stepIC: Function for forward and backward model selection algorithm...

In MangoTheCat/autoPricing: This package is for carrying out automated glm pricing analysis for frequency and severity models

Description

Function for forward and backward model selection algorithm using information criteria to obtain frequency and severity models for actuarial pricing models the routine also dynamically adjusts the categories for variables that are not significant if an adequate mapping table is provided. This function is intended for analysing categorical explanatory varibles.

Usage

stepIC(ratingFact, countVar, sevVar, factLevels=1, timeVar,
    selType="BIC", consistThresh=60, theData, analysisType="frequency",
    myDistr="poisson", theLink="log", theAlg="forward",
    exposureName="Exposure", handicap=0, plotCharts=TRUE,
    myDocumentTitle="Automated Pricing GLM", ...)

Arguments

ratingFact	a character vector denoting the column headers of rating factors in your table (theData). Please be aware that the package autoPricing does not currently support interaction terms in your model.
countVar	a character string denoting the name of the claims count column
sevVar	a character string denoting the column header of the severity variable
factLevels	this is a list of matrices (or data.frames) that denote the mapping of the rating factors from their current categories to their logical aggregatable variables
timeVar	a character denoting the column header of the year variable. This variable must be specified because time consistency analysis is carried out by using interaction terms with the year variable.
IC	a character string denoting whether the information criterion used should be "AIC" (Akaike Information Criterion) or "BIC" (Bayesian Information Criterion)
consistThresh	Set this to the threshold for median consistency, denoting that the model fit of interaction of the variables with Year is consistent to consistThresh or greater before it is accepted
theData	this is the data set that will be used for the analysis containing the rating factors, exposure, severity, claim counts, and year
analysisType	flag indicating whether the analysis is for a "frequency" or "severity" model
myDistr	this is a character string denoting which distribution should be used in the analysis e.g. "poisson", "Gamma"
theLink	this is the link function to be used in the analysis
theAlg	this is the algorithm to be used either "forward" or "backward".
exposureName	this is a character string denoting the column name of the exposure (in Years)
handicap	extra penalty for the information criterion, it is added to the IC of the candiate model to alter how dificult it is to accept variables
myDocumentTitle	This is a character string for the title of the document
plotCharts	this is a logical variable as to whether the charts should be plotted

Value

The output is the frequency or severity model from the chosen algorithm theAlg

Note

please pay attention to the structure of the data requirements, since it is different from how actuarial data for GLM analysis is usually shaped and use the provided dataset "policyTable" as a guide for how the data ashould be formatted.

Author(s)

Chibisi Chima-Okereke [e-mail]

Examples

#Loading the data
data(policyTable)

#Preparing rating factor names and mapping tables 
myRatingFactors <- c("BonusMalus", "WeightClass", "Region", "Age", "Mileage", "Usage")
ratingFactorLevels <- lapply(myRatingFactors, function(x){matrix(as.character(levels(policyTable[,x])))})
names(ratingFactorLevels) <- myRatingFactors
ratingFactorLevels$Mileage <- cbind(ratingFactorLevels$Mileage, c("0-12500", "0-12500", "> 12500"))
ratingFactorLevels$BonusMalus <- cbind(ratingFactorLevels$BonusMalus, as.character(sort(rep(LETTERS[1:7], 2))))
weightClass <- c("650-935", "650-935", "650-935", "650-935", "1030-1315", "1030-1315", "1030-1315", "1030-1315", "1410-1600", "1410-1600", "1410-1600")
ratingFactorLevels$WeightClass <- cbind(ratingFactorLevels$WeightClass, weightClass)

#Example 1: Executing forward algorithm for poisson risk model
outputModelForwardFreq <- stepIC(ratingFact = myRatingFactors, countVar = "NoClaims", 
sevVar = "GrossIncurred", factLevels = ratingFactorLevels, timeVar = "Year", selType = "BIC", 
consistThresh = 60, theData = policyTable, analysisType = "frequency",
myDistr = "poisson", theLink = "log", exposureName = "Exposure",
handicap = 0, myDocumentTitle  = "Automated Pricing GLM", theAlg = "forward", 
plotCharts = TRUE)


# Example 2: Writing process to PDF and log file for documentation purposes
myFolder <- getwd()

pdf(file = file.path(myFolder, "GLMOutput.pdf"), height = 7, width = 11)
par(mfrow = c(1,1), cex.main = 1, cex.axis = .9, cex.lab = 1, cex = 1)
sink(file = file.path(myFolder, "GLMOutput.doc"))
outputModelForwardFreq <- stepIC(ratingFact = myRatingFactors, countVar = "NoClaims", 
sevVar = "GrossIncurred", factLevels = ratingFactorLevels, timeVar = "Year", selType = "BIC", 
consistThresh = 60, theData = policyTable, analysisType = "frequency",
myDistr = "poisson", theLink = "log", exposureName = "Exposure",
handicap = 0, myDocumentTitle  = "Automated Pricing GLM", theAlg = "forward", 
plotCharts = TRUE)
sink()
dev.off()

MangoTheCat/autoPricing documentation built on May 7, 2019, 2:09 p.m.