Stam Formula Sheet
Uploaded by
Olivia KristiantiStam Formula Sheet
Uploaded by
Olivia KristiantiExam STAM
Adapt to Your Exam
SEVERITY,
SEVERITY,FREQUENCY &
FREQUENCY
AGGREGATE MODELS
MODELS
MODELS
& AGGREGATE
á = 1 −−1 1, forfor = 1,1, 2,⋯
Zero-Truncated Distributions
© © E [ ö ] ö ©
Payment per Payment
: payment per payment
E[ ] = () ; E[ ] = E[ ] ⋅ (()
Basic
E[(á)] = 1 −− E[E[]
() = Pr(Pr( ≤ )) = *//- () d [©] = () = E[ −− ∣ > ] = E[E[((−−))•]
CDFs, Survival Functions, and Hazard Functions With ordinary deductible ,
ä = 11−−− ä1−−, forfoär = 1,1,2,⋯, ⋯
Zero-Modified Distributions
() = Pr(() > ) = *- () d Special Shortcuts for ( ) ( )
ℎ() = (- ) E[(ä)] = 1 − E[E[] Exponential(l () −
() = */ℎ() d = − ln(() ; () = ((-) (,,0) Class Property
Uniform(rm (,, ) ++2
= + , forfor = 1,1,2,⋯, ⋯ Pareto (,, ) −1− 1
E[()] = */// () ⋅ () d
Moments
Mixtures and Splices
Pareto (,, ) −1− 1
= * ′ ( ) ⋅ (
( ) d H Var[=]ã=,, (Proba
Bernoulli Shortcut
If
Pro−−Prob
Prbabiobabbiabilitytiylit=y =1 −
)M(1 −− ) , then:
S-P
D
D
raw moment: = =E[E[(]; −−=)]
H E[ö] = (1 ++ ) ™E´ ∧ 1 ++ ≠ − EÆE Æ ∧ 1 ++ Ø
The Ultimate Formula for Insurance
Var[ [
VarVar[ ]
] = M = M M
central moment:
∣ ∼∼ Neg.Bi
Poisson-Gamma Mixture
sonBinomi()al(l ( =,∼Gamma(
PoiPoisson( = )) (,, )
Gamma
where
Var[()])] = E[()) ] − E[E[()])]M
Cov(
Cov ( , ) = E [ ]
] −E
− E [ ]
] E[ ]
If
then Neg.
where
.
,
∞
: deductible (set to 0 if not applicable)
Covariance:
Coefficient ofvari_ ation: = ℎ(( )∣ ) =[ ⋅(()]) ( ) - ( )
Frailty Models
: policy limit (set to if not applicable)
: maximum covered loss,whi
: coinsurance (set to 1 if not applicable)
Skewness = _ ; Kurtosis = = ô − , where where = */ d ss, which equals ++
: inflation rate (set to 0 if not applicable)
Aggregate Loss
Loss Models
(()()0=) =E[E[]] () D
Moment and Probability Generating Functions
ö
Insurance Applications
: payment per loss E=[]∑=¥µ∂E[¥ ]E[]]
Collective Risk Model
If for independent and , then:
(()()1=) =E[E[(]( − 1)1) ⋯ ( − + 1)]
where is the derivative
E[ö(=ö)∧] = =E[(ù,,∧ ))<≥]
Policy Limits, •
• Var[] = E[]Var[] +Va + Varr[]E[]M
where () D
is the derivative
= Pr(Pr( > ) ′
Impact of Deductibles on Claim Frequency
Pr(Pr( ∣ ) = Pr(PrPr((∩))()=) Pr(Pr(Pr∣ () Pr()Pr())
Conditional Distributions = *üü() d + ⋅ () For ,
= * () d E[E[ ∧ ]] Poisson
∣∣() = Pr(Pr( < < )) ,where < < Increaseded Limit FactFactor: = E[ ∧ ]] Binomial ,, ,
Pr( = ) = E[PrPr( = ∣ )]
Law of Total Probability
•
•
: original limit
: increased limit
Neg.
