You are on page 1of 3

Exam C

Adapt to Your Exam

SEVERITY, FREQUENCY & AGGREGATE LOSS


SEVERITY, FREQUENCY & AGGREGATE LOSS Variances Frequency: Exposure & Coverage Modification
Variance of the Sample Mean
Basic Probability Poisson Binomial Neg. Bin.
For 𝑛𝑛 independent and identically distributed
Probability Functions Original
& (i.i.d.) random variables, the variance of the sample
qrs Y Exposure 𝑛𝑛R 𝜆𝜆 𝑚𝑚, 𝑞𝑞 𝑟𝑟, 𝛽𝛽
𝐹𝐹 𝑥𝑥 = 𝑓𝑓 𝑡𝑡 𝑑𝑑𝑑𝑑 = Pr 𝑋𝑋 ≤ 𝑥𝑥 mean is Var 𝑋𝑋 = .
'( d Pr 𝑋𝑋 > 0 = 1
𝑆𝑆 𝑥𝑥 = 1 − 𝐹𝐹 𝑥𝑥 = Pr 𝑋𝑋 > 𝑥𝑥 Bernoulli Shortcut Exp. Mod. 𝑛𝑛? 𝑛𝑛? 𝑛𝑛?
𝑓𝑓 𝑥𝑥 𝑎𝑎 probability = 𝑞𝑞 Exposure 𝑛𝑛? 𝜆𝜆 𝑚𝑚, 𝑞𝑞 𝑟𝑟, 𝛽𝛽
ℎ 𝑥𝑥 = 𝑋𝑋 = 𝑛𝑛R 𝑛𝑛R 𝑛𝑛R
𝑆𝑆 𝑥𝑥 𝑏𝑏 probability = 1 − 𝑞𝑞 Pr 𝑋𝑋 > 0 = 1
?
& Var 𝑋𝑋 = 𝑎𝑎 − 𝑏𝑏 𝑞𝑞 1 − 𝑞𝑞 Cov. Mod.
'8 &
𝐻𝐻 𝑥𝑥 = ℎ 𝑡𝑡 𝑑𝑑𝑑𝑑 = − ln 𝑆𝑆 𝑥𝑥 ; 𝑆𝑆 𝑥𝑥 = 𝑒𝑒 Conditional Variance Formula Exposure 𝑛𝑛R 𝑣𝑣𝑣𝑣 𝑚𝑚, 𝑣𝑣𝑣𝑣 𝑟𝑟, 𝑣𝑣𝑣𝑣
'( VarY 𝑋𝑋 = VarZ EY 𝑋𝑋 ∣ 𝑌𝑌 + EZ Varz 𝑋𝑋 ∣ 𝑌𝑌 Pr 𝑋𝑋 > 0 = 𝑣𝑣
Moments
( Aggregate Losses
Aggregate Loss Models
E 𝑔𝑔 𝑥𝑥 = 𝑔𝑔 𝑥𝑥 𝑓𝑓 𝑥𝑥 𝑑𝑑𝑑𝑑 If 𝑆𝑆 = 𝑋𝑋R + 𝑋𝑋? + ⋯ + 𝑋𝑋{ and 𝑋𝑋o are i.i.d, then: <
'( E 𝑆𝑆 = E 𝑁𝑁 E 𝑋𝑋
kth raw moment: 𝜇𝜇<= = E 𝑋𝑋 < Pr 𝑆𝑆 = 𝑛𝑛 = Pr 𝑁𝑁 = 𝑘𝑘 ⋅ Pr 𝑋𝑋o = 𝑛𝑛
Var 𝑆𝑆 = E 𝑁𝑁 Var 𝑋𝑋 + Var 𝑁𝑁 E 𝑋𝑋 ?
kth central moment: 𝜇𝜇< = E 𝑋𝑋 − 𝜇𝜇 < < oõR

Variance: 𝜎𝜎 ? = 𝜇𝜇? = E 𝑋𝑋 − 𝜇𝜇 ? = E 𝑋𝑋 ? − 𝜇𝜇 ? Policy Limits 𝐸𝐸 𝑆𝑆 − 𝑑𝑑 Ñ = 𝐸𝐸 𝑆𝑆 − 𝐸𝐸 𝑆𝑆 ∧ 𝑑𝑑


