Risk Classification and Claim Prediction: An Empirical Analysis

from Vehicle Damage Insurance in Taiwan

Chu-Shiu Li
Department of Risk Management and Insurance
National Kaohsiung First University of Science and Technology
Kaohsiung, Taiwan

Chwen-Chi Liu
Department of Risk Management and Insurance
Feng Chia University
Taichung, Taiwan

Sheng-Chang Peng＊
Department of Risk Management and Insurance
Ming Chuan University
Taipei, Taiwan

＊
Corresponding author, email: [email protected].
 Risk Classification and Claim Prediction: An Empirical Analysis

from Vehicle Damage Insurance in Taiwan

Abstract

By conducting prediction models by logistic regression, this paper uses a unique vehicle

insurance data set in Taiwan to examine whether rating characteristics are still effective

under a bonus-malus system and to investigate whether extra information can help

predict claim occurrences for vehicle damage insurance. The empirical results show

that all current rating characteristics for vehicle damage insurance are significant factors

to predict claim occurrence. Among all, claim coefficient, car age, and car model are

relative important information for risk classification. In addition, we consider some

extra information regarding both policyholder and automobile third-party liability

insurance coverage to predict claim occurrence. As stated by the accuracy of prediction

models, we find that claim record in the previous policy year is a useful information for

risk classification.

Key Words: risk classification, claim prediction, automobile insurance

1. Introduction

Risk classification is an essential task in the insurance field from both theoretical

and practical views. Unable to acquire complete information of their insureds, insurance

companies determine premium rates through categorizing based on available

information. Not only does risk classification closely relate to the efficiency of

insurance market equilibrium, but it also functions as a necessary process for insurers

to maintain solvency and fairness.

1
 In the case of automobile insurance, it is common for insurers to use a number of

a priori classification variables such as the main driver’s age, gender, and occupation,

and vehicle’s usage and type, to adequately and fairly differentiate risk levels among

policyholders (Lemaire, 1995). However, other characteristics such as driving behavior

and driving history are also important in pricing automobile insurance rates. To include

driving experience into risk classification, the bonus-malus system (BMS) or merit

rating system has been widely used for a long time. A BMS rewards policyholders

without filing any claims by providing a discount (or bonus) and penalizes

policyholders involved in one or more accidents by adding extra premium (or malus).

There are usually two effects when a BMS is adopted by insurers: (i) it may prompt

the insureds to drive carefully and reduce accident occurrences, and (ii) it may link

driving risk to premiums more adequately. Utilizing automobile insurance data before

and after Tunisia introduced a BMS to its vehicle insurance pricing system, Dionne and

Ghali (2005) find a reduced probability of making claims for policyholders who

remained with the same insurance company during the observed period. Moreno et al.

(2006) provide a theoretical model to prove a bonus-malus contract can help eliminate

fraud. Furthermore, a BMS might induce some policyholders not to report small claims

(Lemaire, 1977) or to accumulate small losses to file one claim (Li et al., 2012) in order

to avoid higher future premiums. Therefore, the effects of a BMS on both a

policyholder’s long-term driving behavior and subsequent claim behavior. For instance,

when an insured has made a claim due to a car accident in the previous year, this year,

he or she will drive more carefully to avoid filing a claims and paying higher premiums

in the following year so that he or she can get a discount on future premiums.

We attempt to study the efficiency of pricing characteristics for risk classification

in this paper through analyzing vehicle damage insurance data in Taiwan. In other
2
words, we would like to investigate which characteristics currently applied by insurers

significantly impact the occurrence of auto insurance claims. Especially, drivers’ claim

behavior might be affected by the BMS. Then it becomes a new issue that whether those

characteristics for risk classification are still operative. Furthermore, we also want to

examine whether some additional information which insurers are able to get from

insureds could improve risk classification for vehicle damage insurance.

In regard of examining the characteristics of risk classification, it is common to

test the relationships between a priori variables of risk classification and occurrence of

ex post claims in order to identify which variable can differentiate risk levels of

insureds. An effective risk classification variable, however, should have the ability to

predict claim occurrence. It would be a useful way to examine the characteristics of risk

classification from the results of prediction risk classification characteristics from

analyzing prediction results. In other words, through setting different models to predict

claim occurrence, we could investigate how different rating characteristics affect the

following policy year. Particularly, we could estimate the importance of each different

rating characteristic from comparing the accuracy of prediction models. As insurers

develop their risk classification method for automobile insurance, our results might

provide some ideas for them to underwrite policies or review their current pricing

strategy.

Through conducting prediction models by logistic regression, this paper uses a

vehicle insurance data set to examine whether rating characteristics are still effective

under a BMS and to investigate whether extra information can help predict claim

occurrences for vehicle damage insurance. Our empirical results show that all

characteristics in the current rating system are significant factors to predict claim

occurrence. Among all rating characteristics, claim coefficient, car age, and car model
3
are relative important information for risk classification. For checking useful factors to

progress in risk classification, we consider some additional information regarding both

policyholder and automobile third-party liability insurance coverage. As stated by the

prediction accuracy, claim record in the previous policy year is a useful information for

risk classification.

