beginners guide bayesian
TRANSCRIPT
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 1/30
A Beginner’s Guide to BayesianModelling
Peter England, PhD
EMB
GIRO 2002
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 2/30
Outline
• An easy one paraeter pro!le
• A harder one paraeter pro!le
• Pro!les "ith ultiple paraeters
• Modelling in #inB$G%
• %to&hasti& 'lais Reser(ing• Paraeter un&ertainty in D)A
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 3/30
Bayesian Modelling* General %trategy
• %pe&i+y distri!ution +or the data
• %pe&i+y prior distri!utions +or the paraeters
• #rite do"n the oint distri!ution
• 'olle&t ters in the paraeters o+ interest• Re&ognise the -&onditional. posterior distri!ution/
1es* Estiate the paraeters, or saple dire&tly
o* %aple using an appropriate s&hee
• )ore&asting* Re&ognise the predi&ti(e distri!ution/ 1es* Estiate the paraeters
o* %iulate an o!ser(ation +ro the data distri!ution,&onditional on the siulated paraeters
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 4/30
A One Paraeter Pro!le
• Data %aple 34,5,6,7,6,5,8,5,9,4:
• Distri!uted as a Poisson rando (aria!le/
• $se a Gaa prior +or the ean o+ the
Poisson
• Predi&ting a ne" o!ser(ation/
• egati(e Binoial predi&ti(e distri!ution
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 5/30
Poisson E;aple < Estiation
( )
( )( )
λθ α α θ
θ α
λ θ θ θ
λ α θ
θ
−−
=
−
Γ
=∝
Γ
∏ e y
e y f y f
IPoi y
n
i i
y
i
i<
< =.,->
,?
.-?
( )θ λ α θ n y ei +−−+∑∝ <
( )n yi ++Γ ∑ λ α θ ,?
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 6/30
Poisson E;aple < Predi&tion
( )
<
<<
<<
?
<<
<
<
<
< ,
.,-Binoial egati(e??
,
<
<
<.<?-.-
.?-,>?
<
λ
λ α
λ λ α α
λ λ
λ
α
α θ
α
+==
+=+=
+
++Γ Γ
+Γ =
∑
p x
p x y
n y
y
y y y f
i
y
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 7/30
One Paraeter Pro!le*
%iple 'ase
• #e &an re&ognise the posterior distri!ution
o+ the paraeter • #e &an re&ognise the predi&ti(e distri!ution
• o siulation re@uired
• -#e &an use siulation i+ "e "ant to.
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 8/30
aria!ility o+ a +ore&ast
• In&ludes estiation (arian&e and pro&ess(arian&e
• Analyti& solution* estiate the t"o &oponents
• Bayesian solution* siulate the paraeters, then
siulate the +ore&ast &onditional on the paraeters
2<
(arian&e.estiation(arian&e-pro&esserror predi&tion +=
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 9/30
Main )eatures o+ Bayesian
Analysis• )o&us is on distri!utions -o+ paraeters or
+ore&asts., not ust point estiates• he ode o+ posterior or predi&ti(e
distri!utions is analogous to Ca;iu
lielihood in &lassi&al statisti&s
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 10/30
One Paraeter Pro!le*
Farder 'ase• $se a log lin !et"een the ean and the
paraeter, that is*
• $se a noral distri!ution +or the prior
• #hat is the posterior distri!ution/
• Fo" do "e siulate +ro it/
Mean eθ
=
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 11/30
Poisson E;aple 2 Estiation
( )
( )
( )
2
2<
2
2
<
densitylog
2<
=.,->
.,-?
.-?
2
2
2
2
−
−−∝
∑∝
=∝
∑
∏
−−
−
−−
=
−
σ
µ θ θ
π σ θ θ
σ µ θ
θ
σ
µ θ
θ
σ
µ θ θ
θ
θ
θ
ne y
eee
e y
ee y f y f
N
e IPoi y
i
ne
y
n
i i
e y
i
i
i
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 12/30
Poisson E;aple 2
• %tep <* $se adapti(e ree&tion sapling
-AR%. +ro log density to saple the
paraeter
• %tep 2* )or predi&tion, saple +ro a
Poisson distri!ution "ith ean , "ith
theta siulated at step <
θ e
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 13/30
A MultiParaeter Pro!le
• )ro %&ollni -AAH, 200<.