( (
𝜎𝜎
Coefficient of variation: 𝐶𝐶𝐶𝐶 = E 𝑋𝑋 <
= <
𝑥𝑥 𝑓𝑓 𝑥𝑥 𝑑𝑑𝑑𝑑 = 𝑘𝑘𝑥𝑥 <'R
𝑆𝑆 𝑥𝑥 𝑑𝑑𝑑𝑑 Risk Measures
𝜇𝜇 } } 𝜌𝜌 𝑋𝑋 is coherent if it satisfies (for 𝑐𝑐 > 0):
𝜇𝜇S Å
Skewness: 𝛾𝛾R = S E 𝑋𝑋 ∧ 𝑢𝑢 <
= 𝑥𝑥 < 𝑓𝑓 𝑥𝑥 𝑑𝑑𝑑𝑑 + 𝑢𝑢 < 1 − 𝐹𝐹 𝑢𝑢 • Translation invariance: 𝜌𝜌 𝑋𝑋 + 𝑐𝑐 = 𝜌𝜌 𝑋𝑋 + 𝑐𝑐
𝜎𝜎
𝜇𝜇V }
Å
• Positive homogeneity: 𝜌𝜌 𝑐𝑐𝑐𝑐 = 𝑐𝑐𝑐𝑐 𝑋𝑋
Kurtosis: 𝛾𝛾? = V • Subadditivity: 𝜌𝜌 𝑋𝑋 + 𝑌𝑌 ≤ 𝜌𝜌 𝑋𝑋 + 𝜌𝜌 𝑌𝑌
𝜎𝜎 = 𝑘𝑘𝑥𝑥 <'R 𝑆𝑆 𝑥𝑥 𝑑𝑑𝑑𝑑
Covariance: Cov 𝑋𝑋, 𝑌𝑌 = E 𝑋𝑋 − 𝜇𝜇Y 𝑌𝑌 − 𝜇𝜇Z } • Monotonicity: If 𝑋𝑋 ≤ 𝑌𝑌, then 𝜌𝜌(𝑋𝑋) ≤ 𝜌𝜌(𝑌𝑌)
= E 𝑋𝑋𝑋𝑋 − E 𝑋𝑋 E 𝑌𝑌 If 𝑌𝑌 = 1 + 𝑟𝑟 𝑋𝑋, then: VaR _ 𝑋𝑋 = 𝜋𝜋_ = 𝐹𝐹Y'R 𝑝𝑝
𝐶𝐶𝐶𝐶𝐶𝐶 𝑋𝑋, 𝑌𝑌 𝑢𝑢
E 𝑌𝑌 ∧ 𝑢𝑢 = E 1 + 𝑟𝑟 𝑋𝑋 ∧ 𝑢𝑢 = 1 + 𝑟𝑟 E 𝑋𝑋 ∧ TVaR _ 𝑋𝑋 = E 𝑋𝑋 ∣ 𝑋𝑋 > VaR _ 𝑋𝑋
Correlation coefficient: 𝜌𝜌YZ = 1 + 𝑟𝑟
𝜎𝜎Y 𝜎𝜎Z E 𝑋𝑋 − E 𝑋𝑋 ∧ VaR _ 𝑋𝑋
Percentiles Deductibles = VaR _ 𝑋𝑋 +
1 − 𝑝𝑝
100pth percentile 𝜋𝜋_ is any point satisfying Payment per Loss
𝐹𝐹 𝜋𝜋_' ≤ 𝑝𝑝 and 𝐹𝐹 𝜋𝜋_ ≥ 𝑝𝑝. If 𝐹𝐹 is continuous, 𝜋𝜋_ is 𝑌𝑌 É = 𝑋𝑋 − 𝑑𝑑 Ñ VaR _ 𝑋𝑋 TVaR _ 𝑋𝑋
a unique point satisfying 𝐹𝐹 𝜋𝜋_ = 𝑝𝑝. 𝐹𝐹Z Ö 𝑥𝑥 = 𝐹𝐹Y 𝑥𝑥 + 𝑑𝑑 𝜙𝜙 𝑧𝑧_
( ( Normal 𝜇𝜇 + 𝑧𝑧_ 𝜎𝜎 𝜇𝜇 + 𝜎𝜎
Mode E 𝑋𝑋 − 𝑑𝑑 Ñ = 𝑥𝑥 − 𝑑𝑑 𝑓𝑓 𝑥𝑥 𝑑𝑑𝑑𝑑 = 𝑆𝑆 𝑥𝑥 𝑑𝑑𝑑𝑑 1 − 𝑝𝑝
Mode is x that maximizes 𝑓𝑓 𝑥𝑥 . Ü Ü Φ 𝜎𝜎 − 𝑧𝑧_
Moment Generating Function = E 𝑋𝑋 − E 𝑋𝑋 ∧ 𝑑𝑑 Lognormal 𝑒𝑒 üц° ¢ E 𝑋𝑋
1 − 𝑝𝑝
𝑀𝑀Y 𝑡𝑡 = E 𝑒𝑒 cY Payment per Payment
á
d
𝑀𝑀Y 0 = E 𝑋𝑋 d where 𝑀𝑀 d is the nth derivative 𝑌𝑌 = 𝑋𝑋 − 𝑑𝑑 Ñ ∣ 𝑋𝑋 > 𝑑𝑑 Tail Weight Measures
𝐹𝐹Y 𝑥𝑥 + 𝑑𝑑 − 𝐹𝐹Y 𝑑𝑑 1. More positive moments ⟹ lower tail weight
Probability Generating Function 𝐹𝐹Z à 𝑥𝑥 =
1 − 𝐹𝐹Y 𝑑𝑑 ¶ & © &
𝑃𝑃Y 𝑡𝑡 = E 𝑡𝑡 Y 2. If lim ß > 1 or lim ß > 1, then numerator
d E 𝑋𝑋 − 𝑑𝑑 Ñ E 𝑋𝑋 − E 𝑋𝑋 ∧ 𝑑𝑑 &→( ¶® & &→( ©® &
