BA3202

Actuarial Statistics
Actuarial Statistics
Course opening
Instructor:
Wenjun Zhu, PhD, FSA, CERA
Assistant Professor
Email: wjzhu@ntu.edu.sg
Office: S3-B1B-71
Tel: (65)6592-1859

Division of Banking & Finance

Nanyang Technological University

BA3202
L1
 Actuarial Statistics
• All course material posting on NTULearn

• KEY READING:
• Lecture slides for each week will be uploaded on NTULearn

Wenjun Zhu wjzhu@ntu.edu.sg

• CS1 – Actuarial Statistics 1.
• CS2 – Actuarial Statistics 2.
• Additional materials:
• Practice questions problem set – basic + exam style questions
• Additional tutorial questions – exam style questions

BA3202
L1
Course assessment

Actuarial Statistics
Components Marks
Coursework: Participation) 10

Wenjun Zhu wjzhu@ntu.edu.sg

Coursework: Presentation) 10
Coursework: Mid-term quiz
10
(Likely the week after recess week)
Final examination 70

Total 100

BA3202
L1
Tutorial presentations

Actuarial Statistics
• Group presentation (10 marks)
• Timing: 30 - 45 min
• Material for presentation
• A simple case study, applying previous lecture
concepts/knowledge/methodologies/formulae into real economic/actuarial
scenarios.

Wenjun Zhu wjzhu@ntu.edu.sg

• Can be either quantitative or qualitative
• E.g., introduce a business scenario that apply the learnt knowledge
• E.g., do some simple data analytics examples
• Anything that relates the previous lecture is acceptable
• Practice question problem set
• Needs to be uploaded to NTUlearn after the presentation
• Marking scheme
• Case study & problem set (50%)
• Presentation & communication (50%)
• Mid-term quiz
• After recess week
• Coverage: week 1- week 6 5
• Same question style as the final
BA3202
L1
Expectation of you

Actuarial Statistics
• Catch up with material that’s left behind after each lecture
• Do practice questions after each lecture
• Let me know WHENEVER you have any trouble with my
teaching (speaking too fast/slow, slides or notes not clear

Wenjun Zhu wjzhu@ntu.edu.sg

enough etc.)

• HAVE FUN !!!

BA3202
L1
Expectations of me

Actuarial Statistics
• You are welcome of expressing any of your expectations of me
in terms of:
• Style of class
• Length of the seminars

Wenjun Zhu wjzhu@ntu.edu.sg

• Other questions
• Weekly office hour
• Wednesday 2:00 – 3:00 pm
• Or upon appointment
• Zoom/Teams channel
• Available upon appointment

BA3202
L1
BA3202

Actuarial Statistics
Actuarial Statistics

Lecture 1:
- Decision Theory
- Bayesian Statistics

BA3202
L1
Objectives

Actuarial Statistics
• Decision theory
1. Define the 2 player zero sum games.
2. Determine optimum strategies under the theory of games.
3. Explain what is meant by a decision function and a risk function.

Wenjun Zhu wjzhu@ntu.edu.sg

4. Apply decision criteria to determine which decision functions are
the best.
• Bayesian statistics
1. Use Bayes’ Theorem to calculate simple conditional probabilities.
2. Explain what is meant by a prior distribution, a posterior
distribution and a conjugate prior distribution.
3. Derive the posterior distribution for a parameter in simple cases.
4. Explain what is meant by a loss function.
5. Determine Bayesian estimates of parameters based on loss 9
functions.
BA3202
L1
 Decision Theory

L1

Wenjun Zhu wjzhu@ntu.edu.sg Actuarial Statistics

10

BA3202
Game Theory

Actuarial Statistics
• Zero-sum two-person games
• A game is said to be zero-sum when there are two
players in conflict and whenever one player loses,
the other player wins.
• The net change in benefit/wealth is zero

Wenjun Zhu wjzhu@ntu.edu.sg

• Examples of zero-sum two player games?
• Futures, Options, etc.
• Stock trading
• Other examples?
• Assumptions
• Neither player knows what the other will do
• i.e., perfect understanding and perfect information
• Strategy cannot be changed once chosen 11