Binomial
,, ,
E[] = EE[ ∣ ]
Law of Total Expectation
Deductibles, Binomiiaall ((,,)) π
ã ∼∼Neg.Exponent
Negative Binomial/Exponential Compound Models
Var[]] = EVarVar[ ∣ ] ] +VarE[ ∣ ]
Law of Total Variance E[ö=ö](= −−E[()•−=))ã]−,−0,0=, ,E[]] −≥<E[E[ ∧ ]]
Ordinary deductible:
⇕
E[(ö)] = E[/(( −−•))•] ( ) ª ∼∼Exponent
Binomial ™, 1 ++ º
Pr(EPr[(()=) ⋅,ℎ(ℎ, (=)]=) E=[Pr(Pr()]()]=⋅ E[E)[)ℎ⋅(Pr(Pr)]( = ))
Independence
For independent
•
and ,
= *¶/ −− d ial(l ([1 + ]])
•
Parametric Distributions
= *¶ ( −− )()E[Ed[ ∧ ]] Compound Poisson Models
A collective risk model where the frequency
Special Distribution Shortcuts
(,, ) − −o(,∣ >+) Loss eliminiation ratio: = E[] follows a Poisson distribution.
Pareto
Exponent ial ( ) Paret
Exponent ial ( ) E[ö=ö]ã=,0,E[(≥<−)) ] ++ ⋅ ()
Franchise deductible:
rm (,, )
Uniform( rm (0, − ))
Uniform( •
www.coachingactuaries.com Copyright © 2018 Coaching
Coaching Actuaries.
Actuaries. All Rights Reserved.
Reserved. 1
Risk Measures
Value-at-Risk (VaR)
Choosing from (,,0)
Class
Two methods to fit data to an (,,0) class
( )Plot
Graph the difference between empirical CDF and
VaRæ() = .I() distributions:
• ̅ Gg‹
Method 1: Compare and
fitted CDF
Tail-Value-at-Risk (TVaR)
TVaRæ() = E ∣ > VaRæ ()y • Method 2: Observe the slope of g‹›À
= VaRæ() + VaRæ ()y Distribution Method 1 Method 2
TVaRæ() Poisson ̅ = 0
+ ¿¬ æ √ƒ Binomial ̅ > Negative
Normal
1− Neg. Binomial ̅ < Positive
Lognormal E[] ⋅ ¿Φ¬1 −−æ√ƒ Variance of MLE
Fisher’s Information
One Parameter:
() [′′()]
Var”=y =−E[()].I
Coherence
() is coherent if it satisfies the properties below: Peak:() = g¬ √− ∗¬∗√
• Translation invariance:( +=) =⋅ (()) +
() () = g¬.I√− ¬√
Valley:
• Positive homogeneity:
( + ) ≤ () +() Two Parameters:
() (,) = −E ¿fi,HfiHH—((,,)) fi,HHH— (,)ƒ - Plot
•
•
Subadditivity:
Monotonicity: ≤ () Pr( ≤ ) = 1
, if — (,”) Coordinate: Óg¬ √,∗¬√Ô where g¬√ = + 1
VaR is not coherent because it fails subaddivity. Var[÷]
[(,)].I = flCov÷, Cov÷,y‡
TVaR is coherent.
Tail Weight
”y Var”y Hypothesis Tests: Chi-Square Goodness-of-Fit
Chi-Square Goodness-of-Fit Test
G ¬ − √
À(-) = ∞ lim ÕÀ(-) = ∞⟹ Test statistic: = Ò where
1. Fewer positive raw moments heavier tail Delta Approximation
lim One-Variable:
Var¬”√y ≈ Æ ()Ø Var”y ∂I
2. If or , then numerator
-→ Ã(-) -→ ÕÃ(-) ::#expected
of groups
has a heavier tail. •
3. ℎ()decreases with⟹ heavy tail •
# of observations in group
4. ()increases with⟹ heavy tail Two-Variable:
Var¬÷,”√y ≈ (fiH)VarH [÷] +2” fiH—HCov÷,”y • : actual # of observations in group
+(—) Vary Degrees of freedom
• = − 1 −
: # of estimated parameters
where
CONSTRUCTION
CONSTRUCTIONAND SELECTION OF
AND SELECTION OF
Confidence Interval Chi-Square Goodness-of-Fit Test Properties
PARAMETRIC MODELS
PARAMETRIC MODELS
Maximum Likelihood Estimators
” ±(I•æ)/‰Var”y •
•
Individual and grouped data
Continuous and discrete fit
Steps to Calculating MLE Hypothesis Tests • No adjustments to critical value for censored
1. (()) == l∏n (())
3. H()H = –—– () B
: null hypothesis data
2.