𝑃𝑃Y 1 = E 𝑋𝑋 𝑋𝑋 − 1 ⋯ 𝑋𝑋 − 𝑛𝑛 + 1 á
𝐸𝐸 𝑌𝑌 = 𝑒𝑒Y 𝑑𝑑 = = has higher tail weight.
d 𝑆𝑆Y 𝑑𝑑 𝑆𝑆Y 𝑑𝑑 3. Increasing ℎ 𝑥𝑥 ⟹ lighter tail
where 𝑃𝑃Y is the nth derivative
Special Cases
Conditional Probability 4. Increasing 𝑒𝑒Y 𝑑𝑑 ⟹ heavier tail
Exponential: 𝑒𝑒Y 𝑑𝑑 = 𝜃𝜃
Pr 𝐴𝐴 ∩ 𝐵𝐵 𝜃𝜃 − 𝑑𝑑
Pr 𝐴𝐴 ∣ 𝐵𝐵 = Uniform on 0, 𝜃𝜃 : 𝑒𝑒Y 𝑑𝑑 = , 𝑑𝑑 < 𝜃𝜃
Pr 𝐵𝐵 2
𝑓𝑓 𝑥𝑥, 𝑦𝑦 𝜃𝜃 + 𝑑𝑑 EMPIRICAL MODELS
EMPIRICAL MODELS
𝑓𝑓Y 𝑥𝑥 ∣ 𝑦𝑦 = Two-parameter Pareto: 𝑒𝑒Y 𝑑𝑑 =
𝑓𝑓 𝑦𝑦 𝛼𝛼 − 1 Review of Mathematical Statistics
Bayes’ Theorem 𝑑𝑑 Bias
Pr 𝐵𝐵 ∣ 𝐴𝐴 Pr 𝐴𝐴 Single-parameter Pareto: 𝑒𝑒Y 𝑑𝑑 =
𝛼𝛼 − 1 bias™ 𝜃𝜃 = E 𝜃𝜃 ∣ 𝜃𝜃 − 𝜃𝜃
Pr 𝐴𝐴 ∣ 𝐵𝐵 = Franchise Deductible
Pr 𝐵𝐵 Consistency
𝑓𝑓Z 𝑦𝑦 ∣ 𝑥𝑥 𝑓𝑓Y 𝑥𝑥 0 𝑋𝑋 ≤ 𝑑𝑑
𝑓𝑓Y 𝑥𝑥 ∣ 𝑦𝑦 = 𝑌𝑌íÉ = 𝜃𝜃 is consistent if:
𝑓𝑓Z 𝑦𝑦 𝑋𝑋 𝑋𝑋 > 𝑑𝑑
E 𝑌𝑌íÉ = E 𝑋𝑋 − 𝑑𝑑 Ñ + 𝑑𝑑𝑑𝑑 𝑑𝑑 • lim Pr 𝜃𝜃d − 𝜃𝜃 < 𝛿𝛿 = 1 for all 𝛿𝛿 > 0, or
d→(
Law of Total Probability
• bias → 0 and Var 𝜃𝜃 → 0
• Discrete Loss Elimination Ratio
E 𝑋𝑋 ∧ 𝑑𝑑 Mean Square Error
Pr 𝐴𝐴 = Pr 𝐴𝐴 ∩ 𝐵𝐵o = Pr 𝐵𝐵o Pr 𝐴𝐴 ∣ 𝐵𝐵o LER 𝑑𝑑 = ?
E 𝑋𝑋 MSE™ 𝜃𝜃 = E 𝜃𝜃 − 𝜃𝜃 ∣ 𝜃𝜃
o o
?
• Continuous Severity: Coverage Modifications = Var 𝜃𝜃 + bias™ 𝜃𝜃
Pr 𝐴𝐴 = Pr 𝐴𝐴 𝑥𝑥 𝑓𝑓 𝑥𝑥 𝑑𝑑𝑑𝑑 𝑑𝑑 : deductible, 𝑢𝑢 : maximum covered loss Empirical Distribution for Complete Data
𝑌𝑌 É = (𝑋𝑋 ∧ 𝑢𝑢) − 𝑋𝑋 ∧ 𝑑𝑑 Individual Data
Conditional Expectation Formula E 𝑌𝑌 É = E 𝑋𝑋 ∧ 𝑢𝑢 − E 𝑋𝑋 ∧ 𝑑𝑑 #{𝑥𝑥o ≤ 𝑡𝑡}
EY 𝑋𝑋 = EZ EY 𝑋𝑋 ∣ 𝑌𝑌 If coinsurance factor = 𝛼𝛼: 𝐹𝐹d 𝑡𝑡 =
𝑛𝑛
E 𝑌𝑌 É = 𝛼𝛼 E 𝑋𝑋 ∧ 𝑢𝑢 − E 𝑋𝑋 ∧ 𝑑𝑑 Grouped Data
If inflation rate = 𝑟𝑟: Assume data are uniform on each interval
𝑢𝑢 𝑑𝑑 and interpolate linearly between endpoints
E 𝑌𝑌 É = 𝛼𝛼 1 + 𝑟𝑟 E 𝑋𝑋 ∧ − E 𝑋𝑋 ∧
1 + 𝑟𝑟 1 + 𝑟𝑟 of intervals.

