vi

Contents
Abstract ................................................................................................................................. i

Declaration and Approval...................................................................................................... ii

Dedication ........................................................................................................................... iii

KEYWORDS ........................................................................................................................ iv

Acknowledgements ............................................................................................................... v

1 GENERAL INTRODUCTION ............................................................................................ 1

1.1 Study Background .............................................................................................................. 1
1.2 Problem statement .............................................................................................................. 2
1.3 Study Objective .................................................................................................................. 2
1.3.1 Main objective ............................................................................................................ 3
1.3.2 Specific Objectives ...................................................................................................... 3
1.4 Significance of study ........................................................................................................... 3
2 LITERATURE REVIEW .................................................................................................... 4

3 METHODOLOGY............................................................................................................ 7
3.1 Introduction ....................................................................................................................... 7
3.2 Data Scope ......................................................................................................................... 7
3.3 Generalized Linear Model( GLM) ........................................................................................ 7
3.3.1 component of Generalized linear model ........................................................................... 8
3.3.2 Assumptions of GLM ................................................................................................... 8
3.4 Tweedie distribution............................................................................................................ 9
3.5 XGBoost Model................................................................................................................. 11
4 DATA ANALYTIC AND INTERPRETATION ..................................................................... 14
4.1 Introduction ..................................................................................................................... 14
4.2 Generalized Linear Model................................................................................................. 15
4.3 Generalized linear model Tweedie distribution ................................................................. 15
4.4 Tweedie distribution with xgboost model ........................................................................... 15
4.5 Sensitivity analysis. ........................................................................................................... 16
4.6 Models comparison ........................................................................................................... 19
5 CONCLUSION AND RECOMMENDATIONS................................................................... 20
5.1 Conclusion ....................................................................................................................... 20
5.2 Recommendation .............................................................................................................. 20
Bibliography .............................................................................................................................. 21
Bibliography....................................................................................................................... 21
 1

1 GENERAL INTRODUCTION

1.1 Study Background

Insurance is a tool designed to protect the insured person from financial loss according
to the contract in force. Insurance penetration has been affected because of peoples pur-
chasing power and product design. The insurance penetration in Kenya for the year 2020
was 2.3%. Insurance penetration is the total written premium as a ratio of the country’s
gross domestic product (GDP). 56.39% of the share is owed to non-life policies and 43.61%
a reflection of the life policies Magati[8].

The pricing and distribution of the traditional insurance products has automatically out
priced a section of the population who would like to mitigate themselves from the insur-
able risks but lack the purchasing power. This phenomenon affects the lower-middle and
the lower income class citizens.More than one-third or 36% of Kenyans are exposed to nu-
merous risks such as economic shocks, death, illness and loss of property due to accidents,
natural disasters and calamities. This high incidence of shocks revealed by Mutia[10] un-
derpins the urgent need for risk management solutions for individuals, households and
businesses.

Formal risk management solutions such as insurance can be a potential solution. Yet, the
FinAccess survey also found that only two percent of Kenyans used insurance as a solu-
tion to deal with shocks. These findings highlight the huge gap between risk protection
needs and insurance outreach in Kenya. So, what is the reason for this huge gap? Digging
deeper, there is a huge disparity in insurance uptake among different income segments.
While there is reasonably high usage among the highest income segment (53 percent),
it is quite low among the lower-middle and middle income segments (16 per- cent and
28 percent respectively). These two income segments that can be identified as emerging
consumers (earning between Sh 20,000 to Sh55,000 per month) constitute the highest pro-
portion of Kenya’s population.

However, this segment has stayed away from insurance and opts for traditional, inad-
equate risk mitigation mechanisms such as social networks to meet spiraling medical,
school and funeral costs through harambees. But such traditional community ties are fast
weakening amid the urban bustle and many are wishing they had secured an insurance
 2

payout, rather than relying on the goodwill of others. This means that a huge chunk
of Kenyans remain unprotected against major shocks, while the insurance industry also
stays underdeveloped. Focusing on these emerging consumer groups is key to unlocking
insurance potential as they have growing incomes and constitute the highest proportion
of the population. But first, there is a need for public policy interventions to stimulate
the insurance sector. Among other initiatives, this can include conducive regulations and
government spending to insure vulnerable groups.

