You are on page 1of 6

Exam S

Raise Your Odds with Adapt

PROBABILITY REVIEW
PROBABILITY REVIEW
PROBABILITY REVIEW
PROBABILITY REVIEW PARAMETER ESTIMATION
PARAMETER ESTIMATION
PARAMETER ESTIMATION PARAMETER ESTIMATION HYPOTHESIS TESTING
HYPOTHESIS TESTING
HYPOTHESIS TESTING HYPOTHESIS TESTING
Probability Functions
Probability Functions
Probability Functions Estimator Quality
Estimator Quality
Estimator Quality Hypothesis Testing
Hypothesis Testing
Hypothesis Testing
&& &
Bias Bias
Bias Type I error: Reject
Type I error: Reject
Type I error: Reject even though
even though
even though is true
is true
is true
=== d dd
==Pr =PrPr
'('( '(
bias
biasbias
vv v ==E=EE Type II error: Accept
Type II error: Accept
Type II error: Accept even though
even though
even though RR is true
is true
R is true
==1=11 ==Pr =PrPr > >> An estimator is asymptotically unbiased if the bias =
An estimator is asymptotically unbiased if the bias
An estimator is asymptotically unbiased if the bias =Pr
=PrPr Type I error
Type I error
Type I error==significance level =significance level
significance level
of an estimator goes to 0 as the sample size goes to =
of an estimator goes to 0 as the sample size goes to
of an estimator goes to 0 as the sample size goes to =Pr
=PrPr Type II error
Type II error
Type II error
=== infinity.
infinity.
infinity. 111 = =Pr(=Pr( Pr(Reject
Reject
Reject RR is true
is true
R is true ==power =power
power

&& & Consistency
Consistency
Consistency -value: Probability of the observations if
-value: Probability of the observations if
-value: Probability of the observations if is true
is true
is true
=== d dd
== =ln lnln ; ;; ==='8
'8'8
&& & is consistent if:
is consistent if:
is consistent if: Confidence Intervals and Sample Size
Confidence Intervals and Sample Size
Confidence Intervals and Sample Size
'( '( '( ?? ? ?? ?
lim
lim lim
PrPrPrbb If ~(,
If ~(,
If ~(, ) and
) and
) and is known, then the
is known, then the
is known, then the 111
%
% %
Moments
Moments
Moments b
<<< ==1 for all
=1 for all
1 for all >>0 >0 0
b(
b( b(
(( ( confidence interval for is
confidence interval for is
confidence interval for is /? /? where
/? where
where
EEE === d
d d Sufficient but not necessary for consistency:
Sufficient but not necessary for consistency:
Sufficient but not necessary for consistency: bb b

bias
bias bias 0 and Var
0 and Var
0 and Var 0 as
0 as
0 as Pr (
Pr (
Pr ( <<</? /?
/? ))== ) /2.
=/2.
/2.
'('( '(
kk th thk raw moment:
raw moment:
th raw moment: == =
<<= =E=EE<< <
< Relative Efficiency
Relative Efficiency
Relative Efficiency Confidence Intervals for Means
Confidence Intervals for Means
Confidence Intervals for Means
?? ?
kk th thk central moment:
central moment:
th central moment: = = E=E E
<< <
The relative efficiency of
The relative efficiency of
The relative efficiency of distribution
distribution
distribution
<< < RR with respect to
with respect to
R with respect to ?? ?
Variance:
Variance:
Variance: ?? ?
===??== ?E =E E
?? = ?
=E=EE?? ?
?? ? If
If
If ===iR ?? ?
i i, has a
, has a ?? ?
i , has a distribution. distribution.
distribution.
Var
Var Var ??? iR iR
=== For
For
For i ~
i ~i ~
?? ?
,,,
: : :
Coefficient of variation:
Coefficient of variation:
Coefficient of variation: === VarVarVar RRR
biR b b
iR iRi i
i
?? ???~ ? ?? ?
~ ~
SSS Mean Square Error
Mean Square Error
Mean Square Error bb b