www.coachingactuaries.com Copyright © 2017 Coaching Actuaries. All Rights Reserved. 1


Variance of Empirical Estimators with • If X is uniform kernel-smoothed: • Two variables
Complete Data 𝑏𝑏 ? 𝜕𝜕 ? 𝑙𝑙 𝜃𝜃R , 𝜃𝜃? 𝜕𝜕 ? 𝑙𝑙 𝜃𝜃R , 𝜃𝜃?
Var 𝑋𝑋 = Var 𝑌𝑌 +
𝑆𝑆d 𝑥𝑥 1 − 𝑆𝑆d 𝑥𝑥 3 𝜕𝜕𝜃𝜃R? 𝜕𝜕𝜃𝜃R 𝜕𝜕𝜃𝜃?
Var 𝑆𝑆d 𝑥𝑥 = • If X is triangular kernel-smoothed: I 𝜃𝜃R , 𝜃𝜃? = −EY
𝑛𝑛 𝜕𝜕 𝑙𝑙 𝜃𝜃R , 𝜃𝜃? 𝜕𝜕 ? 𝑙𝑙 𝜃𝜃R , 𝜃𝜃?
?
Cov [𝐹𝐹(𝑥𝑥), 𝐹𝐹(𝑦𝑦) − 𝐹𝐹(𝑥𝑥)] 𝑏𝑏 ?
Var 𝑋𝑋 = Var 𝑌𝑌 + 𝜕𝜕𝜃𝜃? 𝜕𝜕𝜃𝜃R 𝜕𝜕𝜃𝜃??
𝐹𝐹 𝑥𝑥 𝐹𝐹 𝑦𝑦 − 𝐹𝐹 𝑥𝑥 6
=− , 𝑥𝑥 < 𝑦𝑦 Covariance matrix of the MLE’s = 𝐼𝐼 'R 𝜃𝜃R , 𝜃𝜃?
𝑛𝑛 Mortality Table Construction Delta Method
Kaplan-Meier and Nelson-Åalen Estimators Individual-data-based Method • One variable
Kaplan-Meier (a.k.a. Product Limit) Estimator 𝑛𝑛 = number of years ?
Var 𝑔𝑔 𝑋𝑋 ≈ Var 𝑋𝑋 𝑔𝑔= 𝑥𝑥
≥'R • Exact exposure
𝑠𝑠o dܵ • Two variables
𝑆𝑆d 𝑡𝑡 = 1− , 𝑦𝑦≥'R ≤ 𝑡𝑡 < 𝑦𝑦≥ ' ? 𝑑𝑑≥
𝑟𝑟o 𝑞𝑞≥ = 1 − 𝑒𝑒 ¡µ
, Var 𝑞𝑞≥ = 1 − 𝑞𝑞≥ 𝑛𝑛? ∂𝑔𝑔 ?
oõR 𝑒𝑒≥? Var 𝑔𝑔 𝑋𝑋, 𝑌𝑌 ≈ Var 𝑋𝑋
Nelson-Åalen Estimator ∂𝑥𝑥
≥'R
• Actuarial exposure ∂𝑔𝑔 ∂𝑔𝑔 ∂𝑔𝑔 ?
𝑠𝑠o 𝑛𝑛𝑑𝑑≥ 𝑞𝑞≥ 1 − 𝑞𝑞≥ +2Cov 𝑋𝑋, 𝑌𝑌 + Var 𝑌𝑌
𝐻𝐻 𝑡𝑡 = , 𝑦𝑦≥'R ≤ 𝑡𝑡 < 𝑦𝑦≥ 𝑞𝑞≥ = , Var 𝑞𝑞≥ = ∂𝑥𝑥 ∂𝑦𝑦 ∂𝑦𝑦
𝑟𝑟o 𝑒𝑒≥ 𝑒𝑒≥ /𝑛𝑛
oõR
Interval-based Method Fitting Discrete Distributions
Variance of Kaplan-Meier and ¬
𝑃𝑃≥ = 𝑃𝑃≥'R + 𝑛𝑛≥'R ¬
− 𝑑𝑑≥'R − 𝑤𝑤≥'R − 𝑤𝑤≥'R¡
+ 𝑛𝑛≥ƒ Two methods to fit data to an 𝑎𝑎, 𝑏𝑏, 0 class
Nelson-Åalen Estimators distributions:
• Exact exposure
Variance of Kaplan-Meier Estimators • Method 1: Compare 𝜎𝜎 ? to 𝑥𝑥.
Greenwood’s approximation: 𝑒𝑒≥ = 𝑃𝑃≥ + 0.5 𝑛𝑛≥¬ − 𝑤𝑤≥¬ − 𝑑𝑑≥ d“
• Method 2: Calculate 𝑘𝑘 for the first few
𝑠𝑠≥ • Actuarial exposure d“”ß
Var 𝑆𝑆d 𝑡𝑡 = 𝑆𝑆d 𝑡𝑡 ? values of 𝑘𝑘 and observe the slope of the line
𝑟𝑟≥ 𝑟𝑟≥ − 𝑠𝑠≥ 𝑒𝑒≥ = 𝑃𝑃≥ + 0.5 𝑛𝑛≥¬ − 𝑤𝑤≥¬
¥µ ∂c created from these values
Variance of Nelson-Åalen Estimators Method 1 Method 2