This paper is organized as follows. In Section 2, we review literature on risk

classification for automobile insurance. Then we introduce rating characteristics in

Taiwan’s vehicle damage insurance market in the next section. As for Section 4, we

describe our empirical data and methodology. Our estimated models and prediction

results will be presented in Section 5, and Section 6 concludes this paper.

2. Literature review

On the literature of risk classification, early researches focused on theoretical

approaches, the efficiency of risk classification and its influence on social benefits, such

as Hoy (1982), Crocker and Snow (1985, 1986, and 1992), Bond and Crocker (1991),

etc. Crocker and Snow (2000) later reviewed previous theoretical views and

summarized that risk classification can raise the efficiency of insurance markets under

information asymmetry in a condition determined by each market’s information system

and its insureds’ understanding of classification characteristics. Researches in this genre

do not target at analyzing particular insurance types but discussing issues of risk

classification based on dichotomy instead – studying cases of both perfect information

and information asymmetry and of both competitive and non-competitive markets.

As for risk classification of vehicle insurance, it refers to questions oriented to

practices. Each country’s automobile insurance system often conducts risk

classification through characteristics of vehicle – car usage, brand, and style, and
4
characteristics of policyholder – insured’s gender, age, and claim record (Lemaire,

1995). Between these two kinds of classification characteristics, there are little

argument on characteristics of vehicle and accordingly scant discussion. On the

contrary, there is plenty discussion on characteristics of policyholder. The main reason

behind this result is that European and American countries used to argue over whether

discrimination rates should be decided by gender, age, or race, such unalterable

characteristics because discrimination cases took place. Therefore, researchers often

investigate vehicle insurance data by analyzing examples applying gender and age as

classification variables. For example, Butler et al. (1988) argued that pricing rules using

gender to regulate discrimination rates is questionable. They applied actual data to

analyze car accident records for both male and female drivers, then compared the

differences between collected premiums and claim losses from insurance companies,

and found that insured US females were overcharged for their vehicle insurance.

Puelz and Kemmsies (1993) used data of three personal vehicle insurance policies

in Georgia, USA, including vehicle collision coverage, full coverage, and liability

coverage, to evaluate how gender and other demographic variables impact on premium

pricing. Their empirical research results showed that gender significantly affects

premium rates, yet its influential degree is relatively less than other variables such as

driving record, age, location, and vehicle type. Accordingly, it might be unnecessary

for supervising administration to spend much time making laws against insurance rate

pricing resorts to gender. Relying on various existing viewpoints on restricting this kind

of risk classification, Harrington and Doerpinghaus (1993) examined administrative

regulations of risk classification in automobile insurance market. They concluded their

research by pointing out that although the effects of administrative measures are

unclear; they indirectly lead to risk cost or claim control cost increases as they distort

5
claim incentives. As a result, both the supply and the coverage of individual insurance

might be impacted.

Until present day, scholars often propose related analyses of similar car insurance

pricing characteristics. A recent research conducted by Doerpinghaus et al. (2008)

applied the Closed Claim Survey data of year 1997 provided by the Insurance Research

Council (IRC) in USA to study relations between claim filers’ demographic

characteristics and claim payment of third party liability vehicle insurance policies. Via

three possible economic theories discussing diverse risk attitudes and differences and

discrimination on negotiation costs, this study explained how various demographic

characteristics result in different claim amounts and set empirical models to

demonstrate possible relations between demographic characteristics and claim amount.

The results indicated that while controlling other variables, female insureds receive less

claim payments than male ones and married insureds receive more claim payments than

single ones. The relation between age difference and claim payment is insignificant in

Doerpinghaus’ study, but there are other researches arguing over age. Brown et al.

(2007) mentioned previous argument in Canada around whether to include age as a

vehicle insurance pricing characteristic, and six among ten provinces refuse to do so.

Nevertheless, other documented studies which applied car accident data have found

age’s influence on car accident occurrence. For instance, Braver and Tempel (2004) and

Tefft (2008) identified higher accident tendencies for young and elder drivers. To put

their findings into a figure with car accident loss versus age, then the figure shows a

line close to a U shape. Such results respond to the rate regulations in practice in

Taiwan, which apply higher rate coefficients upon young and elder insureds.

In addition, several papers investigate the effectiveness of marriage as a vehicle

insurance risk classification characteristic because similar to the cases of gender and
6
age, there is dispute over whether taking marital status into account is a kind of

discrimination. Gardner and Marlett (2007) retrospected the history of US vehicle

insurance market to envision future trends on vehicle insurance coverage, rates, market

management, and related laws. On the subject of estimating rates according to one’s

marital status, the most common reason to support this application is marital status’

relativity, considering that married drivers are calmer and more responsible than

unmarried drivers. Meanwhile, opposing opinions emphasize that not every driver is

granted the right of marriage; for example, forty-five states in USA forbid same sex

marriage, so only heterosexual drivers are qualified for marital status rate discount.