Then, and most importantly, the insurance industry itself has to expand its horizons by
not just focusing on a limited population of high, higher - middle income individuals. For
insurers, the sheer number of potential customers in the low -income bracket makes this
an attractive market. This is where micro-insurance comes on board. The low-income
segment can be served well by offering products whose coverage is lower in value than
the usual insurance plan and offers considerably smaller premiums that are attractive to
the bottom end of the market. This needs a shift in insurers’ mindset and a deliberate
application of technology and product innovation to offer attractive value propositions.

Unlocking market potential requires targeting uninsured people, particularly the low-
income households. The scarcity of data and even insufficient actuarial skills has led to
a challenge in the pricing of micro-insurance products. How can insurers ensure better
pricing in the micro-insurance market? This is the main focus of this research. Reducing
the price barriers to financial products such as this will greatly benefit the vulnerable
groups.

1.2 Problem statement

One of the key features of a micro-insurance product is the price. The low-income earners
have limited disposable income; they live from hand to mouth. For insurance companies
to tap into this uninsured market it needs to price the product appropriately: not too high
since it will not be taken up by customers; not too low that the company becomes insol-
vent after taking up the risk.

Generalized Linear Modelling is a popular pricing model yet it has its limitations: un-
able to incorporate variables with non-linear trends; unable to effectively model using
zero inflated data. Due to the shortcoming of using the GLM our study will employ a
more sophisticated model; Tweedie distribution in the XGboost Model to eliminate these
shortcomings.
 3

1.3 Study Objective

1.3.1 Main objective

The main objective of the study is to apply Tweedie distribution in the XGboost model in
pricing a micro-insurance product.

1.3.2 Specific Objectives

The following are specific objectives

1. To employ a Generalized Linear Model to the data set that has both linear and non-
linear parameters.

2. To apply a Generalized Linear Model with Tweedie distribution to the data set that
has both linear and non-linear parameters.

3. To test the same data set on the Tweedie distribution with the XGboost model.

4. To compare the goodness of fit by using the RMSE for the three models and compare
the outputs.

1.4 Significance of study

The study is importance since it will increase the efficiency, by use of a single step model,
and accuracy of risk premium computation in Micro-insurance which is a new trend in the
market. Sufficiency in computing the risk premium will in the long run cushion the risk
pool therefore the insurer will be able to conveniently meet its obligation. This will ensure
policyholder’s livelihood is sustained and the shareholder’s interest are safeguarded.
 4

2 LITERATURE REVIEW

Micro-insurance pricing is not highly researched like other fields.Scholars like Obuba(2014)[11],
Platteau(2017)[14], Eling(2014)[7],Brau(2011)[4] Biener(2013)[3], Pena(2019)[13] have dis-
cussed different angles of it.These authors have written an excellent overview on vari-
ous aspects of micro-insurance like Sanchez used some multiple regression in conjunc-
ture with the Tweedie model where the case he has used glm coefficients in his results.
Churchill(2002)[5] is interested in role of commercial insurers. In Maleika(2008)[9] sug-
gested to advice Social funds maybe directed to cater for micro-insurance to benefit poor
people.

In these papers they have touched on various aspects of micro-insurance which have not
been practically efficient. Akotey(2011)[2] explains how the demand for micro-insurance
vehicles is in modern markets of developing countries like Ghana. He figured out the
various factors that may influence the demand of micro-insurance for the sector markets.
The markets of developing countries is entrenched in poverty leading to an untapped
opportunity. There is low-level of income in those households and even that little that
is available is mismanaged. Due to insufficient insurance knowledge by the people may
lead to misunderstanding of the concept and therefore the opportunity to take up those
insurance covers is missed.He collected data from four major markets in Accra, Ghana
and analyzed it using the probit regression model to determine whether the people in the
informal sector would purchase insurance or not. All factors withstanding there is poten-
tial demand for micro-insurance service is high.