𝑠𝑠o PARAMETRIC MODELS
PARAMETRIC MODELS Binomial 𝜎𝜎 ? < 𝑥𝑥 Negative
Var 𝐻𝐻 𝑦𝑦≥ = Poisson 𝜎𝜎 ? = 𝑥𝑥 Zero
𝑟𝑟o? Method of Moments
oõR Neg. Bin. 𝜎𝜎 ? > 𝑥𝑥 Positive
Log-transformed Confidence Interval for 𝑆𝑆d 𝑡𝑡 To fit a 𝑘𝑘-parameter distribution, set:
d
𝑆𝑆d 𝑡𝑡 R/∏ , 𝑆𝑆d 𝑡𝑡 ∏ where 1 Hypothesis Tests
𝐸𝐸 𝑋𝑋 ¬ = 𝑥𝑥o¬ for 𝑚𝑚 = 1, 2, … , 𝑘𝑘 𝐹𝐹 ∗ 𝑥𝑥 = 1 − 𝑆𝑆 𝑥𝑥 𝑆𝑆 𝑑𝑑
𝑧𝑧 RÑ_ Var 𝑆𝑆d 𝑡𝑡 𝑛𝑛
/? oõR 𝐷𝐷 𝑥𝑥 Plots
𝑈𝑈 = exp 𝐷𝐷 𝑥𝑥 = 𝐹𝐹d 𝑥𝑥 − 𝐹𝐹 ∗ 𝑥𝑥
𝑆𝑆d 𝑡𝑡 ln 𝑆𝑆d 𝑡𝑡 Percentile Matching
Smoothed Empirical Percentile 𝑝𝑝-𝑝𝑝 Plots
o
Log-transformed Confidence Interval for 𝐻𝐻 𝑡𝑡 𝜋𝜋_ = 𝑛𝑛 + 1 𝑝𝑝 »… observation Plots empirical distribution, 𝐹𝐹d 𝑥𝑥o = , on x-axis
dÑR
𝐻𝐻 𝑡𝑡 If 𝑛𝑛 + 1 𝑝𝑝 is not an integer, interpolation between and fitted distribution on y-axis.
, 𝐻𝐻 𝑡𝑡 𝑈𝑈 where
𝑈𝑈 the order statistics before and after the Kolmogorov-Smirnov Test
𝑛𝑛 + 1 𝑝𝑝 »… observation. Test statistic, 𝐷𝐷 = max 𝐷𝐷o where
𝑧𝑧 RÑ_ /? Var 𝐻𝐻 𝑡𝑡 Percentile Matching with Incomplete Data 𝐷𝐷o = max 𝐹𝐹d 𝑥𝑥≥ − 𝐹𝐹 ∗ 𝑥𝑥≥ , 𝐹𝐹d 𝑥𝑥≥' − 𝐹𝐹 ∗ 𝑥𝑥≥
𝑈𝑈 = exp • With censored data, select percentiles within the
𝐻𝐻 𝑡𝑡 • Only for individual data
range of the uncensored observations; • Lower critical value if 𝑢𝑢 < ∞
• With truncated data, match the percentiles of • If parameters are fitted, critical value should be
Kernel Smoothing the conditional distribution.
Density and Distribution Functions lowered
𝑘𝑘&º 𝑥𝑥 : kernel density function for point 𝑥𝑥o , Maximum Likelihood • Larger sample size has lower critical value
Steps to Calculating MLE • Uniform weight on all parts of distribution
evaluated at 𝑥𝑥
𝐾𝐾&º 𝑥𝑥 : kernel distribution function for point 𝑥𝑥o , 1. 𝐿𝐿 𝜃𝜃 = 𝑓𝑓 𝑥𝑥 3. Set 𝑙𝑙 = 𝜃𝜃 = 0 Chi-square Test
< ? <
evaluated at 𝑥𝑥 2. 𝑙𝑙 𝜃𝜃 = ln 𝐿𝐿 𝜃𝜃 4. Solve for 𝜃𝜃 𝑂𝑂≥ − 𝐸𝐸≥ 𝑂𝑂≥ ?
Likelihoods Test statistic, 𝑄𝑄 = = − 𝑛𝑛
𝑏𝑏: bandwidth 𝐸𝐸≥ 𝐸𝐸≥
• Individual Data ≥õR ≥õR
𝑓𝑓 𝑥𝑥 : PDF of the kernel-smoothed distribution If 𝑟𝑟 parameters are estimated using the same data
Right-censored at 𝑢𝑢 𝑆𝑆(𝑢𝑢)
𝐹𝐹 𝑥𝑥 : CDF of the kernel-smoothed distribution used to calculate the test statistic, then there are
Left-censored at 𝑑𝑑 𝐹𝐹(𝑑𝑑)
𝑓𝑓 𝑥𝑥 = 𝑓𝑓d 𝑥𝑥o 𝑘𝑘&º 𝑥𝑥 Right-truncated at 𝑢𝑢 𝑓𝑓 𝑥𝑥 𝐹𝐹 𝑢𝑢 𝑘𝑘 − 1 − 𝑟𝑟 degrees of freedom (assuming 𝑛𝑛 is fixed).

Left-truncated at 𝑑𝑑 𝑓𝑓 𝑥𝑥 𝑆𝑆 𝑑𝑑 • May be used for individual or grouped data
𝐹𝐹 𝑥𝑥 = 𝑓𝑓d 𝑥𝑥o 𝐾𝐾&º 𝑥𝑥 Left-truncated at 𝑑𝑑 and • No adjustments on critical value if 𝑢𝑢 < ∞
𝑆𝑆 𝑢𝑢 𝑆𝑆 𝑑𝑑 • If parameters are fitted, critical value is
&º right-censored at 𝑢𝑢
Uniform Kernel • Grouped Data automatically adjusted
1 • Critical value is independent of sample size
𝑘𝑘&º 𝑥𝑥 = , 𝑥𝑥o − 𝑏𝑏 ≤ 𝑥𝑥 ≤ 𝑥𝑥o + 𝑏𝑏 𝐹𝐹 𝑐𝑐≥ − 𝐹𝐹 𝑐𝑐≥'R
2𝑏𝑏 • Higher weight on intervals with low fitted
Grouped data between 𝑑𝑑 and 𝑐𝑐≥ and left-
𝑥𝑥 − 𝑥𝑥o − 𝑏𝑏 probability
𝐾𝐾&º 𝑥𝑥 = , 𝑥𝑥o − 𝑏𝑏 ≤ 𝑥𝑥 ≤ 𝑥𝑥o + 𝑏𝑏 truncated below at 𝑑𝑑: 𝐹𝐹 𝑐𝑐≥ − 𝐹𝐹 𝑑𝑑 𝑆𝑆 𝑑𝑑 Likelihood Ratio Test
2𝑏𝑏
Triangular Kernel MLE = MOM Test statistic = 2(ln 𝐿𝐿R − ln 𝐿𝐿} )
𝑥𝑥 − 𝑥𝑥o + 𝑏𝑏 • Poisson’s 𝜆𝜆 Degrees of freedom = # free parameters in
, 𝑥𝑥o − 𝑏𝑏 ≤ 𝑥𝑥 ≤ 𝑥𝑥o • Binomial’s 𝑞𝑞 (𝑚𝑚 is known) alternate − # free parameters in null
𝑘𝑘&º 𝑥𝑥 = 𝑏𝑏 ?
−𝑥𝑥 + 𝑥𝑥o + 𝑏𝑏 • Negative binomial’s 𝛽𝛽 (r is known) Critical value from Chi-square distribution
, 𝑥𝑥o ≤ 𝑥𝑥 ≤ 𝑥𝑥o + 𝑏𝑏 • Gamma’s 𝜃𝜃 (𝛼𝛼 is known) Schwarz Bayesian Criterion and Akaike
𝑏𝑏 ?
Use area of triangles to calculate 𝐾𝐾&º 𝑥𝑥 . • Normal’s 𝜇𝜇 and 𝜎𝜎 Information Criterion
Variance of Maximum Likelihood Estimators 𝑟𝑟
1 SBC = ln 𝐿𝐿 − ln 𝑛𝑛
Height = , base goes from 𝑥𝑥R − 𝑏𝑏 to 𝑥𝑥R + 𝑏𝑏 Fisher’s Information 2
𝑏𝑏 AIC = ln 𝐿𝐿 − 𝑟𝑟, where
Moments of Kernel-smoothed Distributions • One variable
I 𝜃𝜃 = −EY 𝑙𝑙 == 𝜃𝜃 , Var θ = I 𝜃𝜃 'R 𝑛𝑛 = number of data points
If X is the kernel-smoothed distribution and Y is
𝑟𝑟 = number of parameters in model
the empirical distribution, then use the conditional
Choose model with highest penalized value.
expectation and conditional variance formula to
calculate E 𝑋𝑋 and Var[𝑋𝑋].