?? ? ?? ? ?? ?
~
~ ~ 1 1 1
Skewness:
Skewness:
Skewness: = =
RR R SS S = ? ? ? ?? ? iR
iR iRi i i
MSE
MSE MSEvv v ==E=E E ==Var =Var Var
+++bias bias biasvv v Students t distribution
Students t distribution
Students t distribution
VVV
Kurtosis:
Kurtosis:
Kurtosis: ??= =
? =VV V Uniformly Minimum Variance Unbiased Estimator has a distribution if
Uniformly Minimum Variance Unbiased Estimator
Uniformly Minimum Variance Unbiased Estimator has a distribution if
has a distribution if === with
with
with
?? ?
Covariance: Cov
Covariance: Cov
Covariance: Cov ,
,, ==E=E E
YYY ZZZ (UMVUE)
(UMVUE)
(UMVUE) ~(0,1) and ~
~(0,1) and ~
~(0,1) and ~ ().
().
().
==E=E E E EEEEE A UMVUE is an unbiased estimator that has the
A UMVUE is an unbiased estimator that has the
A UMVUE is an unbiased estimator that has the Testing the mean of a normal population
Testing the mean of a normal population
Testing the mean of a normal population
Cov
Cov Cov ,,
,
lowest variance of any unbiased estimator
lowest variance of any unbiased estimator
lowest variance of any unbiased estimator For a sample of size from a normal population,
For a sample of size from a normal population,
For a sample of size from a normal population,
Correlation coefficient:
Correlation coefficient:
Correlation coefficient: YZYZYZ === regardless of the true value of the estimated
regardless of the true value of the estimated
regardless of the true value of the estimated the statistic to test if the mean is is:
the statistic to test if the mean is is:
the statistic to test if the mean is is:
YYZYZZ
parameter.
parameter.
parameter.

Percentiles
Percentiles
Percentiles ===
The 100p
The 100p
The 100p percentile,
percentile,
percentile, Maximum Likelihood
Maximum Likelihood
Maximum Likelihood
]], is the lowest point where
, is the lowest point where
] , is the lowest point where
thth th
Steps to Calculating MLE:
Steps to Calculating MLE:
Steps to Calculating MLE:

=== 1 1 1
]]]. If is continuous,
. If is continuous,
. If is continuous, ]] is a unique point
is a unique point
] is a unique point 1.1.1. === 3.3.Set
3.Set == =
Set ==0 =0 0
satisfying
satisfying
satisfying ]]]==. =. . Testing the mean of a Bernoulli population
Testing the mean of a Bernoulli population
Testing the mean of a Bernoulli population
2.2.2. ==ln =lnln 4.4.Solve for
4.Solve for
Solve for For a sample of size from a Bernoulli population
For a sample of size from a Bernoulli population
For a sample of size from a Bernoulli population
Mode
Mode Mode The maximum likelihood estimate equals the
The maximum likelihood estimate equals the
The maximum likelihood estimate equals the with mean , the statistic to test if
with mean , the statistic to test if
with mean , the statistic to test if === is:
is:
is:
Mode is x that maximizes
Mode is x that maximizes
Mode is x that maximizes .
. . method of moments estimate for:
method of moments estimate for:
method of moments estimate for:
Moment Generating Function
Moment Generating Function
Moment Generating Function Binomial with fixed
Binomial with fixed
Binomial with fixed ===
YY Y = =E=EEaYaY aY
Poisson
Poisson
Poisson 111
Y
bb b
==E=EEbb, where
b bb b : number of successes
: number of successes
: number of successes
Y Y00 0 , where
, where is the n
is the n
is the n derivative
derivative
th derivative
thth
Negative binomial with fixed
Negative binomial with fixed
Negative binomial with fixed
Conditional Probability
Conditional Probability
Conditional Probability Testing the difference of means from two
Testing the difference of means from two
Testing the difference of means from two
Exponential
Exponential
Exponential
PrPr
Pr populations
populations
populations
Gamma with fixed
Gamma with fixed
Gamma with fixed
PrPrPr ===

If the two populations and are normally
If the two populations and are normally
If the two populations and are normally
Pr PrPr Normal
Normal
Normal
,
,,
distributed with unknown variances that are
distributed with unknown variances that are
distributed with unknown variances that are
Variance of Maximum Likelihood Estimator
Variance of Maximum Likelihood Estimator
Variance of Maximum Likelihood Estimator

YY Y === assumed to be equal, the test statistic is:


assumed to be equal, the test statistic is:
assumed to be equal, the test statistic is:
Fisher information matrix
Fisher information matrix
Fisher information matrix
1.1.Two unpaired normal populations
1.Two unpaired normal populations
Two unpaired normal populations
Bayes Theorem
Bayes Theorem
Bayes Theorem ??lnln ?
ln
x;
x;x;
?? ?;;;

I I
I ==E =EE ==E =E E

PrPr
Pr PrPrPr d
dd
?? ? d dd
? ? ? = == , where
, where
, where
PrPrPr === 11 1 11 1
PrPrPr Rao-Crame
Rao-Crame
Rao-Crame rr lower bound for the variance
lower bound for the variance
r lower bound for the variance +++
YYY ZZZ
ZZ Z
YY Y of an estimator
of an estimator
of an estimator

YY Y === 'R'R'R YYY1 1 1Y?Y?+Y+


?
+ZZ
Z1 1 1Z?Z?Z?