Moreover, because less people orient to marital relationships and the (average) age at

first marriage rise at present day, the effects of marital status discount will become less

critical to insurance purchasers. While most insurance purchasers do not agree that

marital status should be an insurance rate pricing characteristic, the empirical analysis

by Doerpinghaus et al. (2008) has proved that marital status impacts on the total of

claim payment.

On the other hand, researches target on claim experience as a risk classification

characteristic do not take place until recent years. Even though there were early

practical cases of claim experience, most countries adopted this characteristic rather

late. Dionne and Ghali (2005) used the vehicle insurance data of Tunisia before and

after it adopted a BMS and found that the probability for insureds who stayed with the

same insurance company decreases. Also, Moreno et al. (2006) designed theoretical

models which indicated that BMSs can prevent insurance fraud. The research literature

above suggests that applying BMSs induce insureds to lower driving hazards, but BMSs

may change insureds’ claim behavior, encouraging bonus hunger (Lemaire, 1977).

Because of the linkage between premium rates and filed claim numbers, insureds may

7
hesitate to report small claims or accumulate small losses to file one claim before

policies mature (Li et al., 2013) in order to avoid increasing their claim totals.

Except for analyses on those common risk classification characteristics above,

there is also plenty discussion with other factors. Via statistical analysis, Kellison et al.

(2003) examined the relationships between policyholder’s credit record and policies

with claim record, and they identified that those with worse credit scores report greater

losses (including both loss frequency and loss degree), which further proved that credit

record provides information unavailable in the current underwriting system. A related

yet slightly different research by Miller and Smith (2003) analyzed six types of private

car insurance policies and insurance scores only applicable to this kind of vehicle

insurance. However, these two papers did not explain how credit record assists

insurance companies in risk evaluation. To bridge this gap, Brockett and Golden (2007)

first reviewed related literature before investigating the relations between credit record

and car insurance loss from their studying of biological, psychological, and behavioral

attributes and financial assumption of risk regarding these attributes. They concluded

that credit evaluation could be turned into useful underwriting information only when

individual biological and psychological differences are reflected upon the loss risks of

insured vehicles.

Besides, Bair et al. (2012) predicted car accident occurrence based on vehicle

maintenance record through data of compulsory vehicle liability insurance and the

unique maintenance record in Taiwan. They traced lower accidence occurrence

probability from insured cars which follow their maintenance schedules, but they did

not find significant effects on loss degree. Accordingly, Li et al. (2013) analyzed vehicle

insurance data in Taiwan, and they found a significantly less probability of filing vehicle

physical damage claims for insureds who purchase new cars along with vehicle physical
8
damage policies that bundle with high insured amount voluntary vehicle liability

policies. To put it in another way, the purchasing behavior of bundled insurance

coverage also generates useful risk evaluation information.

3. Rating Characteristics for Vehicle Damage Insurance in Taiwan

In Taiwan’s property insurance market, vehicle insurance is the primary business

source for insurance providers. Based on previous statistical data, the income from

automobile insurance premium accounts for approximately fifty per cent of all premium

income; meanwhile, vehicle insurance claims comprise sixty per cent of the claim total.

Judging from both revenue and expense aspects, steadily managing automobile

insurance business or not is critical to each insurance company’s development.

For a long time, the Taiwanese automobile insurance market was regulated, so

premium rates were calculated and announced by supervising administration. Officials

decided basic premium for each insurance policy and related characteristics which

determined discrimination rates. Under rate regulations, the pricing characteristics of

vehicle physical damage insurance resorted to characteristics of policyholder and

characteristics of insured vehicle, while characteristics of policyholder take insured’s

age and gender (gender-age coefficient) and claim record (claim coefficient) into

account. For detailed information on gender-age coefficient, please see Table 1. In

general, the coefficient for male is higher than female, and young drivers are also noted

with relatively higher coefficients. The claim coefficient is calculated based on the

claim record in previous three years estimated as cumulative claim points. As for

characteristics of insured vehicle, official rates consider insured vehicle’s usage, type,

age, brand, and style (manufacture coefficient). Therefore, the administration calculates

premium of vehicle physical damage insurance with this following equation:

9
 Premium ＝ Basic Premium × (Gender-age Coefficient + Claim Coefficient)

× Manufacture Coefficient

Following the global trend of rate liberalization, insurance companies in Taiwan

can determine their own vehicle insurance premium rates since April 2009. However,

a great number of insurance providers stick to the official rates announced by

administrative institutions.

[Insert Table 1]

4. Data and Methodology

Data

This paper uses a data set of private vehicle damage insurance policy and claim

information for the policy years between 2010 and 2012 from the Taiwan Insurance

Institute. The policy information includes demographic characteristics of the

policyholder (age, gender, and marital status), characteristics of the vehicle (car age,

car model, and exhaust), premium, and deductible type. The claim information includes

demographic characteristics of the claimed driver (age, gender, and marital status),

claim date, claim payment, and cause of accident.

The vehicle damage insurance policy covers accidents to the car, including

rollover, lightning, fire, explosion, damage from flying objects, and collision, and it

contains a deductible option. Policyholders can choose policies with or without a

deductible. There are two types of deductibles: increasing per-claim deductible

(3,000/5,000/7,000 New Taiwan Dollar) and straight deductible. In our data set, about

11 percent of policyholders purchase a vehicle damage insurance policy with a

deductible, and 81 percent of them choose increasing per-claim deductible. Since

10
different types of deductibles will affect whether an insured makes a claims for car

accidents, we exclude those policies with straight deductibles to avoid heterogeneity

bias from deductible types.