www.coachingactuaries.com Copyright © 2017 Coaching Actuaries. All Rights Reserved. 2


CREDIBILITY CREDIBILITY Conjugate Priors Estimating 𝑎𝑎

• Poisson/Gamma • Uniform exposures:
Limited Fluctuation Credibility Ì
If 𝑁𝑁|𝛬𝛬~Poisson Λ , and Λ~Gamma 𝛼𝛼, 𝛾𝛾 = 1/𝜃𝜃 : 𝑣𝑣 𝑋𝑋o − 𝑋𝑋 ?
𝑣𝑣
Full Credibility 𝑎𝑎 = 𝑠𝑠 ? − = −
𝛼𝛼 ∗ = 𝛼𝛼 + 𝑛𝑛𝑥𝑥
Basic premise: Pr 𝑅𝑅 − 𝐸𝐸 𝑅𝑅 ≤ 𝑘𝑘𝑘𝑘 𝑅𝑅 ≥ 𝑝𝑝 Posterior Gamma 𝑛𝑛 𝑟𝑟 − 1 𝑛𝑛
𝛾𝛾 ∗ = 𝛾𝛾 + 𝑛𝑛 oõR
where 𝑅𝑅 = frequency, severity, or aggregate loss 𝑟𝑟 = 𝛼𝛼 ∗ • Non-uniform exposures:

Predictive Neg. Bin. Ì ?
𝑁𝑁 = frequency 𝛽𝛽 = 1/𝛾𝛾 ∗ oõR 𝑚𝑚o 𝑋𝑋o − 𝑋𝑋 − 𝑣𝑣 𝑟𝑟 − 1
𝑎𝑎 =
𝑋𝑋 = severity • Normal/Normal 𝑚𝑚 − 𝑚𝑚 'R Ì
oõR 𝑚𝑚o
?
1 + 𝑝𝑝 If 𝑁𝑁|𝛩𝛩~Normal Θ, 𝑣𝑣 , and Θ~Normal 𝜇𝜇, 𝑎𝑎 :
𝑧𝑧 RÑ_ ? = Φ'R 𝑣𝑣𝑣𝑣 + 𝑛𝑛𝑥𝑥𝑎𝑎
2
𝜇𝜇 ∗ =
Number of claims needed for full cred. of 𝑣𝑣 + 𝑛𝑛𝑛𝑛 SIMULATION SIMULATION
Posterior Normal 𝑣𝑣𝑣𝑣
Number of 𝑧𝑧 RÑ_ ? ? 𝜎𝜎{?

𝑛𝑛fi = 𝑎𝑎 ∗ = Inversion Method


claims 𝑘𝑘 𝜇𝜇{ 𝑣𝑣 + 𝑛𝑛𝑛𝑛
𝜇𝜇 = 𝜇𝜇 ∗ 𝑥𝑥o = 𝐹𝐹 'R 𝑢𝑢o
?
𝑧𝑧 RÑ_ ? ? 𝜎𝜎Y Predictive Normal ?

Claim size 𝑛𝑛fi = 𝜎𝜎 = 𝑎𝑎 ∗ + 𝑣𝑣 Special Techniques


𝑘𝑘 𝜇𝜇Y? • Bernoulli/Beta Multiple Decrements
Aggregate 𝑧𝑧 RÑ_ ? ? 𝜎𝜎{? 𝜎𝜎Y? If 𝑁𝑁|𝑄𝑄~Bernoulli 𝑄𝑄 , and 𝑄𝑄~Beta 𝑎𝑎, 𝑏𝑏, 𝜃𝜃 = 1 : o'R
𝑛𝑛fi = + 𝑝𝑝o
losses 𝑘𝑘 𝜇𝜇{ 𝜇𝜇Y? 𝑎𝑎 ∗ = 𝑎𝑎 + 𝑘𝑘 𝑋𝑋o ~Binomial 𝑚𝑚 = 𝑛𝑛 − 𝑥𝑥≥ , 𝑞𝑞 =
Posterior Beta ∗ o'R
𝑏𝑏 = 𝑏𝑏 + 𝑛𝑛 − 𝑘𝑘 1− ≥õR 𝑝𝑝≥
If frequency is Poisson, then 𝜇𝜇{ = 𝜎𝜎{? . 𝑎𝑎 ∗ ≥õR
Predictive Bernoulli 𝑞𝑞 = ∗ Simulating 𝑎𝑎, 𝑏𝑏, 0 Class Distributions
Number of exposures needed for full credibility: 𝑎𝑎 + 𝑏𝑏 ∗
• Exponential/Inverse Gamma ln 1 − 𝑢𝑢<
𝑛𝑛fi 𝑠𝑠< = −
𝑛𝑛¡ = If 𝑁𝑁|𝛩𝛩~Exponential Θ , and 𝜆𝜆<
𝜇𝜇{