ZZ Z
Var
Var Var = = =I I
I ?? ?
===
Law of Total Probability
Law of Total Probability
Law of Total Probability Sufficient Statistics
Sufficient Statistics
Sufficient Statistics YY+Y++ ZZ
Z2 22
Discrete
Discrete
Discrete

= =
=
+ +
+
2
2 2
A statistic is sufficient for a parameter if the
A statistic is sufficient for a parameter if the
A statistic is sufficient for a parameter if the YY Y ZZ Z
2.2.Two paired normal populations
2.Two paired normal populations
Two paired normal populations
Pr PrPr === Pr PrPr i i= = = PrPr Pr PrPr Pr



likelihood of the sample, given , does not
likelihood of the sample, given , does not
likelihood of the sample, given , does not
i ii i ii i
ii i ii i
depend on the parameter:
depend on the parameter:
depend on the parameter: = == , where
, where
, where
Continuous
Continuous
Continuous
,

,
RR R bb b,
;; ;

= =
=
,

, ,
RR R bb b
A statistic is sufficient for a parameter if and
A statistic is sufficient for a parameter if and
A statistic is sufficient for a parameter if and
i i i i i i i i i
= =
=



Pr PrPr === Pr PrPr d d d only if the following factorization is possible:
only if the following factorization is possible:
only if the following factorization is possible: = =sample mean of the
=sample mean of the
sample mean of the i s
i s i s

Conditional Expectation Formula


Conditional Expectation Formula
Conditional Expectation Formula RR,
, R,
bb;;b; == = RR,
, R,
bb b ==sample standard deviation of the
=sample standard deviation of the
sample standard deviation of the i s
i s
i s

EEYYEY ==E=EZZEEZEYYEY Rao-Blackwell Theorem: For an unbiased


Rao-Blackwell Theorem: For an unbiased
Rao-Blackwell Theorem: For an unbiased

= =
=
1
1 1
Variances
Variances
Variances estimator and a sufficient statistic ,
estimator and a sufficient statistic ,
estimator and a sufficient statistic , is is For Bernoulli data for two samples, the
is For Bernoulli data for two samples, the
For Bernoulli data for two samples, the
an unbiased estimator with variance that is less
an unbiased estimator with variance that is less
an unbiased estimator with variance that is less statistic to test for the difference of means of
statistic to test for the difference of means of
statistic to test for the difference of means of
Variance of the Sample Mean
Variance of the Sample Mean
Variance of the Sample Mean
than or equal to
than or equal to
than or equal to . . . two populations with
two populations with
two populations with YY and
and
Y and ZZ successes
successes
Z successes
For independent and identically distributed
For independent and identically distributed
For independent and identically distributed
respectively is:
respectively is:
respectively is:
(i.i.d.) random variables, the variance of the
(i.i.d.) random variables, the variance of the
(i.i.d.) random variables, the variance of the The maximum likelihood estimator is a function of
The maximum likelihood estimator is a function of
The maximum likelihood estimator is a function of YYY ZZZ
lmn lmn lmn
YY Y a sufficient statistic.
a sufficient statistic.
a sufficient statistic.
sample mean is Var
sample mean is Var
sample mean is Var === . . . YYY ZZZ
bb b
=
= = , where
, where
, where
Bernoulli Shortcut
Bernoulli Shortcut
Bernoulli Shortcut 11 1 11 1
111 +++
probability
probability
probability ==
=
YYY ZZZ
= == YY+Y++ ZZZ
probability
probability
probability ==1=11 ==overall sample mean
=overall sample mean
overall sample mean ===
Var Var Var === ???111 YY+Y++ ZZZ
Conditional Variance Formula
Conditional Variance Formula
Conditional Variance Formula
Var VarVarYY Y ==Var =Var VarZZ E ZEYYE Y ++E+ EZZEVar ZVar Var
YY
Y

www.coachingactuaries.com Copyright 2016 Coaching Actuaries. All Rights Reserved. 1