4. Set () = 0 I
: alternative hypothesis
Reject B when test statistic > critical value
• If parameters are estimated, critical value is
automatically adjusted via degrees of freedom
Incomplete Data No change for critical value if s ample size is
()⁄() is true is false
•
large
Left-truncated at
Right-censored at () Reject Type I
Error
Correct
Decision
• Data needs to be grouped according to
• More weights on intervals with poor fit
Grouped data on interval
(,] Pr( < ≤ ) Fail to reject
Correct Type II Hypothesis Tests: Likelihood Ratio
= #−I)of#−free
Test statistic: = 2[( )]
(Bparameters
Decision Error
of free parametersin inI B
Special Cases Degrees of freedom
Hypothesis Tests: Kolmogorov-Smirnov
Distribution Shortcuts Empirical Distribution
Gamma,
fixed ” = ̅ Equal probability for each observation
g () = # of observations ≤ Score-Based Approaches
Two types of criteria:
̂g= ̅ • Schwarz Bayesian Criterion (SBC), a.k.a.
÷ = ∑¥∂I ¥ −̂ Kolmogorov-Smirnov Test Bayesian Information Criterion (BIC)
ÍÎÎ ∗ y
Normal
Test statistic: = max where • Akaike Information Criterion (AIC)
̂ = ∑g¥∂Iln¥ = max¬Ïg ¬√− ¬√Ï,Ïg ¬.I√− ∗¬√Ï√
If data is truncated at , then
SBC/BIC − 2 ln
( ) − ( ) −
÷ = ∑g¥∂I(ln¥) − ̂
Lognormal
∗() = 1 − () ,for ≥ AIC
Poisson ◊ = ̅ Kolmogorov-Smirnov Test Properties
where
::# of estimated parameters
log-likelihood
Binomial,
fixed
÷ = ̅ •
•
Individual data only
Continuous fit only : sample size
◊ = ̅
• Lower critical value for censored data Select model with the highest SBC or AIC value.
Neg. Binomial,
fixed • If parameters are estimated, critical value
should be adjusted
Zero-Truncated Distribution: • Lower critical value if sample size is large
• Match E[á] ̅
to
Zero-Modified Distribution:
•
•
No discretion
Uniform weight on all parts of distribution
• Match
Match
E[Bää] ̅
to the proportion of zero observations
to
•
Uniform Distribution on (0,):
• ” = max(I,,…,g)
www.coachingactuaries.com Copyright © 2018 Coaching Actuaries. All Rights Reserved. 2
CREDIBILITY CREDIBILITY
Classical Credibility
Bühlmann As Least Squares Estimate of
Bayesian
Minimize ∑ÍÎÎ - ´-¬- − ”-√≠ =
where
Bayesian estimate = Bühlmann estimate
Exact Credibility
• Poisson/Gamma
a.k.a. Limited Fluctuation Credibility
”-- : Bayesian estimate given = • Binomial/Beta
Full Credibility
# of exposures needed for full credibility, ı :
: Bühlmann estimate given •
•
Exponential/Inv. Gamma
Normal/Normal
ı = ´(•æ) ⁄≠ ()
Full credibility of aggregate claims: Properties of a Bayesian/Bühlmann graph
• Bühlmann estimates are on a straight line Empirical Bayes Non-Parametric Methods
̂ = ∑¥∂∑⋅∂ ¥
Bayesian estimates are within the range of Uniform Exposures
# of claims needed for ful l credibility, ˆ
:
•
hypothetical means
There are Bayesian estimates above and below
!
ˆ = ´(•æ) ⁄≠ flµµ + ‡ ÷ = ∑¥∂ ∑∂( ¬− ¥1)− ̅¥√
Full credibility of aggregate claims: •
the Bühlmann line !
• Bühlmann estimates are between the sample
•
•
Full credibility of claim frequency: set
˜
Ã̄ = 0
Full credibility of claˆim severity: set ˘¯ = 0
mean and theoretical mean
Conjugate Priors ÷ = ∑¥∂(−̅ ¥ −1 ̅) − ÷
!
ˆ = ı ⋅ µ ; ı = µ Poisson/Gamma
• Model: Poisson ()
Gamma(,) ̂ ∑¥∂ ∑∂ ¥¥
Non-uniform Exposures
• Prior:
(∗ ∣ data )∼ Gamma (∗,∗) = ! "
Credibility premium: ˙ == ̅ ++((1̅−−))
Partial Credibility
∗ = + ∑¥∂¥ ÷ = ∑ ¥∂!