E. Ohlsson(2010)[12] have extensively expounded on the subject of Non-life Insurance

Pricing with Generalized Linear Models as stated. Generalized linear models are very
resent models . They are more appropriate for pricing compared to ordinary linear mod-
els in two ways. The first is, GLMs uses a variety of distributions like normal, Poisson,
Gamma e.t.c considering the characteristics of the data being analysed. This is contrary
to the constraint of the Ordinary Linear Models that assumes normality of data being
modelled. Secondly, the mean of linear models is a linear function of the explanatory
variables while for GLMs some monotone transformation of the mean is a linear function
′
of the xs with linear and multiplicative cases. Other advantages as noted by [12] include:
it has well established techniques for estimating statistical features like standard errors,
confidence intervals and testing model selection: it has software packages that are used
 5

to tariff analysis. In our case we will be using R for our analysis.

Gordon (2002)[15] wrote on fitting Insurance claims Data on Tweedie’s Compound Pois-
son Model this was in a bid to produce a fair and accurate tariff based on aggregated claims
insurance data. They used the Poisson distribution for the claims count and Gamma distri-
bution to model the claims amount data. They noted that it was important to use double
generalized linear model; in their case the number of claims and the claim amount; to
enable one to model one dispersion in the absence of the other. In practice, the explana-
tory variables will have different weightings on the dependent variable. Some variables
will impact the claims frequency more while other variables will affect the severity more
than it would the frequency. Since modelling the aggregated claims insurance data will
likely lead to a display of non-constant dispersion data; modelling the dispersion as well
as the mean to obtain an efficient estimation of µ. The dispersion parameter used in the
sub- model was ψ being that Tweedie models are scale invariant. The residual maximum
likelihood (REML)method was used to estimate the variance estimators though it has a
downward biases effect when the number of parameters used to estimate the fitted values
is large visa vi the sample size. ψ̂ is the maximum likelihood estimator. They concluded
that the Tweedie Compound Poisson method is efficient in analysing aggregated claims
data which is usually used to generate the pure risk premium. Therefore, more terms are
likely to be significant in the model in comparison with approximate methods or methods
used based on univariate likelihoods.

Obuba(2014)[11] in his 2014 paper has noted down factors that affect micro-insurance
product pricing which are:

• Data Availability

• Business acquisition expenses

• Administration cost

• Random stochasticity of the frequency and severity of claims.

• Lack of a suitable actuarial approach

Micro-insurance is a tool that protects the livelihood of low-income earners. This is a

alternative to the traditional insurance products. Features of this kind of product were
highlighted by Ahmed(2006) [1] as below:

• Targets low-income earners in the economy.

 6

• Product is usually bundled with other financial products

• The product is simple in nature.

• Short underwriting process with few requirements.

• The Sums Insured (assured) per policy are small.

• Short claims management turnaround time.

• Simple distribution channel

Guha-Khasnobis advised that the micro insurer should consider the pricing of these prod-
ucts with small periodic premiums. Deblon(2011)[6] has given incite on five risks that can
be considered for modeling micro- insurance products which are: Harvest failure; Third-
party liability; property damage; life-cycle risks and health risks. These risk factors are
individual and predictable hence good enough for them to be mitigated by having an in-
surance pool.

Sanchez(2019) [13] used an innovative approach by mixing two distributions - The com-
pound Poisson distribution and the Gamma distribution to come up with a single risk
premium model called Tweedie model. He noted that the properties of the data being
used will determine the choice of distributions. The model has evidenced its efficiency by
giving different risk premium depending on the group risk profile. This is contrary to the
inefficient burning cost method that is often used to give a flat premium not incorporating
the lapsed policies with a good risk profile or insureds with bad risk profile. The Tweedie
model leaves out the extra step of adding an expense loading which is usually part of
the burning cost model. The data used to price the micro-insurance product was from
both internal and external sources. This is particularly helpful because it covers for the
deficiency that actuaries encounter when modeling the price of a micro-insurance prod-
uct. He has recommended for the model to be replicated with data from different regions
and different micro-insurance products. Therefore, this model can be replicated with data
from the Kenyan market.
 7

3 METHODOLOGY

3.1 Introduction

In this chapter, we will present the methods for arriving at the earlier stated project ob-
jectives. It looks at the model and model development with a clear presentation of the
model assumptions. The chapter also presents the sources of data.