Chi Square Tests NONPARAMETRIC METHODS FOR
NONPARAMETRIC METHODS FOR HYPOTHESIS For the normal approximation, evaluate =
The chi-square test statistic is: TESTING HYPOTHESIS TESTING
'
, where
<
i i ?
<
i? Order Statistics

<'R = = R + ? + 1
The density function of the order statistic is: E R = R
i i ! 2
iR iR <'R b'<
Z = 1 Y Y () R + ? + 1
= 1 1 ! ! Y Var R = R ?
= number of groups
The expected value of the order statistic from a 12
When testing equality of distributions in two sample of from a uniform distribution on 0, is The Mann-Whitney statistic is the sum of the
dimensions, = 1 ( 1) where is the < counts of the number of items in the first sample
.
number of rows and is the number of columns. bR less than each item in the second sample. There
Sign Test should be ? terms in the sum.
Confidence Intervals for Variances
The confidence level that the 100pth percentile is The relationship between and the rank-sum
The unbiased sample variance from a normal
between the th and th order statistics from a statistic is
population, ? , has the distribution sample of size is equal to Pr < , where
? 1 , where ~ ? 1 . ? ? + 1
~Binomial , . = ?
Confidence intervals for the variances at the 2
To test if the median is equal to , subtract from R R + 1
100 1 % confidence interval are: each sample item to produce positive numbers = R ? + R
1. 0, ? 1 2
after subtraction. For large sample sizes, calculate

2. ? 1 R' , <'b/? For the normal approximation,


the statistic = . Determine the -value using R ?
3. ? 1 R' ? , ? 1 ? b/?
E =
An (R , ? ) random variable is defined by the normal distribution. 2
R + ? + 1
R R ? ? , R ~ ? (R ) and ? ~ ? ? . To test if two populations have the same median Var = R ?
Uniformly Most Powerful Critical Regions 12
based on pairs of sample data i , i , calculate Rank Correlation Coefficients
A level test is uniformly most powerful if its the differences i = i i . Let be the number of
power is greater than or equal to the power of all Pearsons correlation coefficient
positive differences. For large sample sizes, b
other level tests. iR i i
calculate the statistic =
<'b/?
. Determine the - =
Likelihood Ratio Tests b/? b ? b
i
iR i ?
1. Calculate the maximum likelihood as a function value using the normal distribution. iR

of the observations under . Wilcoxon Tests b


iR i i
Signed Rank Test =
2. Calculate the maximum likelihood as a function
b ? b ?
of the observations under R . To test = , subtract from each sample item iR i ? iR i ?
3. The critical region is the set of observations for i : i = i . To test if means of two samples of Spearmans
which the ratio of the first expression over the paired data are equal, i.e. R = ? , calculate b
i i
iR
second expression is below a constant k. differences i = i i . In either case: =
This method reduces to Neyman-Pearson 1. Calculate the ranks of i . If there are ties, b
i ? b
i ?
iR iR
when and R are both simple. assign the average of the ranks to the tied group. b
2. The Wilcoxon signed rank statistic, , is the iR i i
=
sum of ranks of positive i . b ? b ?
BAYESIAN ESTIMATION

iR i ? iR i ?
BAYESIAN ESTIMATION
b ?
Bayesian Estimation Null and Alternate 6 iR i i
Test -value =1 if there are no ties
Bayesian Hypotheses ? 1
Type of Loss ,
Estimate, Two- : = ; R = ? Kendalls
Mean of min 2, 2 1
Squared-error
? sided R : ; R ? =
posterior : ; R ? 1 2
Median of qq-Plot
Absolute One- R : < ; R < ?
posterior sided : ; R ? Let x-coordinates be the fitted values and y-
0 if = Mode of 1 coordinates be the observed data. The fit is good if
Zero-one R : > ; R > ?
1 if posterior The probability under the aforementioned null the point-connected line is close to the 45 line.
Conjugate Priors hypotheses that the statistic equals , or , With n data points, k-th order statistic is
Poisson/Gamma is the coefficient of a in biR 1 + ia , divided by considered + 1 quantile.
If ~Poisson , and ~Gamma , = 1/ : 2b . If the line is steeper than 45, the fitted
= + distribution puts too little weight in that
Posterior Gamma = + For the normal approximation, evaluate = interval; vice versa.
'
Normal/Normal , where

If ~Normal , , and ~Normal , :
+ 1
+ E =
= 4
Posterior Normal + + 1 2 + 1
Var =
= 24
+ Rank Sum Test
Bernoulli/Beta To test if two populations with samples of sizes R
If ~Bernoulli , and ~Beta , : and ? have the same distribution, sum the ranks
= + of the first sample within the combined sample to
Posterior Beta
= + get R , the Wilcoxon rank sum statistic. Sum the
When the model is binomial , , treat it as a ranks of the second sample to get ? .
series of m Bernoullis.