∑ ¥¬¥ − ̅¥√
∂"
Posterior
= Ó— + Ô ∗ ∗ ÷ = ∑¥∂ ∑¥¥∂(̅ ¥(−¥−̅ )1)− ÷( − 1)
• !
where
: manual premium
: credibility factor/credibility
•
Neg.Binomial ( = , = )
!
− ∑¥∂ ¥ !
Square Root Rule: = ¸ı = ¸′ˆ
Predictive
Estimate E M as: ̂ = ∑∑¥∂¥∂¥¥̅ ¥
Balancing the Estimators
Binomial/Beta
• Model: ( ∼ Beta(,
∣ ) ∼ Bin,1omi) al (,) H
!
′
where
: actual # of exposures
: actual # of claims
• Prior:
(∗ ∣ data ) ∼ Beta(∗,∗,1) ÷
Empirical Bayes Semi-Parametric Methods
Bayesian Credibility
Posterior • ∗ == ++ [∑¥∂()¥− ∑¥∂ ¥] To estimate :
Model Distribution •
PoiNeg.sson()
B i
Model
n omi a l ( , ) ̅ ( 1 +̅ )
%
Model density function: ( ∣ )
Distribution of model conditioned on a parameter Predictive -
Gamma (,) ̅
Prior Distribution
• Model:
(∼ Inv.∣ )G∼amma(,)
Exponential/Inv. Gamma
Exponential () To estimate ̂ ÷
and , use the non-parametric
Prior density function: ( )
Initial distribution of the parameter
• Prior:
(∗∣ data) ∼ Inv.Gamma(∗,∗)
method formulas shown above.
Posterior Distribution
() ⋅∣ (data)
Revised distribution of the parameter Posterior • ∗ == ++ ∑¥∂ ¥
( ∣ data) = ∫// (data ∣ ) ⋅ ()) d
( data ∣
Posterior density function: •
Predictive Pareto( = ∗, = ∗)
(∼Normal(,)
∣ ) ∼ Normal (,)
Normal/Normal
Predictive Distribution
Revised unconditional distribution (w.r.t. model) • Model:
Prior:
Predictive Mean = Bayesian(Premi∣ data)
•
of the model
Predictive density function:
um (∗ ∣ data ) ∼ Normal (∗,∗)
Bühlmann Credibility
Posterior •
•
∗ == (1−̅ +)(1− )
= EE[ ∣ ]
Expected Hypothetical Mean (EHM):
Predictive Normal( = ∗, = + ∗)
= EVar[ ∣ ]
Expected Process Variance (EPV):
(∼ ∣ ) ∼ Uni(,) form (0, )
Uniform/S-P Pareto
• Model:
= VarE[ ∣ ]
Variance of Hypothetical Mean (VHM): • Prior: S-P Pareto
( ∗ ∣ data ) ∼ (∗,∗)
Bühlmann : = S-P Pareto
∗ == max(,
+ ,…,)
Bühlmann Credibility Factor: = + Posterior •
˙ == +̅ +((1̅ −− ))
Bühlmann Credibility Premium: Predictive -
www.coachingactuaries.com Copyright © 2018 Coaching Actuaries. All Rights Reserved. 3
Insurance Coverages
SHORT-TERM INSURANCES
Variable Expense Ratio: =ç ä
Expenses and Profit
Compensation: = /mimin,n(, ⋅),7, ≥< FiPerxedmiExpens e Rat i o : =
Homeowners Coinsurance
where
ssiëble Loss Ratio: PLR = 1− −ë
,
is the target profit and contingencies ratio
Disappearing Deductible Premium
,
Deductible decreases linearly over a specific range:
= 0,−7, < > ≤ − ≤ Premium
Claim Payment:
0, < ≤ ≤
−,
Aggregation
• Calendar Year (CY)
Current Rate Level
• Extension of Exposures
= − −− 7, < ≤ • Policy Year (PY) Method
• Parallelogram Method
⎩ , >
Loss Reserving
1.
2.
K=M. =KM.−⋅S
Expected Loss Ratio Method
Uí = Uì −U +Uí\
Unearned premium for CY :
Premium at Current Rates
Chain-Ladder Method
1. KUMUM..== ∏U,ZXWYZ[\⋅ UM.W
a.k.a. Loss Development Triangle Method Extension of Exposures Method
Recalculates the premiums of historical policies under the current rate l evel
Parallelogram Method
2.