3.2 Data Scope

The developing and introduction of micro-insurance products into the market has led to
increase in market share. It has however proven impossible to get granulated data from
the respective insurance companies who are sighting data protection policies. Access to
data possess as the greatest huddle hence testing the model’s effectiveness. One particular
difficulty of predicting such claims is that there are many policyholders with no accidents
and consequently, no claims. The resulting distribution is therefore positively-skewed,
with a mass at 0. Such zero-inflated distributions may not fare well if loss functions such
as Gaussian distribution are used. We therefore use secondary data that have the following
variables:

• Age

• Gender

• Number of Children

• Smoking status

3.3 Generalized Linear Model( GLM)

Generalized linear models represent the class or family of regression models which mod-
els the response variable, Y , and the random error term (ε) based on exponential family
of distributions such as normal, Poisson, Gamma, Binomial. GLM assumes that the distri-
bution of the response variable is a member of the exponential family of distribution. This
is different from the general linear models (linear regression / ANOVA) where response
variable, Y , and the random error term (ε)have to be based solely on the normal distribu-
tion. Linear models can be expressed in terms of expected value of response variable as
 8

the following

g(µ) = β0 + ∑ βi Xi (1)
i=1

where µ can be expressed as E(Y ) but for generalized linear models g(µ) takes the link
function associated to the exponential family distribution i.e for Poisson model

log (Y ) = β0 + ∑ βi Xi (2)
i=1

3.3.1 component of Generalized linear model

GLM has three components namely

• Random Component - specifies the probability distribution of the response variable;

e.g., normal distribution for Y in the classical regression model, or binomial distribu-
tion for Y in the binary logistic regression model. This is the only random component
in the model; there is not a separate error term.

• Systematic Component - specifies the explanatory variables (x1 , x2 , · · · , xk ) in the

model, more specifically, their linear combination; e.g.,β0 + β1 x1 + · · · + βk xk , as we
have seen in a linear regression, and as we will see in the logistic regression in this
lesson.

• Link Function g(µ), specifies the link between the random and the systematic com-
ponents. It indicates how the expected value of the response relates to the linear com-
bination of explanatory variables; e.g, g(µ)=g(E(Yi ))=E(Yi ) for classical regression,
or g(µ)=log(µ) for Poisson regression.

3.3.2 Assumptions of GLM

The following are assumptions of the generalized linear models

• The data Y1 ,Y2 , · · · ,Yn are independently distributed.

• The dependent variable Yi does NOT need to be normally distributed, but it typically
assumes a distribution from an exponential family (e.g. binomial, Poisson, multino-
mial, normal, etc.).
 9

• A GLM does NOT assume a linear relationship between the response variable and
the explanatory variables, but it does assume a linear relationship between the trans-
formed expected response in terms of the link function and the explanatory variables.

• Explanatory variables can be nonlinear transformations of some original variables.

• The homogeneity of variance does NOT need to be satisfied. In fact, it is not even
possible in many cases given the model structure.

• Errors need to be independent but NOT normally distributed.

• Parameter estimation uses maximum likelihood estimation (MLE) rather than ordi-
nary least squares (OLS).

3.4 Tweedie distribution

The Tweedie distribution is named after M. C. K. Tweedie 1984. Tweedie studied the
exponential dispersion model with power variance functions. This kind of distribution
can have a cluster of data items at zero. This property makes Tweedie distribution so
useful in modelling any zero-inflated data sets such as micro-insurance. It is a mixture of
a a Poisson and Gamma distribution.

Let N; be the number of claims;it follows a Poisson distribution with mean λ ; it is bound
between 0 to ∞.

Let Yi be the claim amount that is independent and identically distributed. Data-set will
generate a Gamma distribution with parameters α and γ.