The relationship between R and ? is


R + ? R + ? + 1
? = R

2

www.coachingactuaries.com Copyright 2016 Coaching Actuaries. All Rights Reserved. 2


DISCRETE MARKOV CHAINS
DISCRETE MARKOV CHAINS THE POISSON PROCESS
THE POISSON PROCESS CONTINUOUS MARKOV CHAINS
CONTINUOUS MARKOV CHAINS
Chapman-Kolmogorov Equations Probabilities of Events Notations
b
For a Poisson random variable with mean , i : The rate at which the process transitions from
ia = a'
i< i< ' b state i.
<R b = i : The probability that the process transitions to
!
Gamblers Ruin Probabilities In a homogeneous Poisson process with parameter state j from state i, given that it transitions.

, = 1 , the number of events over a time period is a i : The instantaneous transition rate from state i

= 'R , where = . Poisson random variable with mean . to state j, defined by i = i i .
, 1 In a non-homogeneous Poisson process with Relationship: bR i = i .
'R
intensity function (), the number of events

Algorithmic Efficiency between times and + is a Poisson random Birth and Death Processes
'R
variable with mean () d. A continuous Markov chain in which the states are
1
E = nonnegative integers and the only transitions
Thinning
possible are + 1 < and 1 < .
iR Given a Poisson process with intensity function
'i Rate of leaving state = < + < .
1 1 () and probability function () for a certain
Var = Expected time to stay in state = 1 < + <
1 type of an event:
Expected time starting in state k until process
iR That type of event follows a Poisson process R
Both E and Var approach ln j as . with intensity (). reaches state + 1 < = + <'R

Classification Of States They are independent of any other events For M/M/1 queuing system:
Absorbing: Cannot be exited. Homogeneous Poisson Processes Characteristics 1 <R
Accessible from state i to j: ib > 0 for some n. Given two Poisson processes with rate R and ? : ,
E < =
Two states communicate: each state is accessible The probability that the next event is from the + 1
, =
from the other. first process is

Class of states: A maximal set of states that Processes < <
The probability that k events from process 1
communicate with each other; an absorbing state Poisson 0
occur before l events from process 2 is
is a class by itself. <'R Yule 0
Irreducible chain: Only one class exists. + 1 i <'R'i M/M/1
1
Recurrent state: Probability of reentering the state , 1 <
i<
is 1. A finite Markov chain must have at least one M/M/s
where = R R + ? . ,
recurrent class.

Sums and Mixtures Limiting Probabilities


Transient state: A non-recurrent state.
Sum of independent Poisson random variables Transition probabilities for pure birth processes
Long-Run Proportions is a Poisson random variable. <
b b <,<R = ' a ' a
A linear combination of Poisson random <R <
i = i , i = 1 variables is not a Poisson random variable Kolmogorovs backward equations
R iR unless all the R are positive numbers. i= = i< < i i

Time In Transient States A mixture of Poisson random variables is not a <i


Expected time in transient states Poisson random variable; raw moments of the Kolmogorovs forward equations
= 'R mixture are weighted averages of individual
raw moments. i= = i< < i
Probability of ever transitioning from state i to j <i
i i, A gamma , mixture of Poisson random
i = Balance equations
variables is a negative binomial , with

= and = . < < =
For a 2-transient-state (1 and 2) Markov chain:
Compound Poisson Processes <i
R? General limiting probabilities for birth and death
R? =
1 RR = i processes

1
Branching Processes iR
=
A type of Markov chain representing the growth or E = E
1 + iR i<R <'R
(

extinction of a population. Var = E[ ? ] <


b
E b = b The sum of compound Poisson processes R , R i'R
and ? , ? is a compound Poisson process with b =
b'R ?b'R i
? , 1 parameter R + ? and secondary distribution a iR
Var b = 1
?
mixture with weight R R + ? on R and the
, = 1 balance on ? .