• 4 Group "orers &opensation poli&ies
• E;posure easured using payroll as a pro;y
• u!er o+ &lais a(aila!le +or ea&h o+ last
8 years
• Pro!le is to des&ri!e &lai +re@uen&ies in
the +ore&ast year
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 14/30
%&ollni E;aple <
Year Group 1 Group 2 Group 3 Group 1 Group 2 Group 3 Group 1 Group 2 Group 3
1 280 260 9 6 0.032 0.023
2 320 275 145 7 4 8 0.022 0.015 0.055
3 265 240 120 6 2 3 0.023 0.008 0.025
4 340 265 105 13 8 4 0.038 0.030 0.038
5 285 115
Average 0.029 0.019 0.039
Payroll P(i,j) lai!" #(i,j) Pro$a$ili%ie"
.<,26-?
.6,6-?
.,-?
.-?
Gamma
Gamma
Gamma
P Poi X
i
iijij
β
α
β α θ
θ
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 15/30
%&ollni E;aple <
Posterior Distri!utions
( )
( )
( )
( )
α α α
α θ α
β β θ θ θ α
θ α α θ θ θ β
β α β α θ θ θ
β α β α θ θ θ
β α β α θ θ θ
684
<
4
42<
42<
442<4
224<2
<<42<
.-.,,,,>-
<,284?,,,,>
,<?,,,,>
,<?,,,,>
,<?,,,,>
−
=
Γ ∝
++
+++
+++
+++
∏
∑∑ ∑∑ ∑
∑ ∑
e X f
Gamma X
P X Gamma X
P X Gamma X
P X Gamma X
i
i
i
j j
j j
j j
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 16/30
%&ollni E;aple <
• $se Gi!!s %apling
Iterate through ea&h paraeter in turn
%aple +ro the &onditional posterior
distri!ution, treating the other paraeters as
+i;ed
• %apling is easy +or• $se AR% +or
β θ θ θ ,,, 42<
α
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 17/30
#inB$G%
• #inB$G% is an e;pert syste +or Bayesian analysis
• 1ou spe&i+y he distri!ution o+ the data
he prior distri!utions o+ the paraeters• #inB$G% "ors out the &onditional posterior
distri!utions
• #inB$G% de&ides ho" to saple the paraeters
• #inB$G% uses Gi!!s sapling +or ultiple paraeter pro!les
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 18/30
%to&hasti& 'lais Reser(ing
• 'hanges the +o&us +ro a C!est estiate o+ reser(esto a predi&ti(e distri!ution o+ outstanding lia!ilities
• Most sto&hasti& ethods to date ha(e only&onsidered 2nd oent properties -(arian&e. in
addition to a C!est estiate• Bayesian ethods &an !e used to in(estigate a +ull
predi&ti(e distri!ution, and in&orporate udgeent-through the &hoi&e o+ priors.
• )or ore in+oration, see England and errall -BAH,2002.