[Insert Table 2]

Our data are reorganized by policy year from 2010 to 2012. The number of

policies and claim ratio are shown in Table 2. Depending on claim information, all

proportions (claim ratios) of claimed policy numbers in all policies by policy year are

about 48 percent. Moreover, we also observe the claim coefficient, which can serve as

the long-term claim history. They are all negative in average for three policy years,

which shows that many policyholders had a discount for vehicle damage insurance

premium.

[Insert Figure 1]

As shown in Figure 1, the shares of new cars exceed 40 percent, and one-year

cars are about 20 percent. The shares of other car ages display a declining tendency.

Accordingly, we observe the claim history of both new car and old car policies, shown

in Table 2. New car policies have higher claim ratios than old car policies for three

policy years. Similarly, compared with renewed policies, there are relative high claim

ratios for new policies which include new car policies and those policies transferred

from other insurers. For insurance companies, renewed policyholders seem have lower

risk. The claim coefficient displays consistent results mentioned above.

Methodology

This paper introduces logistic regression to conduct prediction models. After

estimating coefficients of predictors, we proceed with estimation and holdout samples

for prediction models. Depending on our data set, we use samples in the 2010 and 2011

11
policy years to estimate prediction models and make in-sample predictions,

respectively. Subsequently, for out-sample predictions, samples in 2011 and 2012 are

respectively used (correspondingly applied). Therefore, we have two sub data sets of

2010-2011 and 2011-2012 policy years to make estimations and predictions to confirm

for predictive consistency. For examining the efficiency of prediction characteristics

for claim occurrence, the model is set as follow:

Pr(Claim＝1| X1 , X2 , X3)＝F(X1β1＋X2β2＋X3β3) (1)

where Claim is a binary variable for whether a claim was filed during the insurance

period; X1 are demographic characteristics of the policyholder (such as age, gender,

and marital status); X2 are characteristics of the vehicle (such as car age, exhaust, car

model, and domestic car); X3 are other variables (such as insured district, insurance

company, claim experience in the last policy year, and so on); β1, β2, and β3 are

estimated coefficient vectors.

Prediction accuracy by the logistic regression model can be examined by a

classification table. A comparison of the predicted probability with the cutpoint can

distinguish whether a predicted event will occur. If the probability of the predicted

event is greater than or equal to the cutpoint, this defines that the predicted event will

occur; otherwise it will not occur. The classification table applied in our study is shown

in Table 2. Of total claimed policies (A+B), A is the number of policies that the

predicted probability correctly forecast a claim filed during the policy period, and B is

the number of policies falsely predicted. The sensitivity (= A / (A+B)) is the percentage

of prediction accuracy of total claimed policies. Similarly, the specificity (= D / (C+D))

is the percentage of prediction accuracy of total non-claimed policies (C+D). Total

correct (= (A+D) / (E+F)) is the percentage of prediction accuracy of total policies. E

12
and F are numbers of claimed and non-claimed policies, respectively. The cutpoint is

usually 0.5, but this paper would like to use a sensitivity analysis to determine it through

the highest percentage of prediction accuracy from total policies.

[Insert Table 3]

5. Empirical Results

After controlling for characteristics of the insured and vehicle, we predict the claim

probability using prediction models and then determine whether a claim is made during

the policy period by the cutpoint which is determined by sensitivity analysis. We apply

the percentages of correct prediction to the estimation sample (in the estimated policy

year) and holdout sample (in the next policy year) in order to evaluate the accuracy of

the prediction models. Prediction results are divided into the 2010 and 2011 policy

years. We only display prediction results for the 2010 policy year, and others are shown

in Appendix.

Firstly, we examine individual effect of rating characteristics, and results for the

2010 policy year are shown in Table 4. From estimation results, all rating factors, which

include insured age, gender, claim coefficient, car age, exhaust, and car model, have

significant effects on claim occurrence via Wald test, shown in Model 1. Then we

proceed prediction for claim occurrence through individually excluding one of six

rating characteristics from basic model, Model 1, and the results are shown in Models

2-7. When claim coefficient is removed out of the prediction model, the ratio of total

correct in Model 3 is slightly smaller than it in Model 1 for estimation sample, and the

difference rises for holdout sample. We have similar findings for car age and car model.

These results demonstrate that claim coefficient, car age, and car model are relative

important among current rating characteristics. There are consistent results in the 2011
13
policy year, shown in Appendix A.

Secondly, we implement claim predictions for including additional information,

shown in Table 5. Model 1 is basic model which contains all factors in rating system.

