346    41.1995    48.8075    69.3235   118.883   100
MackChainLadder(dat) 24165.159 25623.6585 26337.6920 27088.3740 34485.472   100

#### Content

Those already used to the ChainLadder package will find the form of our function familiar. The methods of execution of our chainladder function is given below. As with the MackChainLadder function the tail can be fully or partially specified

# Fully specify the tail factor
output = chainladder(dat, tail = 1.05, tail.se = 0.02, tail.sigma = 71)
# Specify tail factor f only
output2 = chainladder(dat, tail = 1.05)
# Specify that the default tail factor should be calculated using interpolation
output2 = chainladder(dat, tail = 1)
# No tail factor calculated

There is a wghts parameter that allows a weights matrix to be specified; the function also allows the $$\alpha$$ (alpha) parameter to be specified as in the paper.

We can take a closer look at the object returned, we used the same data set as in Mack's 1999 paper so feel free to check the outputs.

names(output)
[1] "f"        "f.se"     "sigma"    "Mack.S.E" "C"        "C.se"     "R"
[8] "F"        "F.se"

Firstly we have the development factors

output$f [1] 11.104259 4.092273 1.707913 1.275920 1.138912 1.068697 1.026335 [8] 1.022683 1.050000 Then we have the associated standard error output$f.se
[1] 2.24376318 0.51681801 0.12200144 0.05117008 0.04207692 0.02303354 0.01465199
[8] 0.01222874 0.02000000

Then $$\sigma$$ as defined in the paper:

output$sigma [1] 1336.96847 988.47643 440.13971 206.98511 164.19978 74.60176 35.49316 [8] 16.88652 71.00000 The standard error of the total reserve output$Mack.S.E
[1] 4053668

The completed claims triangle, if requested the last column is the tail factor derived ultimate ...

output$C [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [1,] 58046 127970 476599.0 1027692 1360489 1647310 1819179 1906852 1950105 2047610 [2,] 24492 141767 984288.0 2142656 2961978 3683940 4048898 4115760 4209118 4419573 [3,] 32848 274682 1522637.0 3203427 4445927 5158781 5342585 5483281 5607658 5888041 [4,] 21439 529828 2900301.0 4999019 6460112 6853904 7324745 7517641 7688163 8072571 [5,] 40397 763394 2920745.0 4989572 5648563 6433219 6875160 7056216 7216272 7577086 [6,] 90748 951994 4210640.0 5866482 7485161 8524943 9110579 9350505 9562602 10040732 [7,] 62096 868480 1954797.0 3338623 4259816 4851558 5184844 5321386 5442091 5714195 [8,] 24983 284441 1164010.4 1988029 2536565 2888926 3087386 3168691 3240567 3402595 [9,] 13121 145699 596240.1 1018326 1299303 1479792 1581449 1623096 1659913 1742908 and the associated standard error output$C.se
[,1]     [,2]     [,3]    [,4]      [,5]      [,6]      [,7]      [,8]       [,9]     [,10]
[1,]    0      0.0      0.0       0       0.0       0.0       0.0       0.0       0.00  106544.1
[2,]    0      0.0      0.0       0       0.0       0.0       0.0       0.0   60883.43  179976.6
[3,]    0      0.0      0.0       0       0.0       0.0       0.0  113393.5  139670.27  249707.6
[4,]    0      0.0      0.0       0       0.0       0.0  251132.9  295260.4  319019.65  417857.0
[5,]    0      0.0      0.0       0       0.0  456927.9  544256.1  575178.4  596210.29  670156.0
[6,]    0      0.0      0.0       0  584337.0  862501.4  967279.5 1007399.6 1037861.76 1127984.1
[7,]    0      0.0      0.0  659974  938781.2 1135845.3 1230032.2 1267288.3 1298251.31 1377496.2
[8,]    0      0.0 547296.5 1058013 1384868.1 1602335.3 1718387.5 1765323.0 1806031.70 1901740.3
[9,]    0 155949.8 745195.1 1319331 1697070.1 1942624.8 2078338.8 2133664.3 2182258.43 2293436.8

The reserves

output$R [1] 97505.25 303813.39 545456.06 1218667.38 1928522.57 4174250.14 3759398.29 [8] 3118154.01 1729787.46 The individual development factors (calculated from the triangle) ... output$F
[,1]     [,2]     [,3]     [,4]     [,5]     [,6]     [,7]     [,8]
[1,]  2.204631 3.724303 2.156303 1.323830 1.210822 1.104333 1.048194 1.022683
[2,]  5.788298 6.942998 2.176859 1.382386 1.243743 1.099067 1.016514 0.000000
[3,]  8.362214 5.543272 2.103868 1.387866 1.160339 1.035629 0.000000 0.000000
[4,] 24.713280 5.474043 1.723621 1.292276 1.060957 0.000000 0.000000 0.000000
[5,] 18.897294 3.825999 1.708322 1.132074 0.000000 0.000000 0.000000 0.000000
[6,] 10.490523 4.422969 1.393252 0.000000 0.000000 0.000000 0.000000 0.000000
[7,] 13.986086 2.250826 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000
[8,] 11.385382 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000

... and the associated standard errors:

output\$F.se
[,1]     [,2]      [,3]       [,4]       [,5]       [,6]       [,7]        [,8]       [,9]
[1,]  5.549259 2.763199 0.6375496 0.20417738 0.14077485 0.05812479 0.02631521 0.012228739 0.05084279
[2,]  8.542973 2.625299 0.4436387 0.14140439 0.09540731 0.03886806 0.01763909 0.008323676 0.03460691
[3,]  7.376778 1.886041 0.3566912 0.11564629 0.07787381 0.03284547 0.01535567 0.007211405 0.02998248
[4,]  9.131012 1.357997 0.2584455 0.09257564 0.06460298 0.02849575 0.01311440 0.006158846 0.02560631
[5,]  6.651914 1.131337 0.2575394 0.09266324 0.06908815 0.02941270 0.01353640 0.006357029 0.02643029
[6,]  4.438157 1.013093 0.2144947 0.08545751 0.06001668 0.02555073 0.01175903 0.005522333 0.02295992
[7,]  5.365242 1.060685 0.3148036 0.11328056 0.07955677 0.03386948 0.01558751 0.007320281 0.03043515
[8,]  8.458607 1.853405 0.4079549 0.14680059 0.10309784 0.04389155 0.02019990 0.009486373 0.03944100
[9,] 11.671796 2.589632 0.5700067 0.20511416 0.14405137 0.06132658 0.02822390 0.013254643 0.05510813

We hope that the next or future updates of the R and Julia Reserving packages will provide tools to carry out the Bornhuetter-Ferguson method as well as an update to the Chain Ladder functionality in Julia. It is possible that future versions of the package in Julia will be written in pure Julia depending on how things evolve, and every effort will be made to maintain the quality and performance of the functions. Any decisions or changes will be provided in the package documentation.

We hope that these packages prove valuable for your work and we are commited to their further development and release of new packages.

#### Release

The package is scheduled for release before the end of the working week - Friday 29th August 2014 at the latest and will be available for download at our GitHub page.