RELIABILITY RELIABILITY
1, 1
( Structures
= is the structure function for a system.
, > 1
= 1 if the system functions

= 0 otherwise.
Reverse-Time Markov Chain Series System
Q is the reverse-time Markov chain for ergodic P if = biR i
and only if i i = i . Parallel System

Monte Carlo Methods = max i = 1 biR 1 i


, Minimal Path Set
, = min ,1 A set of components for which the system
,
functions if all are functioning but not if one of the
components does not.
Minimal Cut Set
A set of components for which the system fails
if all fail but does if at least one of the
components functions.

www.coachingactuaries.com Copyright 2016 Coaching Actuaries. All Rights Reserved. 3


Probabilities LIFE CONTINGENCIES
LIFE CONTINGENCIES For a single variable classic linear model,
is the reliability function for a system. Survival Model i = + R i + i ,
If is a polynomial of the i s, then is &R = & & = R
the same polynomial of the i s. &b R =
& '& '
= &
b & = = & b' & & '&
Inclusion/Exclusion bounds &
Using minimal path sets: & &b & is the sample covariance of x and y
b & = 1 b & = & , are the sample standard deviations of x
&
and y.
i E &b & = = b&
&b & = = b& b & and R are the least squares estimates of
( and R .
i i
& = E & = < &
i = i i is called the residual. By using the
+ <R and R solved above, biR i = 0.
Contingent Payments

i i i<
( For a multiple variables classic linear model,
Using minimal cut sets: & = <R < & &< = & + & &R = = 'R =
1 1 <
=
(

i 1 & For a weighted least squares model,


& = < < & = 1 + & &R =
1 1 1 = = 'R =
<
i i & 1 Measures of Fit
= = : Number of parameters (not includes constant)
1 1 1 & &
Error Sum of Squares (SSE)
i i ?
SSE = iR i i and has df = + 1 .
+ 1 CLASSICAL LINEAR MODEL
CLASSICAL LINEAR MODEL Regression Sum of Squares (SSR)
?
i< Variables SSR = iR i i and has df = .
Bounds Using Intersections Response: the variable we are trying to predict Total Sum of Squares (SST)

Explanatory Variables: variables used to predict SST = ?
iR i i = SSE + SSR and has df =
1 1 1 1 response 1 .
iR iR
Three Types of Variables (and examples) Coefficient of Determination ( ? )
Time to Failure Continuous variables: claim amount; time to The proportion of the sum of squares explained by
Ross used for survival function instead of settlement of claim. the regression.
; he also used the terminology failure rate Categorical variables: Bernoulli variables (yes/no); SSR SSE ?
function instead of commonly known hazard auto usage; gender. ? = =1 = Y,Z
SST SST
rate function . Count variables: number of claims Y,Z correlation coefficient of and
(

E System Life = Graph Types Standard Error of the Regression ()


Scatter Plot: Examine a continuous variable ( axis) = SSE + 1
For k-out-of-n systems whose components have as an explanation of another continuous variable ( ? = ?
identical distribution exp , axis).
b : diagonal element of 'R
1 Box Plot: Examine a categorical variable as an
= t Statistic
explanation of a continuous variable with a
Test the significance of estimator = .
i< rectangle above each category.
Failure Rate Function Frequency table and Mosaic plot: Examine a
b' ]R =
= categorical variable as an explanation of a
categorical variable. F Statistic
Hazard Function qq-plot and pp-plot: Assess the reasonability of the Test the significance of entire regression.
a
= d distribution assumption in GLM. SSR + 1 ?
Runoff Triangles: Present the development of ],b' ]R = =
SSE + 1 1 ?
Increasing Failure Rate Distribution (IFR) claims which may take years to resolve.

is a non-decreasing function of t.
Test the significance of additional variables
Data Issues SSE SSE
Decreasing Failure Rate Distribution (DFR) ,b' ]R =
Censoring: Have information on the number of
is a decreasing function of t. SSE + 1
claims above or below a certain number, but dont ?
Increasing Failure On The Average (IFRA) ?
know their amounts. It results in under- =
/ is a non-decreasing function of t. ?
1 + 1
representation of large claims.
The monotone systems lifetime distribution Variance Inflation Factor (VIF)
Lack of Independence: Multiple exposures are
is IFRA if the lifetime of all components are Test collinearity of variable j.
correlated. Eg. Multiple claims from a single
IFRA. ? VIF ? VIF
accident.