Θ~Inverse gamma 𝛼𝛼, 𝛽𝛽 : Distribution 𝝀𝝀𝒌𝒌
Partial Credibility 𝛼𝛼 ∗ = 𝛼𝛼 + 𝑛𝑛 Poisson 𝜆𝜆
𝑃𝑃fl = 𝑍𝑍𝑋𝑋 + 1 − 𝑍𝑍 𝑀𝑀 = 𝑀𝑀 + 𝑍𝑍 𝑋𝑋 − 𝑀𝑀 Posterior Inverse gamma 𝛽𝛽 ∗ = 𝛽𝛽 + 𝑛𝑛𝑥𝑥
Binomial −𝑚𝑚 ln 1 − 𝑞𝑞 + 𝑘𝑘 ln 1 − 𝑞𝑞
𝑛𝑛 = Number of claims observed 𝛼𝛼 = 𝛼𝛼 ∗ Neg. Bin. 𝑟𝑟 ln 1 + 𝛽𝛽 + 𝑘𝑘 ln 1 + 𝛽𝛽
Predictive 2-parameter Pareto 𝜃𝜃 = 𝛽𝛽 ∗
𝑛𝑛 𝑛𝑛fi , 𝑛𝑛 < 𝑛𝑛fi Start at 𝑘𝑘 = 0. Calculate d<õ} 𝑠𝑠< until the sum
𝑍𝑍 =
1, 𝑛𝑛 ≥ 𝑛𝑛fi exceeds 1. The simulated value is 𝑛𝑛.
= Bühlmann Credibility
𝑛𝑛 = Number of exposures observed Normal Random Variables: The Polar Method
𝜇𝜇 = Expected Hypothetical Mean EHM
𝑛𝑛= 𝑛𝑛¡ , 𝑛𝑛= < 𝑛𝑛¡ Transform uniform random numbers on 0, 1 , 𝑢𝑢R
𝑍𝑍 = = E E 𝑋𝑋 ∣ 𝜃𝜃
1, 𝑛𝑛= ≥ 𝑛𝑛¡ and 𝑢𝑢? , to uniform random number on −1, 1 :
𝑣𝑣 = Expected Process Variance EPV
𝑣𝑣R = 2𝑢𝑢R − 1, 𝑣𝑣? = 2𝑢𝑢? − 1
Bayesian Estimation and Credibility = E Var 𝑋𝑋 ∣ 𝜃𝜃
Density 𝑎𝑎 = Variance of Hypothetical Mean VHM 2 ln 𝑆𝑆
If 𝑆𝑆 = 𝑣𝑣R? + 𝑣𝑣?? ≤ 1, then 𝑇𝑇 = −
𝜋𝜋 𝜃𝜃 = Prior density = Var E 𝑋𝑋 ∣ 𝜃𝜃 𝑆𝑆
𝑓𝑓 𝑥𝑥 ∣ 𝜃𝜃 = Conditional density 𝑣𝑣 𝑛𝑛 𝑛𝑛𝑛𝑛 Standard normal random numbers: 𝑣𝑣R 𝑇𝑇 and 𝑣𝑣? 𝑇𝑇
𝑘𝑘 = , 𝑍𝑍 = =
𝑓𝑓 𝑥𝑥R , ⋯ , 𝑥𝑥d = Unconditional density 𝑎𝑎 𝑛𝑛 + 𝑘𝑘 𝑛𝑛𝑛𝑛 + 𝑣𝑣

𝑃𝑃fl = 𝑍𝑍𝑥𝑥 + 1 − 𝑍𝑍 𝜇𝜇 Number of Data Values to Generate


= 𝜋𝜋 𝜃𝜃 ⋅ 𝑓𝑓 𝑥𝑥R , ⋯ , 𝑥𝑥d ∣ 𝜃𝜃 d𝜃𝜃 (Continuous) Note: 𝑛𝑛 is the denominator of the 𝑥𝑥 term used to • Number of runs, 𝑛𝑛, needed for sample mean to
calculate 𝑃𝑃fl . be within 100𝑘𝑘% of true mean with level of
= 𝜋𝜋 𝜃𝜃 ⋅ 𝑓𝑓 𝑥𝑥R , ⋯ , 𝑥𝑥d ∣ 𝜃𝜃 (Discrete)
confidence 𝑝𝑝 is such that:
™ Empirical Bayes Non-Parametric Methods 𝑠𝑠d d ?
𝜋𝜋 𝜃𝜃 ∣ 𝑥𝑥R , ⋯ , 𝑥𝑥d = Posterior density oõR 𝑥𝑥o − 𝑥𝑥
Uniform Exposures 𝑧𝑧 RÑ_ ? ≤ 𝑘𝑘𝑥𝑥, where 𝑠𝑠d? =
𝜋𝜋 𝜃𝜃 𝑓𝑓 𝑥𝑥R , ⋯ , 𝑥𝑥d ∣ 𝜃𝜃 Ì d 𝑛𝑛 𝑛𝑛 − 1
= 1
𝑓𝑓 𝑥𝑥R , ⋯ , 𝑥𝑥d 𝜇𝜇 = 𝑥𝑥 = 𝑋𝑋o≥ • Number of runs, 𝑛𝑛, needed for sample
𝑟𝑟𝑟𝑟 proportion to be within 100𝑘𝑘% of 𝐹𝐹 𝑥𝑥 with
𝑓𝑓 𝑥𝑥dÑR ∣ 𝑥𝑥R , ⋯ , 𝑥𝑥d = Predictive density oõR ≥õR