BA3202
L1
Game Theory

Actuarial Statistics
• Represent the game with Loss Matrix
• Player A’s options: I, II, III ...
• Player B’s options: 1, 2, 3 etc.
• L(I, 1), L(II, 1), L(I, 2) and L(II, 2) are the values of the loss function

Wenjun Zhu wjzhu@ntu.edu.sg

• L(I, 1) represents the loss to A (and i.e., the gain to B) when A
chooses option I and B chooses option 1
• The entries in the matrix are called payoffs of the game
• The objectives of Game Theory: Player A
• Optimal strategies I II
• Value/payoff of the game 1 L(I,1) L(II,1)
Player B
2 L(I,2) L(II,2)

12

BA3202
L1
Domination

Actuarial Statistics
• Strategy I dominates II if Player A (LOSSES)
• I is at least as good as II I II III IV
• I is better in some cases 1 2 0 < 1 4
Player B
2 1 2 < 2 3

Wenjun Zhu wjzhu@ntu.edu.sg

• A dominated strategy can (GAINS)
always be discarded. 3 4 1 < 3 2

• Strategy II dominates III

• What will happen finally? • III is redundant

• How to determine the Player A (LOSSES)

optimal strategy? I II IV
1 2 0 < 4
Player B 13
2 1 2 < 3
(GAINS)
3 4 1 < 2 BA3202
L1
The minimax criterion

Actuarial Statistics
• Each player chooses strategy that has the Player A (LOSSES)
lowest “maximum-loss” I II III
• Player A’s maximum losses:
1 6 2 2
• I: 6; II:3; III: 5 Player B
2 4 3 5
• 3 is the lowest of all (GAINS)
3 -3 1 0

Wenjun Zhu wjzhu@ntu.edu.sg

• A will choose strategy II
• Player B’s maximum losses (lowest gains)
• S1:-2; S2: -3; S3:3 (Note: negative signs since
we look at losses)
• “-3” is the lowest of all
• B will choose strategy 2
• The game now has an equilibrium point
• Saddle Points (Local minimum/maximum)
is where element L(a,b) is the
14
• largest in column and
• smallest in row
BA3202
L1
Randomized strategy

Actuarial Statistics
• Alternative strategy: minimize the “maximum expected loss”
• Example: determine the optimum strategy for Player A
• No saddle point
• 𝑝: Probability of strategy I for player A Player A

Wenjun Zhu wjzhu@ntu.edu.sg

• If player B chooses S1, A’s expected loss: I II
𝐿! 𝑝 = 7𝑝 − 6 1 − 𝑝 = 13𝑝 − 6 1 7 -6
• If player B chooses S2, A’s expected loss: Player B
2 1 5
𝐿" 𝑝 = 𝑝 + 5 1 − 𝑝 = 5 − 4𝑝
• Find 𝑝 such that max{𝐿! 𝑝 , 𝐿" (𝑝)} is minimized
10
à achieved when
𝐿! 𝑝 = 𝐿" (𝑝)
5
!!
13𝑝 − 6 = 5 − 4𝑝 ⇒ 𝑝 =
!# 0
0 0.2 0.4 0.6 0.8 1
!! !! $!
• The value of the game=𝐿! = 𝐿" = -5
!# !# !# 15
-10

L1 L2
BA3202
L1
Statistical games

Actuarial Statistics
• Purpose of all statistical inference:
• learn more about the population
• Process: study a collected sample
• 2 person game: player A is the Statistician; player B is the
Nature.