Models 2-6 individually join additional information into basic model, including

insured's married status, claim filed in the last policy year, liability coverage for third-

party's body and property, and claim coefficient of liability coverage. In Model 3, we

find that the information of claim record in the previous policy year could improve

prediction accuracy. Even if we contain all additional information in the prediction

model (Model 6), the ratios of total correct for estimation and holdout sample are

similar with those in Model 3. This finding exhibits that claim history in the last policy

year is an important information for improving risk classification. Other additional

information, however, cannot have consistent and robust estimation results, and also

cannot improve prediction accuracy. We have consistent outcomes in the 2011 policy

year, shown in Appendix B.

According to significantly different claim ratios among new car policies, new

policies, and renewed policies, this paper also proceeds prediction analyses for three

sub-samples, shown in Table 5. In Panels A and B, there are similar prediction results.

The ratios of sensitivity (prediction accuracy of claimed policies) are all higher than 70

percent for estimation and holdout samples. On the contrary, the ratios of specificity

(prediction accuracy of non-claimed policies) are lower than 60 percent. As insurers

cannot get more claim information for new car policies and new policies, other

additional information cannot improve prediction accuracy as mentioned above. As for

renewed policies, the ratios of sensitivity (prediction accuracy of claimed policies) are

about 40 percent, and the ratios of specificity (prediction accuracy of non-claimed

policies) are higher than 80 percent. These prediction results of renewed policies are
14
greatly different compared with new policies or new car policies. From those ratios of

total correct in Models 2-6, we can have the same finding that claim history in the last

policy year is an important information for improving risk classification.

6. Conclusions

A driver’s past claim history has been considered as one of the most important

variable to predict the future number of claims. However, not only can a BMS more

adequately link driving risk to premiums, it can also prompt the insured to drive more

carefully and reduce accident occurrences. As policyholders might change their claim

behavior due to a BMS, it is interesting to examine whether rating characteristics are

still effective. In addition, this paper uses an automobile insurance data set to investigate

whether extra information can help predict claim occurrences and improve risk

classification for vehicle damage insurance. We conduct prediction models by logistic

regression. The percentages of correct prediction for estimation and holdout samples

are used to evaluate the accuracy of prediction models.

The empirical results show that all characteristics in the current rating system are

significant factors in prediction models, and claim coefficient, car age, and car model

are relative important information for risk classification. As stated by the highest

prediction accuracy, the predictor of claim record in the previous policy year is more

helpful than other extra information among all prediction models. From the sub-sample

of renewed policies, we have a consistent and robust result.

15
Reference

Andersson, H., 2005, “The Value of Safety as Revealed in the Swedish Car Market: An
Application of the Hedonic Pricing Approach,” Journal of Risk and Uncertainty,
30(3): 211-239.

Bair, Shyi-Tarn, Rachel J. Huang, and Kili C.Wang, 2012, “Can Vehicle Maintenance
Records Predict Automobile Accidents?” Journal of Risk and Insurance, 79(2):
567-584.

Bond, Eric W. and Keith J. Crocker, 1991, “Smoking, Skydiving and Knitting: The
Endogenous Categorization of Risks in Insurance Markets with Asymmetric
Information,” Journal of Political Economy, 99: 177-200.

Braver, E. R. and R. E. Trempel, 2004, “Are Older Drivers Actually at Higher Risk of
Involvement in Collisions Resulting in Deaths or Non-Fatal Injuries among Their
Passengers and Other Road Users?” Injury Prevention, 10: 27–32.

Brockett, P. L., and L. L. Golden, 2007, “Biological and Psychobehavioral Correlates

of Credit Scores and Automobile Insurance Losses: Toward an Explication of
Why Credit Scoring Works,” Journal of Risk and Insurance, 74(1): 23-63.

Brown, R. L., D. Charters, S. Gunz, and N. Haddow (2007), “Colliding Interests－Age

as an Automobile Insurance Rating Variable: Equitable Rate-Making or Unfair
Discrimination?” Journal of Business Ethics, 72: 103-114.

Butler, P., T. Butler, and L. Williams, 1988, “Sex-Divided Mileage, Accident, and
Insurance Data Show that Auto Insurers Overcharge Most Women,” Journal of
Insurance Regulation,” 6: 243-284, 373-416.

Crocker, Keith J. and Arthur Snow, 1985, “The Efficiency of Competitive Equilibria in
Insurance Markets with Asymmetric Information,” Journal of Public Economics,
26: 201-219.

Crocker, Keith J. and Arthur Snow, 1986, “The Efficiency Effects of Categorical
Discrimination in the Insurance Industry,” Journal of Political Economy, 94: 321-
344.

16
Crocker, Keith J. and Arthur Snow, 1992, “The Social Value of Hidden Information in
Adverse Selection Economies,” Journal of Public Economics, 48: 317-347.

Crocker, K. and A. Snow, 2000, “The Theory of Risk Classification,” in G. Dionne

(ed.), Handbook of Insurance, 245–276, London: Kluwer.

Dionne, G. and O. Ghali, 2005, “The (1992) Bonus-Malus System in Tunisia: An

Empirical Evaluation,” Journal of Risk and Insurance, 72: 609-633.

Doerpinghaus, H., J. Schmit and J. J-H. Yeh, 2008, “Age and Gender Effects on Auto
Liability Insurance Payouts,” Journal of Risk and Insurance, 75(3): 527-550.