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 19/30
he Bornhuetter)erguson Method
• $se+ul "hen the data are unsta!le
• )irst get an initial estiate o+ ultiate
• Estiate &hainladder de(elopent +a&tors
• Apply these to the initial estiate o+
ultiate to get an estiate o+ outstanding
&lais
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 20/30
'on&eptual )rae"or
P r e & i ' % i v e i " % r i $ u % i o
* a r i a $ i l i % y( P r e & i ' % i o + r r o r )
, e " e r v e e " % i ! a % e
( - e a " u r e o . l o ' a % i o )
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 21/30
10000 14000 18000 22000 26000 30000 34000
/o%al ,e"erve"
igure 1. Pre&i'%ive Aggrega%e (i"%ri$u%io) o /o%al ,e"erve"
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 22/30
Estiates o+ outstanding &lais
o estiate ultiate &lais using the &hain ladder te&hni@ue, you"ould ultiply the latest &uulati(e &lais in ea&h ro" !y f , a
produ&t o+ de(elopent +a&tors
Fen&e, an estiate o+ "hat the latest &uulati(e &lais should !e iso!tained !y di(iding the estiate o+ ultiate !y f %u!tra&ting this
+ro the estiate o+ ultiate gi(es an estiate o+ outstanding
&lais*
<Estiated $ltiate <
f
× − ÷
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 23/30
he Bornhuetter)erguson Method
Jet the initial estiate o+ ultiate &lais +or
a&&ident year i !e
he estiate o+ outstanding &lais +or a&&ident
year i is
−
+−+− ninin
i M
λ λ λ
42
<<
( )<<
42
42
−= +−+−+−+−
ninin
ninin
i M λ λ λ
λ λ λ
i M
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 24/30
'oparison "ith 'hainladder
repla&es the latest &uulati(e
&lais +or a&&ident year i, to "hi&h the usual &hainladder
paraeters are applied to o!tain the estiate o+ outstanding
&lais )or the &hainladder te&hni@ue, the estiate o+ outstanding
&lais is
ninin
i M
λ λ λ
42
<
+−+−
( )<42<, −+−+−+− nininini D λ λ λ
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 25/30
Multipli&ati(e Model +or 'hainJadder
( )<
? - .
- .
"ith <
is the e;pe&ted ultiate +or origin yearis the proportion paid in de(elopent year
ij ij
ij ij ij
n
ij i j k
k
i
j
C IPoi
C
E C x y y
x i y j
µ
µ η
=
Ε = =
= =∑
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 26/30
B) as a Bayesian Model
Put a prior distri!ution on the ro" paraeters
he Bornhuetter)erguson ethod assues there
is prior no"ledge a!out these paraeters, andthere+ore uses a Bayesian approa&h he prior
in+oration &ould !e suarised as the
+ollo"ing prior distri!utions +or the ro"
paraeters*
( )iii x β α ,tindependen? Γ
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 27/30
B) as a Bayesian Model
• $sing a per+e&t prior -(ery sall (arian&e.
gi(es results analogous to the B) ethod
• $sing a (ague prior -(ery large (arian&e.
gi(es results analogous to the standard
&hain ladder odel
• In a Bayesian &onte;t, un&ertainty
asso&iated "ith a B) prior &an !e
in&orporated
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 28/30
Paraeter $n&ertainty in D)A
• O+ten, in D)A, +ore&asts are o!tained usingsiulation, assuing the underlying
paraeters are +i;ed -+or e;aple, a
standard appli&ation o+ #ilie’s odel.• In&luding paraeter un&ertainty ay not !e
straight+or"ard in the a!sen&e o+ a Bayesian
+rae"or, "hi&h in&ludes it naturally• Ignoring paraeter un&ertainty "ill
underestiate the true un&ertainty=
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 29/30
%uary
• Bayesian odelling using siulationethods &an !e used to +it &ople; odels
• )o&us is on distri!utions o+ paraeters or+ore&asts
• Mode is analogous to Ca;iu
lielihood• It is a natural "ay to in&lude paraeter
un&ertainty "hen +ore&asting -eg in D)A.
8/18/2019 Beginners Guide Bayesian
http://slidepdf.com/reader/full/beginners-guide-bayesian 30/30
Re+eren&es
%&ollni, DPM -200<. Actuarial Modeling wit MCMC
and !"G# , orth Aeri&an A&tuarial Hournal, 6 -2., pages
7K<28
England, PD and errall, RH -2002. #toca$tic Claim$
%e$er&ing in General In$urance, British A&tuarial Hournal
olue 5 Part II -to appear.
%piegelhalter, DH, hoas, A and Best, G -<777.,'in!"G# (er$ion )*+ "$er Manual, MR' Biostatisti&s
$nit