SN = Y1 +Y2 +Y 3 +Y 4 +Y5 + · · · +YN (3)

The variable SN is a Poisson sum of Gamma. The probability of not having a claim is:

P(SN = 0) = e−λ

The p.d.f of the Tweedie distribution for y > 0 is therefore

 10

∞
λ n γ nα nα−1 −yγ
fs (y) = ∑ e−λ n! Γ(nα)
y e (4)
n=1

The mean and variance will be:

α
E(SN ) = λ (5)
γ
and
λα
Var(SN ) = (1 + α) (6)
γ2

Changing the Tweedie parameters λ , α, 1

γ to Poisson-Gamma parameters we get:

µ 2−p
λ= (7)
φ [2 − p]
2− p
α= (8)
p−1
1
= φ (p − 1)µ p−1 (9)
γ

Substituting Tweedie’s new parameters above to equations (2) will be;

−1 µ 2−p
   
y
fs (y) = exp + + S(y, φ ) (10)
φ 2 − p [p − 1]µ [ p − 1]

If 1 < p < 2, then Y > 0, with a positive mass at Y = 0; refer to the histogram to make
this observation.

Note that if:

• p = 0 it becomes a normal distribution

• p = 1 it becomes a Poisson distribution

• p = 2 it becomes a Gamma distribution

• P = 3 it becomes an inverse Gaussian Distribution

Note that 1 < p < 2 We will apply this model on the data in the next chapter.
 11

3.5 XGBoost Model

This is a tree-based algorithm. It is an improvement from the initial decision tree; to

Bagging; to random forest; Boosting; gradient boosting; then to the popular XGBoost.
Decision tree showed a graphical representation of the of problem solutions under set con-
ditions then conclusions are drawn. The development to Bagging which groups together
meta-algorithms that combines predictions from multiple-decision trees. Then came an-
other improvement; random forest where only a subset of features are selected at random
to build a collection of decision trees. The came Boosting technique that reduces the error
from previous models while increasing performance models. A later improvement to Gra-
dient Boosting technique led to deployment of a gradient descent algorithm to minimise
errors as the model progresses. Then came a later improvement to the XGBoost which
we will apply in the project to solve a risk premium prediction.

The running of the model will lead to getting the best parameters that will best fit the
explanatory variables. The XGBoost has the following features that make it a viable option
for actuaries to use when pricing the micro-insurance products. These features are:

• Prevents over-fitting

• Accommodates missing values in your data-set

• Can be run on large data - It is a black box

• Can cross-validate at each iteration

• Auto-pruning is possible

We will derive the XGBoost model before we apply it to our data.

Ob j(θ ) = l(θ ) + Ω(θ ) (11)

Ob j(θ )-Objective function.

l(θ -Loss function.

 12

Ω(θ )-Regularization function.

The loss function in equation () measures how predictive the model is in relation with the
data being modelled.

l(θ ) = ∑i (yi − ŷi )2

The regularization function Ω(θ ) in equation () controls the complexity of the model
while performing iterations to avoid over-fitting.

n t
Ob j = ∑ l (yi − ŷi )t + ∑ ω( fi )
i=1 i=1

The objective function at step t will be:

n
 
1 2
Ob j(t) = ∑ gi ft (xi ) + hi ft (xi ) + ω( ft ) (12)
i=1 2

where;

(13)
 
gi = δyˆi (t−1) l yi , ŷi (t−1) (14)
 
2 (t−1)
hi = δyˆ (t−1) l yi , ŷi (15)
i

ω( f )− is the complexity of the tree.

Note that the function of the tree is:

ft (x) = ωq(x) , ω ∈ RT , q: Rd → [1, 2, 3, · · · , T ] (16)

ω− vector of scores on leaves.

q− Function assigning each data point to its corresponding leaf.
T − Number of leaves.
Therefore:

T
1
ω( f ) = γT + λ
2 ∑ ω 2j (17)
j=1
 13

Substituting equation () into equation () we will arrive at;

n
1 1 T 2
Ob j(t) = ∑ [gi ωq(xi ) + hi ωq(x
2
i)
] + γT + λ ∑ ωj
i=1 2 2 i=1
" ! !#
T
1
=∑ ∑ gi ω j + 2 ∑ hi + λ ω 2j + γT (18)
j=1 i∈I j i∈I j

Where I j = (i/q(xi ) = j) - Are the set of indices of data points assigned to the jth leaf.

Gj = ∑ giH j = ∑ hi
i∈I j i∈I j

Hence, the objective function will be:

T
1
Ob j(t) = ∑ [Giω j + 2 (hI + λ )ω 2j ] + γT (19)
j=1

Pruning is a very useful tool of the model since it prevents over-fitting.