1
Ì d level of confidence 𝑝𝑝 is such that:
?
= 𝑓𝑓 𝑥𝑥dÑR |𝜃𝜃 ⋅ 𝜋𝜋 𝜃𝜃 ∣ 𝑥𝑥R , ⋯ , 𝑥𝑥d d𝜃𝜃 Continuous 𝑣𝑣 = 𝑋𝑋o≥ − 𝑋𝑋o
𝑟𝑟 𝑛𝑛 − 1 𝑧𝑧 RÑ_ ? 𝑉𝑉𝑉𝑉𝑉𝑉 𝐹𝐹 𝑥𝑥 ≤ 𝑘𝑘𝐹𝐹 𝑥𝑥 ,
oõR ≥õR
= 𝑓𝑓 𝑥𝑥dÑR |𝜃𝜃 ⋅ 𝜋𝜋 𝜃𝜃 ∣ 𝑥𝑥R , ⋯ , 𝑥𝑥d (Discrete) Ì
1 𝑣𝑣 𝐹𝐹 𝑥𝑥 1 − 𝐹𝐹 𝑥𝑥
™ 𝑎𝑎 = 𝑋𝑋o − 𝑋𝑋 ?
− where 𝑉𝑉𝑉𝑉𝑉𝑉 𝐹𝐹 𝑥𝑥 =
Bayesian credibility estimate of 𝑥𝑥dÑR 𝑟𝑟 − 1 𝑛𝑛 𝑛𝑛

oõR
= Mean of predictive distribution Non-uniform Exposures Applications
= E 𝑥𝑥dÑR 𝑥𝑥R , ⋯ , 𝑥𝑥d Ì dº
If 𝑦𝑦≥ = 𝑗𝑗 »… order statistic from the sample and 𝑘𝑘 =
oõR ≥õR 𝑚𝑚o≥ 𝑋𝑋o≥
𝜇𝜇 = 𝑋𝑋 = 𝑛𝑛𝑛𝑛 + 1, then:
Loss Functions 𝑚𝑚 d
Bayesian Ì dº ? oõ< 𝑦𝑦o
Type of Loss 𝒍𝒍(𝜽𝜽, 𝜽𝜽) oõR ≥õR 𝑚𝑚 o≥ 𝑋𝑋o≥ − 𝑋𝑋 o VaR _ 𝑋𝑋 = 𝑦𝑦< , TVaR _ 𝑋𝑋 =
Estimate, 𝜽𝜽 𝑣𝑣 = Ì 𝑛𝑛 − 𝑘𝑘 + 1
oõR 𝑛𝑛o − 1 ? ?
? Mean of Ì ? 𝑠𝑠_ + 𝑝𝑝 TVaR _ 𝑋𝑋 − VaR _ 𝑋𝑋
Squared-error 𝜃𝜃 − 𝜃𝜃 oõR 𝑚𝑚o 𝑋𝑋o − 𝑋𝑋 − 𝑣𝑣 𝑟𝑟 − 1 Var TVaR _ 𝑋𝑋 =
posterior 𝑎𝑎 = 𝑛𝑛 − 𝑘𝑘 + 1
Median of 𝑚𝑚 − 𝑚𝑚'R ÌoõR 𝑚𝑚o? d
𝜃𝜃 − 𝜃𝜃 1 ?
Absolute Note: If 𝑎𝑎 < 0, then set 𝑎𝑎 = 0 and 𝑍𝑍 = 0. where 𝑠𝑠_ ? = 𝑦𝑦o − TVaR _ 𝑋𝑋
posterior 𝑛𝑛 − 𝑘𝑘
1 if 𝜃𝜃 ≠ 𝜃𝜃 Mode of Empirical Bayes Semi-Parametric Methods
oõ<
Zero-one
0 if 𝜃𝜃 = 𝜃𝜃 posterior Estimating 𝜇𝜇 and 𝑣𝑣 Bootstrap Approximation

Model 𝜇𝜇 𝑣𝑣 𝜃𝜃: Parameter of distribution 𝐹𝐹
𝑋𝑋o ~Poisson 𝜆𝜆 𝑥𝑥 𝑥𝑥 𝑔𝑔: Estimator based on sample of size 𝑛𝑛

𝑋𝑋o ~Neg. Bin. 𝑟𝑟, 𝛽𝛽 𝑥𝑥 𝑥𝑥 1 + 𝛽𝛽 The bootstrap approximation of the mean square
𝑋𝑋o ~Gamma 𝛼𝛼, 𝜃𝜃 𝑥𝑥 𝑥𝑥𝜃𝜃 error of 𝑔𝑔 as an estimator of 𝜃𝜃 is:
MSEÚ 𝜃𝜃 = Eí 𝑔𝑔 − 𝜃𝜃 ?

Copyright © 2017 Coaching Actuaries. All Rights Reserved. 3


www.coachingactuaries.com Copyright
Personal © 2017
copies Coaching
permitted. Actuaries.
Resale All Rights
or distribution Reserved. 3
is prohibited.

You might also like