Wenjun Zhu wjzhu@ntu.edu.sg

• Unlike the zero-sum game, statistician has some indication of
how the Nature behave, given information in the sample
• She is making decisions against the potential outcome of nature
• The resulting decision matrix (a 2×2 example):
• Statistician can make decisions 𝑎! and 𝑎" Statistician
• The nature has states 𝜃! and 𝜃" 𝑎! 𝑎"
𝜃! 𝐿(𝑎! , 𝜃! ) 𝐿(𝑎" , 𝜃! )
Nature 16
𝜃" 𝐿(𝑎! , 𝜃" ) 𝐿(𝑎" , 𝜃" )
BA3202
L1
Example

Actuarial Statistics
• A coin either has 1 head or 1 tail (Normal) or 2 heads (Biased).
• The statistician cannot inspect the coin, but can make a
decision based on one single sample toss.
• 2×2 matrix of available strategy
• 𝜃! : coin is two-headed (i.e., biased, denoted as B)

Wenjun Zhu wjzhu@ntu.edu.sg

• 𝜃" : coin has 1 head and 1 tail (i.e., normal, denoted as N)
• 𝑎! : Statistician’s decision is that coin is two-headed (B)
• 𝑎" : Statistician’s decision is that coin is normal (N)
• 𝐿(𝑎! , 𝜃" ): the loss function of strategy
𝑎! with nature 𝜃" Statistician
𝑎! 𝑎"
𝜃! 𝐿(𝑎! , 𝜃! ) 𝐿(𝑎" , 𝜃! )
Nature
• 2×2 matrix of available strategy 𝜃" 𝐿(𝑎! , 𝜃" ) 𝐿(𝑎" , 𝜃" ) 17

BA3202
L1
Example (Cont’d)

Actuarial Statistics
• If the statistician makes the wrong decision, there is a penalty
of $1.
Statistician
𝑎! 𝑎"

Wenjun Zhu wjzhu@ntu.edu.sg

𝜃! 𝐿(𝑎! , 𝜃! ) 𝐿(𝑎" , 𝜃! )
Nature
𝜃" 𝐿(𝑎! , 𝜃" ) 𝐿(𝑎" , 𝜃" )

Statistician
𝑎! 𝑎"
𝜃! 0 1
Nature
𝜃" 1 0 18

BA3202
L1
Example (Cont’d)

Actuarial Statistics
0, ℎ𝑒𝑎𝑑
• Sample random variable : 𝑥 = %
1, 𝑡𝑎𝑖𝑙
• 𝑥 is what the statistician observes based on one sample toss
• 𝑑# (𝑥): decision functions based on observation of 𝑥

Wenjun Zhu wjzhu@ntu.edu.sg

• Subscript 𝑖 indicates different decision functions
• Possible decision functions are
𝑑! 0 = 𝑎! and 𝑑! 1 = 𝑎" (see a head->B; see a tail->N)
𝑑" 0 = 𝑎! and 𝑑" 1 = 𝑎! (see a head->B; see a tail->B)
𝑑% 0 = 𝑎" and 𝑑% 1 = 𝑎" (see a head->N; see a tail->N)
𝑑$ 0 = 𝑎" and 𝑑$ 1 = 𝑎! (see a head->N; see a tail->B)
• Some decision functions may not be very meaningful.

19

BA3202
L1
Example (Cont’d)

Actuarial Statistics
• 𝑅(𝑑# , 𝜃$ ): expected loss under 𝜃$ nature using decision function 𝑑#
𝑅 𝑑# , 𝜃$ = 𝐸(𝐿(𝑑# 𝑥 , 𝜃$ ))
where 𝐿(𝑑# 𝑥 , 𝜃$ ) is the loss function under decision function 𝑑# and
the nature is 𝜃$ .

Wenjun Zhu wjzhu@ntu.edu.sg

• Let’s look at 𝑅 𝑑! , 𝜃$ :
• Under 𝜃! : Pr 𝑥 = 0 = 1, Pr 𝑥 = 1 = 0
• Under 𝜃" : Pr 𝑥 = 0 = 0.5, Pr 𝑥 = 1 = 0.5
• 𝑅 𝑑! , 𝜃! = Pr 𝑥 = 0 𝐿 𝑑! 0 , 𝜃! + Pr 𝑥 = 1 𝐿 𝑑! 1 , 𝜃! = 0
• 𝑅 𝑑! , 𝜃" = Pr 𝑥 = 0 𝐿 𝑑! 0 , 𝜃" + Pr 𝑥 = 1 𝐿 𝑑! 1 , 𝜃" = 0.5

• After class try to calculate 𝑅 𝑑# , 𝜃$ for 𝑖 = 2, 3,4. Now we have:

• 𝑅 𝑑! , 𝜃! = 0；𝑅 𝑑! , 𝜃" = 0.5
• 𝑅 𝑑" , 𝜃! = 0；𝑅 𝑑" , 𝜃" =1
• 𝑅 𝑑% , 𝜃! = 1；𝑅 𝑑% , 𝜃" =0 20
• 𝑅 𝑑$ , 𝜃! = 1；𝑅 𝑑$ , 𝜃" = 0.5
BA3202
L1
Example (Cont’d)

Actuarial Statistics
• Write into an expected payoff Statistician
matrix: 𝑑! 𝑑" 𝑑% 𝑑$
• 𝑑! dominates 𝑑" 𝜃! 0 0 1 1
• 𝑑# dominates 𝑑$ Nature
𝜃" 0.5 1 0 0.5
• Reduce to a new 2×2 payoff matrix

Wenjun Zhu wjzhu@ntu.edu.sg

Statistician
𝑑! 𝑑%
• Optimal randomized strategy:
𝜃! 0 1
" Nature
• select 𝑑! with probability and 𝑑# 𝜃" 0.5 0
#
!
with probability #
• Value of game:
1
𝐸 𝑙𝑜𝑠𝑠 𝜃! = 𝐸 𝑙𝑜𝑠𝑠 𝜃" = 21
3

BA3202
L1
Example (Cont’d)

Actuarial Statistics
• Minimax Criterion
• Under the minimax criterion, the optimal strategy is 𝑑!
• Bayes criterion:
• Bayes risk: 𝐸[𝑅(𝑑, 𝜣)]

Wenjun Zhu wjzhu@ntu.edu.sg

• 𝜣 is regarded as a random variable
• Expectation is taken with respect to 𝜣
• Bayes criterion seeks to minimize Bayes risk
• Suppose Pr 𝜃! = 𝑝 , Pr 𝜃" = 1 − 𝑝
• Bayes risk for 𝑑! : 0 ∗ 𝑝 + 0.5 1 − 𝑝 = 0.5 − 0.5𝑝
• Bayes risk for 𝑑% : 1 ∗ 𝑝 + 0 ∗ 1 − 𝑝 = 𝑝
Statistician
! !
• When 𝑝 > , choose 𝑑! ; when 𝑝 < , choose 𝑑% 𝑑! 𝑑%
% %

𝜃! 0 1
Nature 22
𝜃" 0.5 0
BA3202
L1
 Bayesian Statistics

L1

Wenjun Zhu wjzhu@ntu.edu.sg Actuarial Statistics

23

BA3202
Intro

Actuarial Statistics
• In classical statistics the parameter values (𝜃, or 𝛼, or 𝜆…any
notations) are considered to be fixed but unknown
• In Bayesian statistics the parameter values (𝜃, or 𝛼, or 𝜆…any
notations) are considered to be random numbers

Wenjun Zhu wjzhu@ntu.edu.sg

• Recall in probability: 𝑃 𝑋 𝑌 𝑃 𝑌 = 𝑃 𝑋 ∩ 𝑌 = 𝑃 𝑌 𝑋 𝑃(𝑋)

Bayes Theorem:
If 𝐵! , 𝐵" , … , 𝐵% constitute a partition of a sample space 𝑆 and 𝑃 𝐵# ≠
0 for 𝑖 = 1, 2, … , 𝑘, then for any event 𝐴 in 𝑆 such that 𝑃 𝐴 ≠ 0,
𝑃 𝐴 𝐵& 𝑃 𝐵&
𝑃 𝐵& 𝐴 =
𝑃 𝐴
where 𝑃 𝐴 = ∑%#'! 𝑃 𝐴 𝐵# 𝑃 𝐵# for 𝑖 = 1, 2, … , 𝑘.
24
• We use this theorem to find 𝑃 𝐵 𝐴 if we know 𝑃(𝐴|𝐵)
BA3202
L1
Bayes Theorem

Actuarial Statistics
Example:
3 manufacturers supply clothing to a retailer.
• 60% stock comes from manufacturer 1, 30% from manufacturer 2
and 10% from manufacturer 3.
• 10% of the clothing from manufacturer 1 is faulty, 5% from

Wenjun Zhu wjzhu@ntu.edu.sg

manufacturer 2 is faulty and 15% from manufacturer 3 is faulty.
• What is the probability that a faulty garment comes from
manufacturer 3?