Exposure Differences: Factors that may affect ? = ? = ?
& 1 b

exposures must be taken into account. iR i

Length bias: The counterpart of censoring: less


aggressive claims are not detected, which results in
over-representation of large claims.
Estimating Parameters
Assumptions of classical linear models
Homoscedasticity: i = 0; i = ?
Normality: i is normally distributed
Linearity: dependent variables are linearly related
to coefficients of the model
No correlation: i are uncorrelated

www.coachingactuaries.com Copyright 2016 Coaching Actuaries. All Rights Reserved. 4


Model Selection/Comparison GENERALIZED LINEAR MODEL (GLM)
GENERALIZED LINEAR MODEL (GLM) Measures of Fit
F statistic can be used to compare a model with a Generalized linear model extends the classical i satisfies i = i .
set of variables to another model with a subset of linear model in two ways: i satisfies i = i .
that same set of variables. The response variable does not need to be

Adjusted ? ? normal. Saturated Model


The higher, the better. ]
The linear expression i is equal to a The model that each is matched perfectly. If are
SSE + 1 ? function of the mean of the response variable, , the fitted betas for this model, = = .
? = 1 = 1 ?
SST 1 , instead of the mean itself. Deviance
? : Sample variance of i . Exponential Family b
i i i i + i
Log-likelihood = 2
= , exp
+ 1 ? iR
= ln 0.5ln 2 E = = In SAS, is called scaled deviance; is called
2 ?
In practice, ? would be estimated by ? . Var = the deviance.


V = Likelihood Ratio Test
Akaike Information Criterion Common exponential family distributions
AIC = 2 + 2 Likelihood ratio statistic = 2
Bayesian Information Criterion : Log-likelihood of the unconstrained model
BIC = 2 + ln() ln
$
v : Log-likelihood of the constrained model
Binomial ln 1 + 1
Where = + 2 in a linear model R'$
Likelihood ratio statistic follows a chi-square
Model Validation Poisson ln v
1 with q degrees of freedom to test significance.
Leverage: ii 1

Leverage in 1-variable Regression Neg. Bin. ln ln 1 v 1 Wald Test


1 +
1 i ? = 'R
= 'R

ii = + Normal 1 2 ? ?
?

b
Gamma 1 ln 1 Wald statistic reduces to for testing a single
R (1
Estimated Variance of Residuals coefficient = 0.
Var i = 1 ii ? Inv. Gauss. 1 2 ? Deviance Residuals

Studentized/Standardized Residuals 2 ? Test whether specific data points are unusual.
i = i 1 ii Estimation 2 i i i i + i
Outlier: A point with unusually high residuals ]
Link: = i i =
High-leverage points: Diagonal elements ii that
Canonical Link: The for which = or a
are more than 2 or 3 times the average value. The sign on the square root is taken as the sign
constant multiple of .
Cooks distance: Combines outliers and high- of i i .
How To Generate
leverage points into a single measurement.
? Given estimated coefficients and a set of values Anscombe Residuals
b
R i ii for the variables in a GLM: i i
i = = i ? ]
+ 1 ? + 1 1 ii 1. Generate = i . i i
Sum of Squared Prediction Errors (SSPE) 'R
2. Apply to the result to obtain . where = 'R S .
b
3. Calculate given the fact = .

SSPE = i i ? 4. Generate randomly from the parameters of Models for Counts


ib R the response distribution. Quasi-likelihood
Predicted Residual Sum of Squares (PRESS)
Used for model that otherwise would be modeled
b b
i ? Canonical link using Poisson or negative binomial distributions
? Response Link name
PRESS = i i = but its variance is greater than mean.
1 ii
iR iR Binomial ln Logit ( = 1) b
i i di
SSPE and PRESS are both used to compare models. Poisson ln Log =
The lower SSPE/PRESS is, the better. SSPE/PRESS i d
Neg. Bin. ln 1 + iR
value has no absolute meaning. Normal Identity

Models for Continuous Responses


Gamma 1 Power Method
Inv. Gauss. 1 2 ? Power Use GLM and utilize gamma or inverse Gaussian in
Estimation: Fisher Scoring the exponential family with a log link.
Let be the diagonal matrix with entries

'R Tweedie Distribution


i ? i and be the diagonal matrix with A compound distribution with Poisson claim
entries i . counts and gamma claim sizes.