Harrington, S. and H. Doerpinghaus, 1993, “The Economics and Politics of Automobile

Insurance Rate Classification,” Journal of Risk and Insurance, 60(1): 59-84.

Hoy, Michael, 1982, “Categorizing Risks in the Insurance Industry,” Quarterly Journal
of Economics, 97: 321-336.

Hoy, Michael, Chu-Shiu Li, Chwen-Chi Liu and Sheng-Chang Peng, 2012, “Risk
Classification, Rating Innovation, and Multiple Contracts in Automobile
Insurance Market,” 39th Seminar of the European Group of Risk and Insurance
Economists, Palma de Mallorca, Spain.

Iversen, H. and T. Rundmo, 2002, “Personality, Risky Driving and Accident

Involvement among Norwegian Drivers, Personality and Individual Differences,”
33: 1251-1263.

Kellison, B., P. Brockett, S.-H. Shin, and S. Li, 2003, “A Statistical Analysis of the
Relationship Between Credit History and Insurance Losses,” Bureau of Business
Research: McCombs School of Business, The University of Texas at Austin.

Lemaire, J., 1977, La soif du bonus, Astin Bulletin, 9: 181-190.

Lemaire, J., 1985, Automobile Insurance: Actuarial Models, Boston, MA: Kluwer
Academic Publishers.

Lemaire, J., 1995, Bonus-Malus Systems in Automobile Insurance, Boston, MA:

Kluwer Academic Publishers.

17
Li, Chu-Shiu, Chih-Hao Lin, Chwen-Chi Liu, and A. Woodside, 2012, “Dynamic
Pricing in Regulated Insurance Markets with Heterogeneous Insurers: Strategies
Nice versus Nasty for Customers,” Journal of Business Research, 65: 968-976.

Li, Chu-Shiu, Chwen-Chi Liu, and Sheng-Chang Peng, 2012, “Is Claim Experience a
Good Factor to Predict Automobile Insurance Claim Occurrences?” Asia-Pacific
Risk and Insurance Association 16th Annual Conference, Seoul, South Korea.

Li, Chu-Shiu, Chwen-Chi Liu, and Sheng-Chang Peng, 2013, “The Expiration Date
Effects of Automobile Insurance Contracts: The Curious Case of Last Policy
Month Claims in Taiwan,” Geneva Risk and Insurance Review, 38: 23-47.

Li, Chu-Shiu, Chwen-Chi Liu, and Sheng-Chang Peng, 2013, “Bundled Automobile
Insurance Coverage and Accidents,” Accident Analysis and Prevention, 50: 64-
72.

Li, Chu-Shiu, Chwen-Chi Liu, and Jia-Hsing Yeh, 2007, “The Incentive Effects of
Increasing Per-Claim Deductible Contracts in Automobile Insurance,” Journal of
Risk and Insurance, 74: 441-459.

Miller, M., and R. A. Smith, 2003, “The Relationship of Credit-based Insurance Scores
to Private Passenger Automobile Insurance Loss Propensity,” Actuarial Study,
Epic Actuaries online at http://www.epicactuaries.com.
Moreno, I., F Vázquez, and R. Watt, 2006, “Can Bonus-Malus Allieviate Insurance
Fraud?” Journal of Risk and Insurance, 73: 123-151.

Puelz, R., and W. Kemmsies, 1993, “Implications for Unisex Statutes and Risk-pooling:
The Costs of Gender and Underwriting Attributes in the Automobile Insurance
Market,” Journal of Regulatory Economics, 5(3): 289-301.

Tefft, Brian C., 2008, “Risks Older Drivers Pose to Themselves and to Other Road
Users,” Journal of Safety Research, 39(6): 577-582.

Wang, Jennifer L., Ching-Fan Chung, and Larry Y. Tzeng, 2008 “An Empirical
Analysis of the Effects of Increasing Deductibles on Moral Hazard,” Journal of
Risk and Insurance, 75: 551-566.

18
Table 1. Gender-age coefficients
Age Male Female
Under 20 1.89 1.70
20 or above but under 25 1.74 1.57
25 or above but under 30 1.15 1.04
30 or above but under 60 1.00 0.90
60 or above but under 70 1.07 0.96
70 or above 1.07 0.96

Table 2. Claim history

Number of policies Claim ratio Claim coefficient
Policy year 2010 2011 2012 2010 2011 2012 2010 2011 2012
Total policies 135,082 141,809 130,932 48.09% 48.07% 48.70% -0.15 -0.17 -0.17

New cars 61,236 62,963 49,957 59.13% 58.61% 60.80% -0.01 -0.02 -0.03
Old cars 73,846 78,846 80,975 38.94% 39.64% 41.24% -0.27 -0.29 -0.27

New policies 79,112 86,611 76,146 55.97% 55.67% 55.04% -0.05 -0.06 -0.07
Renewed policies 55,970 55,198 54,786 36.96% 36.13% 39.88% -0.31 -0.33 -0.33
Note: Claim ratio is the proportion of claimed policy numbers in all policies.