Solution:
• Let 𝐴 be the event that a garment is faulty. Let 𝐵# be the event that
the garment comes from manufacturer 𝑖.
) *! ∩, ) ,|*! ) *!
• 𝑃 𝐵( 𝐴 = = =
) , ) ,|*" ) *" .) ,|*# ) *# .) ,|*! ) *!
/.!1∗/.!
= 0.167 25
/.!∗/.3././1∗/.(./.!1∗/.!

BA3202
L1
Prior/Posterior distributions

Actuarial Statistics
Suppose we have a random variable 𝑋 with known pdf 𝑓(𝑥|𝜃)
• Study target: 𝜃, unknown parameter that needs to be estimated
• Prior distribution of 𝜃: assume 𝜃 is known to have a pdf of 𝑓 𝜃
• The prior distribution comes from some other information known
about 𝜃 before drawing any sample from the population
• In the previous clothing example, the proportion of different

Wenjun Zhu wjzhu@ntu.edu.sg

manufacturers can be viewed as priors
• 𝑋 = (𝑋! , 𝑋" , … , 𝑋4 ) is a random sample from 𝑋 (population)
𝒇 𝑿𝜽 𝒇𝜽
• Posterior pdf of 𝜽: 𝒇 𝜽 𝑿 = ∝𝒇 𝑿𝜽 𝒇 𝜽
𝒇 𝑿
• With the sample 𝑋, we have more information about 𝜃 and therefore
obtained the posterior distribution
• 𝑓 𝑋 𝜃 = ∏'%&! 𝑓(𝑋% |𝜃): the likelihood of the sample is a function of
unknown 𝜃
• 𝑓 𝑋 = ∫ 𝑓 𝑋 𝜃 𝑓 𝜃 𝑑𝜃 : a constant that does not involve 𝜃
• Posterior is proportional to the likelihood function times the prior. 26

BA3202
L1
Example

Actuarial Statistics
If 𝑋~𝐸𝑥𝑝 𝜆 where 𝜆 is the unknown parameter that needs to be
estimated, and 𝑥! , 𝑥" , … , 𝑥4 is a random sample from 𝑋. Suppose
the prior distribution of 𝜆 is 𝐸𝑥𝑝 𝛼 , where 𝛼 is a known parameter.
Find the posterior distribution 𝑓 𝜆 𝑥 of 𝜆.
Solution:
• 𝑓 𝜆 = 𝛼𝑒 89: is the prior pdf of 𝜆

Wenjun Zhu wjzhu@ntu.edu.sg

• 𝑓 𝑥 𝜆 = 𝜆𝑒 8:; is the conditional pdf of 𝑋
• The likelihood function given 𝑥! , 𝑥" , … , 𝑥4 is
4
∑&
𝑓 𝑥𝜆 = N 𝜆𝑒 8:;$ = 4
𝜆 𝑒 8: $%" ;$

#'!
∑&
• Hence, 𝑓 𝜆 𝑥 ∝ 𝑓 𝑥 𝜆 𝑓 𝜆 = 𝛼𝜆4 𝑒 8 $%" ;$ .9 :

• Note: 𝛼, 𝑥! , 𝑥" , … , 𝑥' are known; 𝜆 is the only unknown variable

(!
• Recall: Gamma distribution Γ 𝑎, 𝑏 has a pdf: 𝑓 𝑥 = 𝑥 *+! 𝑒 +(,
) *
• Posterior distribution of 𝜆 is Γ 𝑛 + 1, ∑'%&! 𝑥% + 𝛼 27

BA3202
L1
Prior/Posterior distributions

Actuarial Statistics
• Step 1:
• Write down prior pdf of the unknown parameter 𝜃: 𝑓 𝜃
• Write down the conditional pdf of 𝑋: 𝑓 𝑥 𝜃 (i.e., the likelihood)
• Step 2:

Wenjun Zhu wjzhu@ntu.edu.sg

Write down the likelihood function 𝑓 𝑥 𝜃 given the random
sample {𝑥! , 𝑥" , … , 𝑥4 }
• Step 3:
Write down posterior pdf 𝑓 𝜃 𝑥 ∝ 𝑓 𝑥 𝜃 𝑓(𝜃)
• Step 4:
Identify the form of the posterior pdf that may be the same as any
standard distribution
• Note that in this step, only the form involving the random
variable matters. Any constants will be absorbed by the “∝” 28

BA3202
L1
Special cases

Actuarial Statistics
• Conjugate priors
• When the posterior and the prior distribution belong to the same
“general” family of distributions, this prior distribution is called
the conjugate prior.
• Conjugate distributions can be found by selecting a family of

Wenjun Zhu wjzhu@ntu.edu.sg

distribution with the same form as the likelihood function.
• Improper prior distributions
• Uninformative prior distribution: 𝑈(−∞, ∞)
• Assumes that an unknown parameter is equally likely to take any
value
• Does it make sense in practice?
• No because in this case the pdf of the prior is 0 everywhere
• In practice, usually assume a prior distribution with 𝑈(−𝑁, 𝑁)
!
where 𝑁 is a large number: 𝑓 𝜃 = 29
"=

BA3202
L1
The loss function

Actuarial Statistics
• Objective: To obtain the estimator of 𝜃

• Definitions:
• 𝑔 𝑥 : the estimator of 𝜃

Wenjun Zhu wjzhu@ntu.edu.sg

• A function of the sample
• 𝐿 𝑔 𝑥 , 𝜃 : the loss function
• Measures how much the estimates deviate from the
“true” value of 𝜃
• Expected posterior loss (EPL):
𝐸𝑃𝐿 = 𝐸 𝐿 𝑔 𝑥 , 𝜃 = ∫ 𝐿 𝑔 𝑥 , 𝜃 𝑓 𝜃 𝑥 𝑑𝜃
• Bayesian estimator: function 𝑔(𝑥) that minimizes the EPL
30

BA3202
L1
Common loss functions

Actuarial Statistics
#
• Quadratic loss: 𝐿 𝑔 𝑥 , 𝜃 = 𝑔 𝑥 − 𝜃
• Bayesian estimator: 𝑔 = 𝐸(𝜃|𝑋), mean of posterior

• Absolute error loss: 𝐿 𝑔 𝑥 , 𝜃 = 𝑔 𝑥 − 𝜃

Wenjun Zhu wjzhu@ntu.edu.sg

• Bayesian estimator: median of posterior
0, 𝑖𝑓𝑔 𝑥 = 𝜃
• “All or nothing” error: 𝐿 𝑔 𝑥 , 𝜃 = -
1, 𝑖𝑓𝑔 𝑥 ≠ 𝜃
• Bayesian estimator: mode of posterior

31

BA3202
L1
Quadratic loss function

Actuarial Statistics
• Bayesian estimator of quadratic loss function is the posterior
mean, 𝑔 = 𝐸(𝜃|𝑋).
• Scratch of proof:
"
• 𝐿 𝑔 𝑥 ,𝜃 = 𝑔 𝑥 − 𝜃

Wenjun Zhu wjzhu@ntu.edu.sg

"
• 𝐸𝑃𝐿 = ∫ 𝐿 𝑔 𝑥 , 𝜃 𝑓 𝜃 𝑥 𝑑𝜃 = ∫ 𝑔 𝑥 − 𝜃 𝑓 𝜃 𝑥 𝑑𝜃
• First order condition: take partial differentiation w.r.t. 𝑔
𝜕𝐸𝑃𝐿
= 2∫ 𝑔 − 𝜃 𝑓 𝜃 𝑥 𝑑𝜃 = 0
𝜕𝑔
⇨ 𝑔∫ 𝑓 𝜃 𝑥 𝑑𝜃 = ∫ 𝜃𝑓 𝜃 𝑥 𝑑𝜃
• Also note that,
∫ 𝑓 𝜃 𝑥 𝑑𝜃 = 1; ∫ 𝜃𝑓 𝜃 𝑥 𝑑𝜃 = 𝐸(𝜃|𝑥),
32
• Hence, 𝑔(𝑥) = 𝐸(𝜃|𝑥).
BA3202
L1
Absolute loss function