' ' Has a point mass at 0
Score Vector:
( Continuous on positive numbers
' A member of exponential family
Information Matrix:
( Var = ] with 1 < < 2
Iterative Formula Binary Responses
< = <'R + = 'R = Logit link (canonical)
Local Dependent Variable
= + = ln
1
Iterative Formula in Iterative Weighted Least Probit link
Squares = 'R
<R = = 'R = Complementary log-log link
= ln ln 1

www.coachingactuaries.com Copyright 2016 Coaching Actuaries. All Rights Reserved. 5


Three Goodness of Fit Measures for Logistic Autocovariance (acvf) Autoregressive Models
Deviance < = a a< Assume the time series have a mean of 0:
Not a good measure. Autocorrelation (acf) AR = a = R a'R + ? a'? + + ] a'] + a
i. Does not directly depend on data < If the mean is not 0:
< = ?
ii. Not approximately chi-square AR = a = R a'R + ? a'?
iii. Depends on whether data are grouped Sample acvf +
b'<
Chi-square 1 + ] a'] + a
Not reliable. < = a a<
b
Alternate Expression
aR
i i ? Sample acf
a = ] a , where ] is the characteristic
=
i 1 i < polynomial.
iR < = Equations for stationary AR(1) processes
ROC curve and AUC
Cross-covariance of time series (ccvf) < = 0
A Receiver Operating Characteristic (ROC) curve
< , = a< & a < ?
plots sensitivity (y) against the complement < =
of specificity (x).

1 ?
Cross-correlation of time series (cvf) < = <
Sensitivity: Relative frequency of predicting an < ,
< , = Moving Average Models
event when it occurs. & MA = a = a + R a'R + + a'
Specificity: Relative frequency of predicting a
Sample ccvf Mean = 0
non-event when an event does not occur. b'<
Area Under the Curve (AUC) is 1 for a perfect 1 Always Stationary
< , = a a Var a = Var a + iR i? Var a'i

fit and 0.5 if the ROC is the 45-degree line.
aR ? ?

= i i
Ordinal Responses Sample cvf
< , Autocovariance
Model cumulative distribution by creating bands '<
< , =
i'R , i ; and then use cumulative versions of the , ,
< = ? i i<
three links for binary responses. White Noise
i
Alternate option: use continuation ratio A stationary times series in which each term is Autocorrelation
independent and has mean 0 and constant 1, = 0
= variance ? . '<
<R < R = R i i i<
< = , 1
Nominal Responses ? , = 0 i i
?

Set to the probability of base levels; for the < =


0, > 0 0, >
other 1 levels, build relative odds models 1, = 0 ARMA/ARIMA/SARIMA Models
(odds of class j relative to base level). < =
0, > 0 ARMA
$
ln $ = = , = 1, , 1. Random Walks ARMA , = a = R a'R + + ] a']
2 A non-stationary time series which is the +R a'R + + a' + a
accumulation of white noise.
Alternate Expression
R = R ] a = a
TIME SERIES TIME SERIES
R = a'R + a
A time series is a series of data indexed by time.
= 0 Equations for ARMA(1, 1) processes
The study of time series tries to find relationships
? = ? = < 1 + 2 + ?
between the values of the series in order to
a5 R = ?
forecast future values. < = = 1 ?
R</a
a5 a< 5 1 +
Correlation < = ? + <'R
For higher t, the correlogram will decrease 1 ?
Ergodic: The mean of the observed time series
slowly from 1 to 0. For any k, the correlogram will <'R + 1 +
approaches the theoretical mean as time goes < =
always decrease. 1 + 2 + ?
to infinity.
Autocorrelation: The correlation of a times series < = <'R for 2.

with itself. ARIMA
Lag: The distance between terms ARIMA , , : 1 7 a = a
Strictly Stationary: No moments of a time series It becomes ARI(p, d) with no moving
vary with t. average terms.

It becomes an IMA(d, q) model with no
autoregressive terms.
SARIMA (or ARIMA(p, d, q)(P, D, Q))
An ARIMA model which incorporates seasonal
effects by differencing the series with lag s

Copyright 2016 Coaching Actuaries. All Rights Reserved. 6


www.coachingactuaries.com Copyright
Personal 2016
copies Coaching
permitted. Actuaries.
Resale All Rights
or distribution Reserved. 6
is prohibited.

You might also like