Figure 1. Car age pattern of the whole sample for each policy yea

19
Table 3. Classification table
Predicted
Actual % Claim No Claim
Claim Sensitivity A B
No Claim Specificity C D
Total correct E F

Table 4. Examination results of individual effect of rating characteristics for the 2010 policy year
Predictor Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
Wald tests of individual effects (Wald Chi-Square and P-value)
Insured age 20.78 Y Y Y Y Y
(0.0004)
Male insured 17.53 Y Y Y Y Y
(<.0001)
Claim coefficient 1304.23 Y Y Y Y Y
(<.0001)
Car age 731.11 Y Y Y Y Y
(<.0001)
Exhaust 88.28 Y Y Y Y Y
(<.0001)
Car model 953.98 Y Y Y Y Y
(<.0001)
Others Deductible, Insured district, Insurance company
Classification Accuracies for Estimation and Holdout Samples (%)
Estimation sample
Sensitivity 64.00 63.90 63.90 63.60 64.20 64.20 64.20
Specificity 71.50 71.50 71.60 71.40 71.10 71.30 70.60
Total Correct 67.90 67.90 67.90 67.70 67.80 67.90 67.50
Observations 135,082
Holdout sample
Sensitivity 64.20 64.18 64.21 65.37 63.79 64.37 64.86
Specificity 67.78 67.77 67.73 65.76 67.74 67.57 66.47
Total Correct 66.06 66.05 66.04 65.57 65.84 66.03 65.69
Observations 141,809
Notes: The sensitivity is the percentage of prediction accuracy of the total claimed policies. The specificity is
the percentage of prediction accuracy of the total non-claimed policies. Total correct is the percentage of
prediction accuracy of the total policies.

20
Table 5. Estimation and prediction results of additional information for the 2010 policy year
Predictor Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Logit Predictor Coefficients for Estimation Sample
Basic risk factors Y Y Y Y Y Y

Married insured 0.043 *** 0.041 **

(0.016) (0.016)
Claim filed in last 0.487 *** 0.486 ***
policy year (0.016) (0.016)
Body liability -0.003 -0.004 *
coverage (0.002) (0.002)
Property liability 0.127 *** -0.071 **
coverage (0.032) (0.035)
Claim coefficient of -0.181 *** -0.072
liability coverage (0.058) (0.058)
Classification Accuracies for Estimation and Holdout Samples (%)
Estimation sample
Sensitivity 64.00 64.00 65.40 63.90 64.40 65.80
Specificity 71.50 71.50 71.30 71.50 71.20 71.10
Total Correct 67.90 67.90 68.50 67.90 67.90 68.50
Observations 135,082
Holdout sample
Sensitivity 64.20 64.24 65.44 64.25 63.44 64.42
Specificity 67.78 67.74 67.55 67.70 68.77 68.68
Total Correct 66.06 66.06 66.53 66.04 66.20 66.63
Observations 141,809
Notes: ***, **, and * denote statistical significance at the 1, 5, and 10 percent levels. Basic risk factors include
a limited number of a priori classification variables for calculating premiums. The sensitivity is the percentage
of prediction accuracy of the total claimed policies. The specificity is the percentage of prediction accuracy of
the total non-claimed policies. Total correct is the percentage of prediction accuracy of the total policies.

21
Table 6. Prediction results of additional information for different sub-samples for the 2010 policy year
Variable Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Basic risk factors Y Y Y Y Y Y Y Y Y Y Y Y
Married insured Y Y Y Y
Claim filed in last policy year Y Y Y Y
Body liability coverage Y Y Y Y
Property liability coverage Y Y Y Y
Claim coefficient of liability coverage Y Y Y Y
Estimation sample Holdout sample
Panel A: New car policies
Sensitivity 77.90 78.00 78.00 78.30 78.30 74.31 74.35 74.38 73.47 73.37
Specificity 55.10 55.30 55.20 54.50 54.50 53.45 53.37 53.35 55.04 55.27
Total Correct 68.60 68.70 68.70 68.70 68.70 65.68 65.67 65.68 65.84 65.87
Observations 61,236 62,963
Panel B: New policies
Sensitivity 77.00 77.00 77.00 77.60 78.00 71.38 71.39 71.39 70.33 70.25
Specificity 56.10 56.10 56.10 55.30 55.50 54.40 54.39 54.44 56.24 56.54
Total Correct 67.80 67.80 67.80 67.90 68.20 63.85 63.85 63.87 64.08 64.18
Observations 79,112 86,611
Panel C: Renewed policies
Sensitivity 40.50 40.50 41.00 40.70 41.00 41.50 43.15 43.23 44.83 43.50 43.29 44.71
Specificity 84.80 84.80 85.50 84.60 84.40 85.20 85.04 85.04 85.66 84.83 84.93 85.52
Total Correct 68.40 68.40 69.10 68.40 68.20 68.90 69.90 69.93 70.90 69.90 69.88 70.78
Observations 55,970 55,198
Notes: *** denotes statistical significance at the 1 percent level. Basic risk factors include a limited number of a priori classification variables for calculating premiums. The
sensitivity is the percentage of prediction accuracy of the total claimed policies. The specificity is the percentage of prediction accuracy of the total non-claimed policies. Total
correct is the percentage of prediction accuracy of the total policies.
22
Appendix A. Examination results of individual effect of rating characteristics for the 2011 policy year
Predictor Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
Wald tests of individual effects (Wald Chi-Square and P-value)
Insured age 24.92 Y Y Y Y Y
(<.0001)
Male insured 42.83 Y Y Y Y Y
(<.0001)
Claim coefficient 1079.57 Y Y Y Y Y
(<.0001)
Car age 1121.49 Y Y Y Y Y
(<.0001)
Exhaust 100.47 Y Y Y Y Y
(<.0001)
Car model 1548.85 Y Y Y Y Y
(<.0001)
Others Deductible, Insured district, Insurance company
Classification Accuracies for Estimation and Holdout Samples (%)
Estimation sample
Sensitivity 59.40 59.30 59.20 59.00 58.80 59.70 59.40
Specificity 73.40 73.50 73.50 73.20 73.30 73.00 71.80
Total Correct 66.70 66.60 66.60 66.30 66.40 66.60 65.90
Observations 141,809
Holdout sample
Sensitivity 52.01 52.00 51.96 52.07 52.17 51.89 52.42
Specificity 65.40 65.34 65.45 64.04 64.81 65.57 63.94
Total Correct 58.88 58.85 58.88 58.21 58.66 58.91 58.33
Observations 130,932
Notes: The sensitivity is the percentage of prediction accuracy of the total claimed policies. The specificity is
the percentage of prediction accuracy of the total non-claimed policies. Total correct is the percentage of
prediction accuracy of the total policies.