Actuarial Statistics
• Bayesian estimator under the absolute loss function is the median of the posterior.
• Scratch of proof:
• 𝐿 𝑔 𝑥 ,𝜃 = 𝑔 𝑥 − 𝜃
• 𝐸𝑃𝐿 = ∫ 𝐿 𝑔 𝑥 , 𝜃 𝑓 𝜃 𝑥 𝑑𝜃 = ∫ 𝑔 𝑥 − 𝜃 𝑓 𝜃 𝑥 𝑑𝜃
# "
= 9 (𝑔 − 𝜃)𝑓 𝜃 𝑥 𝑑𝜃 + 9 𝜃 − 𝑔 𝑓 𝜃 𝑥 𝑑𝜃
!" #
• First order condition: take partial differentiation w.r.t. 𝑔

Wenjun Zhu wjzhu@ntu.edu.sg

𝜕𝐸𝑃𝐿
𝜕𝑔
# "
= 9 𝑓 𝜃 𝑥 𝑑𝜃 + 𝑔 − 𝑔 𝑓 𝑔 𝑥 ⋅ 1 + 9 −1 ⋅ 𝑓 𝜃 𝑥 𝑑𝜃 − 𝑔 − 𝑔 𝑓 𝑔 𝑥 ⋅ 1 (∗∗∗)
!" #
# "
𝜕𝐸𝑃𝐿
⇨ = 9 𝑓 𝜃 𝑥 𝑑𝜃 − 9 𝑓 𝜃 𝑥 𝑑𝜃 = 0
𝜕𝑔 !" #
# "
⇨ ' 𝑓 𝜃 𝑥 𝑑𝜃 = ' 𝑓 𝜃 𝑥 𝑑𝜃
!" #
⇨ Pr 𝜃 ≤ 𝑔 𝑥 𝑥 = Pr(𝜃 > 𝑔(𝑥)|𝑥)
• Hence, 𝑔 𝑥 is the posterior median.

• Note that Equation (***) is because:

𝑑 ((&) ((&)
𝜕 33
9 𝑓 𝑢, 𝑥 𝑑𝑢 = 9 𝑓 𝑢, 𝑥 𝑑𝑢 + 𝑓 𝑏 𝑥 , 𝑥 𝑏) 𝑥 − 𝑓 𝑎 𝑥 , 𝑥 𝑎) 𝑥
𝑑𝑥 $(&) $(&) 𝜕𝑥
BA3202
L1
All-or-nothing loss function

Actuarial Statistics
• Bayesian estimator under the All-or-nothing loss function is the
mode of the posterior.
• Scratch of proof:
0, 𝑖𝑓𝑔 𝑥 = 𝜃
• 𝐿 𝑔 𝑥 ,𝜃 = ?
1, 𝑖𝑓𝑔 𝑥 ≠ 𝜃

Wenjun Zhu wjzhu@ntu.edu.sg

• Now consider a small neighbor around 𝜃: 𝜃 − 𝜀, 𝜃 + 𝜀 ,
• 𝐿 𝑔 𝑥 , 𝜃 becomes
0, 𝑔 − 𝜀 < 𝜃 < 𝑔 + 𝜀，
𝐿 𝑔 𝑥 ,𝜃 = ?
1, otherwise

𝐸𝑃𝐿 = ∫ 𝐿 𝑔 𝑥 , 𝜃 𝑓 𝜃 𝑥 𝑑𝜃
FHG
= 1 − ∫F+G 𝑓 𝜃 𝑥)𝑑𝜃 = 1 − 2𝜀𝑓(𝑔|𝑥) for small 𝜀
• EPL is minimized when 𝑓(𝑔|𝑥) is maximized 34
• Hence, 𝑔 𝑥 is the posterior mode.
BA3202
L1
 L1

Actuarial Statistics
BA3202