3. = KM. −S Calculates average factors to be a pplied to the aggregate historical premiums
to make them on-level
=KMK. M. 1 − 1M.7 where
Bornhuetter-Ferguson Method Ratemaking
Indicated Avg. Rate Change = 1− +− ë −1
Loss Ratio Method
AlterMnat. ively,
•
•
is calculated based on the expected loss ratio method
is calculated based on the c hain-ladder method
Indicated RelativityU = Current RelativityU ⋅1+IòôöõnUdicated Avg.Rate Change
= ⋅ + (1 −) ⋅ where = 1M. Indicated Base Rate =InCurdicratentedBasAvg.e RatReleat⋅ivity Off-Balance Factor
Frequency-Severity Method
Alternate Method: Off-Balance Factor = CurrentAvg. Relativity
2.
3.
K=M. =KMk. M−.⋅SKM.
1. Apply the chain-ladder method to frequency and severity separately
Pure Premium Method
†
+
Indicated Avg.RateAvg.= 1R−ate − ë† †ç †
Closure Method:
UU,W,W == k̂Wq.k,UM,. −U,W\z
Frequency Avg.Rela†tivityU = Base RateUU
1.
2. Adj.†U = Avg.RelativityUU ⋅ Expos† u†reU
1.
2.
|U,W==∑U[U,WWÄ⋅ |UU,,WW
Aggregate
, where is the valuation CY
Indicated RelativityU = AdjIAdjndi.†.còôöõatU†ed Avg.Rate
Data Preparation
Losses
Indicated Base Rate = Indicated Avg.Relativity
Losses
New Relativity = (Indicated Relativity) + (1 −)(Current Relativity)
Credibility-Weighted Relativities
Other Topics
Aggregation
• Calendar Year (CY)
Develop to
Ultimate
Trending
• Trend Period = (()+
Increased Limit Factor
)+ ß®
• Accident Year (AY)
• Policy Year (PY)
• Loss Development
Factors
• Trend Factor
•
• : original limit
: increased limit
Rate of policy variation with limit = ß ⋅ Indicated Base Rate
Projected Losses
ÉU = SU + U − U\ © = ̅(−
)̅ − ()()
Loss Elimination Ratio
U
Incurred losses for CY :
where is the reserves at the end of CY
ÉU = SU + U
•
• : original deductible
: increased deductible
= (1− ©)⋅ Indicated Base Rate
U
Incurred losses for AY or PY :
where is the reserves as of the valuation date
Rate of policy variation with deductible
Copyright © 2018 Coaching Actuaries. All Rights Reserved. 4
www.coachingactuaries.com
www.coachingactuaries.com Personal copies
Copyright permitted.
© 2018 Resale
Coaching or distribution
Actuaries. is prohibited.
All Rights Reserved. 4
You might also like
- Institute and Faculty of Actuaries: Subject CT6 - Statistical Methods Core TechnicalNo ratings yetInstitute and Faculty of Actuaries: Subject CT6 - Statistical Methods Core Technical5 pages
- First Order Ordinary Differential EquationsNo ratings yetFirst Order Ordinary Differential Equations17 pages
- Insured Motor Vehicle Details Insured Declared Value (Idv) (In RS.)No ratings yetInsured Motor Vehicle Details Insured Declared Value (Idv) (In RS.)1 page
- Ordinary Differential Equations (Odes) : Department of Mathematics Iit Guwahati Ra/Rks/Mgpp/KvkNo ratings yetOrdinary Differential Equations (Odes) : Department of Mathematics Iit Guwahati Ra/Rks/Mgpp/Kvk16 pages
- Exam LTAM: You Have What It Takes To PassNo ratings yetExam LTAM: You Have What It Takes To Pass14 pages
- Engineering Agreement for Brazos CountyNo ratings yetEngineering Agreement for Brazos County10 pages
- N1067 Walkways and Ramps Excel BoQ 6-6-2022No ratings yetN1067 Walkways and Ramps Excel BoQ 6-6-202254 pages
- Berco: Model Number 700423-1 Tractor Winter CabNo ratings yetBerco: Model Number 700423-1 Tractor Winter Cab16 pages
- Calculus, Probability, and Actuarial ReviewNo ratings yetCalculus, Probability, and Actuarial Review29 pages
- Summary Lesson 1st Quarter Entrepreneurship 2025No ratings yetSummary Lesson 1st Quarter Entrepreneurship 20254 pages
- Canada Direct Deposit Enrolment Form - Direct Deposit - Receiver General For Canada - PSPC - Canada - CaNo ratings yetCanada Direct Deposit Enrolment Form - Direct Deposit - Receiver General For Canada - PSPC - Canada - Ca7 pages