23
Appendix B. Estimation and prediction results of additional information for the 2011 policy year
Predictor Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Logit Predictor Coefficients for Estimation Sample
Basic risk factors Y Y Y Y Y Y

Married insured 0.107 *** 0.084 ***

(0.015) (0.016)
Claim filed in last 0.355 *** 0.397 ***
policy year (0.016) (0.016)
Body liability 0.036 *** 0.037 ***
coverage (0.001) (0.001)
Property liability -0.456 *** -0.569 ***
coverage (0.026) (0.030)
Claim coefficient of -0.089 0.039
liability coverage (0.057) (0.058)
Classification Accuracies for Estimation and Holdout Samples (%)
Estimation sample
Sensitivity 59.40 59.30 60.30 61.80 60.80 64.50
Specificity 73.40 73.50 73.50 72.20 72.60 71.40
Total Correct 66.70 66.60 67.20 67.20 66.80 68.00
Observations 141,809
Holdout sample
Sensitivity 52.01 52.08 53.06 58.51 51.24 58.64
Specificity 65.40 65.38 65.23 60.83 66.35 61.68
Total Correct 58.88 58.90 59.31 59.70 58.99 60.20
Observations 130,932
Notes: *** denotes statistical significance at the 1 percent level. Basic risk factors include a limited number of
a priori classification variables for calculating premiums. The sensitivity is the percentage of prediction
accuracy of the total claimed policies. The specificity is the percentage of prediction accuracy of the total non-
claimed policies. Total correct is the percentage of prediction accuracy of the total policies.

24
Appendix C. Prediction results of additional information for different sub-samples for the 2011 policy year
Variable Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Basic risk factors Y Y Y Y Y Y Y Y Y Y Y Y
Married insured Y Y Y Y
Claim filed in last policy year Y Y Y Y
Body liability coverage Y Y Y Y
Property liability coverage Y Y Y Y
Claim coefficient of liability coverage Y Y Y Y
Estimation sample Holdout sample
Panel A: New car policies
Sensitivity 74.70 77.70 78.80 78.20 80.30 61.80 66.84 69.85 64.46 68.61
Specificity 54.10 51.20 52.50 51.10 51.20 47.61 44.67 39.96 47.24 41.73
Total Correct 66.20 66.80 67.90 67.20 68.50 56.24 58.15 58.13 57.71 58.07
Observations 62,963 49,957
Panel B: New policies
Sensitivity 80.20 79.80 77.00 81.20 78.70 73.04 73.11 71.69 71.75 70.54
Specificity 45.70 46.30 51.60 44.60 50.10 45.10 45.09 43.46 46.95 45.39
Total Correct 64.90 65.00 65.80 65.30 66.30 60.49 60.51 59.00 60.61 59.23
Observations 86,611 76,146
Panel C: Renewed policies
Sensitivity 47.20 47.50 48.50 48.40 48.50 50.20 33.08 33.08 34.05 36.91 33.00 38.40
Specificity 83.40 83.20 84.90 82.50 82.60 83.40 82.56 82.50 83.75 79.42 82.51 80.27
Total Correct 70.30 70.30 71.70 70.20 70.10 71.30 62.82 62.79 63.93 62.46 62.76 63.57
Observations 55,198 54,786
Notes: Basic risk factors include a limited number of a priori classification variables for calculating premiums. The sensitivity is the percentage of prediction accuracy of the
total claimed policies. The specificity is the percentage of prediction accuracy of the total non-claimed policies. Total correct is the percentage of prediction accuracy of the
total policies.
25