Pruning of over-fitted data is done when Gain<λ i.e the split will not occur.

G2l G2r (Gl + Gr )2

 
1
Gain = + − −γ (20)
2 Hl + λ Hr + λ Hl + Hr + λ

We will therefore go to the next chapter to implement this model.

 14

4 DATA ANALYTIC AND INTERPRETATION

4.1 Introduction

In this section we will compare the two pricing models with the same data set. The Data
was run with RStudio programming language. The codes are attached for reference.

Figure 4.1,shows the characteristics of the response and explanatory variables that have
been used to run the two models.

A look at the below histogram figure

The data contains 1338 entries has more that a quarter of it with no claim reported; the
histogram has captured the same visually. The age range of the policyholders is between
28 years and 64yrs having an almost equal proportion in the gender disparity. The Body
Mass Index goes from a minimum of 16 to 53.10. The policyholders have between 1 to
5 children recognised in the policy details. 3% of them are smokers. The independent
 15

variables presented contribute to the frequency and severity of claims being brought forth.
We will use three methods to investigate the same data and later infer the model with
the best results. The three models are: Simple Generalised Linear model; Generalised
Linear Model with Tweedie distribution; Tweedie distribution with xgboost model. We
will interrogate one at a time

4.2 Generalized Linear Model

A GLM was run on the same data that produced the below result.

The model runs on assumptions like:linear relations between the dependent and the in-
dependent variables; the independent variables don’t have multicollinearity correlation;
the error components are normally distributed which might not be the case for all the
data sets. Before the model is fitted; all these aspects need to be investigated. Therefore,
scouting for a more robust model.
The Beta of the smoking status has the highest number concluding that this explanatory
variable has the highest relation with the claim amount. The least contributing factor is
Body Mass Index. The age of the policyholder has the second highest weighting. The
RMSE of this regression model is 722.8002. Its AIC equally on the higher side.

4.3 Generalized linear model Tweedie distribution

RMSE for Generalised Linear Model with the Tweedie distribution is 641.591. Its AIC is
15220. This model runs the Generalised Linear Model including the Tweedie distribution
because of the unique characteristics of the micro insurance data.
 16

4.4 Tweedie distribution with xgboost model

The data that is being used to run this model must be changed to sparse metrics. Gender
and smoking status are the variables that were affected since they were factor variables.

The ability of an actuary to change the parameters while modeling is a significant step
towards finding the best fit model. This can be done automatically by use of the function
Grid − search We did change the parameters eta and maxd epth and the output difference
can be seen in the figures from(fig:4) to figure(fig:6)below.

nrounds defines the maximum number of iterations/trees.

maxd epth instructs the number of branched the tree can have; increases the chance of
the model to view interactions. It is worth noting that the a higher value increases the
complexity of the model. The recommended range is 1 − 5 but it can take values 1 → ∞.

eta- learning rate; defines the gradient of the errors. The lower the value of eta the higher
the nrounds. The recommended range is between 0 and 0.3.

gamma is set at zero to prevent from over-fitting the model. var.power while running
insurance date ranges between [1,2].

4.5 Sensitivity analysis.

XGBoost comes with a built-in cross validation tool for each iteration. However, to see
the effect of different model parameters I have a run sensitivity analysis as below
 17

figure 1 : eta = 0.1 and the max depth= 3

The graph generated with these parameters have lead to under-fitting of the train data
compared to the test data all through the hundred iterations.
figure 2 : eta = 0.1 and the max depth= 5

The trend noted with the inputed parameters indicate a sudden variation of the standard
error hence not the best parameters to set on the model.
figure 3 : eta = 0.3 and the max depth= 3
 18

It is the best model parameters. The standard error between the train data and the test
data is minimal even as the number of rounds increases. The RMSE of the train data was
optimal at round 62 at a value of 478.7296. This makes the approximation of the claim
amount considering the respective explanatory variables reliable. It will neither be over-
fitted nor under-fitted.

figure 4 : eta = 0.3 and the max depth= 5

The graph shows that there was a great deal of over fitting. With these parameters of out-
put has the most drastic standard error therefore becoming unfit to be used to simulate
 19

the claims amount.

4.6 Models comparison

MODEL RMSE
GLM 722.8002
Tweedie GLM 641.591
Tweedie GLM in XGBoost Model 478.7296

It is evident that the RMSE of the three models have been improving as we went along.Therefore
the Xgboost with Tweedie distribution is the best model since it has the lowest RMSE.
 20

5 CONCLUSION AND RECOMMENDATIONS

5.1 Conclusion

The Simple GLM model doesn’t account for correlation between the independent vari-
ables and non-linear terms. This characteristic therefore takes away from the model’s
efficiency.

Tweedie distribution with Xgboost model is a robust model choice due to its ability to
easily tweak the different model until the desired results are achieved.

5.2 Recommendation

By studying the topic I have come across limitations like; scanty academic research in this
area and data inadequacies. This is to encourage others to study this area using available
big data from the social sciences since getting primary data from the industry has proven
impossible due to data protection. Xgboost model has capacity to a manipulate big data.

The government is strongly pushing the Universal Health Coverage agenda; this can be
treated a micro-insurance after population segmenting.This can be an areas of future re-
search. IRA is also offering incubation for start-ups with innovative insurance solutions
that would increase penetration.

Insurers should consider making strategic partnerships that will help it with the distribu-
tion of the product hence increasing its sales at the minimum cost possible.

A researcher can carry out Sensitivity analysis on λ , tweedie variance power.

 21

Bibliography

[1] M Ahmed. Market research on microinsurance demand. Technical Assistance Consul-

tant’s Report-Market Research on Microinsurance Demand. Project, (4761-SRI), 2006.

[2] Oscar Joseph Akotey, Kofi A Osei, and Albert Gemegah. The demand for micro
insurance in ghana. The Journal of Risk Finance, 2011.

[3] Christian Biener. Pricing in microinsurance markets. World Development, 41:132–

144, 2013.

[4] James C Brau, Craig Merrill, and Kim B Staking. Insurance theory and challenges
facing the development of microinsurance markets. Journal of Developmental En-
trepreneurship, 16(04):411–440, 2011.

[5] Craig Churchill. Trying to understand the demand for microinsurance. Journal of
International Development, 14(3):381, 2002.

[6] Yvonne Deblon and Markus Loewe. The potential of microinsurance for social pro-
tection. Craig Churchill/Michal Matul (2012)(eds.): Protecting the Poor: A Microinsur-
ance Compendium, 2, 2011.

[7] Martin Eling, Shailee Pradhan, and Joan T Schmit. The determinants of microinsur-
ance demand. The Geneva Papers on Risk and Insurance-Issues and Practice, 39(2):224–
263, 2014.

[8] Steve KB Magati. Influence of distribution channels on consumer purchase behavior

in the life insurance industry: a case of Britam Life Assurance Company Kenya. PhD
thesis, Strathmore University, 2021.

[9] Marc Maleika and Anne T Kuriakose. Microinsurance: extending pro-poor risk man-
agement through the social fund platform. 2008.

[10] Cecilia K Mutia. Access to Credit and Household Savings in Kenya Evidence From
Kenya National Finaccess 2019 Survey. PhD thesis, University of Nairobi, 2020.

[11] Chache Willys Obuba. The effect of product pricing on the growth of micro insurance
by insurance underwriters in Kenya. PhD thesis, University of Nairobi, 2014.

[12] Esbjörn Ohlsson and Björn Johansson. Non-life insurance pricing with generalized
linear models, volume 2. Springer, 2010.
 22

[13] Inmaculada Peña-Sanchez. Applying the tweedie model for improved microinsur-
ance pricing. The Geneva Papers on Risk and Insurance-Issues and Practice, 44(3):365–
381, 2019.

[14] Jean-Philippe Platteau, Ombeline De Bock, and Wouter Gelade. The demand for
microinsurance: A literature review. World Development, 94:139–156, 2017.

[15] Gordon K Smyth and Bent Jørgensen. Fitting tweedie’s compound poisson model to
insurance claims data: dispersion modelling. ASTIN Bulletin: The Journal of the IAA,
32(1):143–157, 2002.