reserving methods: future - actuaries institute · reserving methods: future trends greg taylor...
TRANSCRIPT
![Page 1: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/1.jpg)
Reserving methods: future trends
Greg Taylor
Taylor Fry Consulting ActuariesUniversity of Melbourne
University of New South Wales
![Page 2: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/2.jpg)
Overview• Discussion focuses on stochastic reserving
models• Some comments on current stochastic
reserving practices• Discussion of some of the more advanced
models currently available• Examination of some extensions of these that
are within reach in the near future
![Page 3: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/3.jpg)
Stochastic models
Opening observations
![Page 4: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/4.jpg)
General framework• Data vector Y• Model Y=f(β)+ε
• Estimate of β• Future observation vector Z=g(β)+η• Forecast Z*=g( ) of Z• Prediction error Z-Z*=[g(β) - g( )] + η
β
Parameter vector
Stochastic error vector
β
β
![Page 5: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/5.jpg)
Estimating prediction error• Prediction error Z-Z*=[g(β) - g( )] + η
• Both errors estimated in terms of residuals of data with respect to model
R=Y- =Y- f( )• Ultimately distributional properties of Z*
depend on R(f) and g
Y
Parameter error
Process error
β
β
![Page 6: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/6.jpg)
Pictorially
Data Fitted
Forecast
Model f Forecast g
![Page 7: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/7.jpg)
Pictorially
Data Fitted
Residual
Model f Forecast g
Forecast g
Forecast
Forecast error
![Page 8: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/8.jpg)
Pictorially
Data Fitted
Residual
Model f Forecast g
Forecast g
Call this case of forecast error estimation coherent
Forecast
Forecast error
![Page 9: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/9.jpg)
Pictorially
Data Fitted
Residual
Model f Forecast g
Forecast h???
Forecast
Forecast error
![Page 10: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/10.jpg)
For example
Data Fitted
Residual
Model PPCI Forecast PPCI
Forecast chain ladder (Mack)
Forecast
Forecast error
![Page 11: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/11.jpg)
Pictorially
Data Fitted
Residual
Model f Forecast g
Forecast h???
Call this case of forecast error estimation incoherent
Forecast
Forecast error
![Page 12: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/12.jpg)
Forecast
Forecast error
Pictorially
Data Fitted
Residual
Model f Forecast g
Forecast h???
![Page 13: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/13.jpg)
Conclusion 1• Any incoherent estimation of stochastic
properties of a loss reserve is meaningless
![Page 14: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/14.jpg)
Available options for forecast error estimation
• Only two– Internal estimation
• Based on measured error between data and model (such as just illustrated)
• Good for capturing features inherent in the model– Parameter error– Process error
– External estimation• Based on
– Identification of specific components of forecast error (see O’Dowd, Smith & Hardy, 2005) e.g.
» Future changes in superimposed inflation» Generally systemic changes that are not well represented in
past data– Judgmental assessment of their contributions
![Page 15: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/15.jpg)
Conclusion 2• Ideally, forecast error should be composed of
– Internal estimates• Parameter error• Process error
– External estimates• Model specification error• Errors due to other systemic effects
![Page 16: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/16.jpg)
Internal estimation of forecast error
Data Fitted
Residual
Model f Forecast g
Forecast gLarge residuals imply large forecast error
Forecast
Forecast error
![Page 17: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/17.jpg)
Conclusion 3• Good models produce low forecast error
(CoV)– Economic in use of capital
• Poor models produce high forecast error– Uneconomic in use of capital
![Page 18: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/18.jpg)
Internal estimation of forecast error
Data Fitted
Residual
Model f Forecast g
Forecast g
Forecast
Forecast error
![Page 19: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/19.jpg)
Internal estimation of forecast error
Data Fitted
Residual
Model f Forecast g
Forecast g
This is what the bootstrap does
Forecast
Forecast error
![Page 20: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/20.jpg)
Bootstrapping• One internal form of forecast error estimation• Are there others?• Very rarely
– Due to intractable mathematical complexity in mapping residuals to forecast error
• So need to make the bootstrap work
![Page 21: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/21.jpg)
Bootstrap
Data Fitted
Residual
Model f Forecast g
Forecast g
Forecast
Forecast error
![Page 22: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/22.jpg)
Bootstrap
Data Fitted
Residual
Model f Forecast g
Forecast gPseudo-residual Resample
Forecast
Forecast error
![Page 23: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/23.jpg)
Bootstrap
Data Fitted
Residual
Model f Forecast g
Forecast gResample
Forecast
Forecast error
Generate pseudo-data
Pseudo-residual
![Page 24: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/24.jpg)
Bootstrap
Data Fitted
Residual
Model f Forecast g
Forecast gResample
Forecast
Forecast error
Generate pseudo-data
Pseudo-residual
Assumed iid
![Page 25: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/25.jpg)
Bootstrap
Data Fitted
Residual
Model f Forecast g
Forecast gResample
Forecast
Forecast error
Generate pseudo-data
Pseudo-residual
Assumed iid
![Page 26: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/26.jpg)
What happens when residuals not iid? - example
ResidualPseudo-residual
Resample
Assumed iid
Actually large residuals
Residuals assume more influential positions – can distort model and forecast
![Page 27: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/27.jpg)
Conclusion 4• Particular care is needed to ensure that
model residuals are consistent with iidassumption if ludicrous bootstrap results are to be avoided
![Page 28: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/28.jpg)
Individual claim reservingand
Statistical case estimation
![Page 29: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/29.jpg)
Reserving data treatment
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Raw data Date of accident
Date of notification
Age
Gender
Income
etc
![Page 30: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/30.jpg)
Reserving data treatment
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Raw data Summary dataAccident period
Development period
Information lost
![Page 31: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/31.jpg)
Why does quantity of data matter?
0
5
10
15
20
25
1 2
Data set
Volu
me
of d
ata
00.5
11.5
22.5
33.5
44.5
1 2
Data set
CoV
of r
eser
ve
00.5
11.5
22.5
33.5
44.5
1 2
Data set
Cap
ital
empl
oyed
![Page 32: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/32.jpg)
Individual claim modelling• Data vector Y• Model Y=f(β)+ε• Let components Yi of Y relate to individual
claims– Yi denotes some outcome for the i-th claim, e.g.
finalised size, paid to date, etc.• Call this model an individual claim model• Call a reserve based on such a model an
individual claim reserve
![Page 33: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/33.jpg)
Example Yi = finalised individual claim sizeYi ~ GammaE[Yi]
= exp {function of operational time+ function of accident period (legislative change)+ function of finalisation period (superimposed inflation+ joint function (interaction)of operational time & accident period
(change in payment pattern attributable to legislative change)}
Discussion in Taylor & McGuire (2004)
![Page 34: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/34.jpg)
Pictorially No longer
Now
Data Fitted ForecastModel f
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Forecast (reserve)
β
![Page 35: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/35.jpg)
For example
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Accident period 1
Accident period 2Accident period 3
Accident period m
:
:
β
![Page 36: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/36.jpg)
Alternatively
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Special case of individual claim reserving – statistical case estimation
β
![Page 37: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/37.jpg)
Can bootstrap individual claim reserve
Just in the usual wayForecast
g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:Model
Y=f(β)+ε
Forecast (reserve)
Claim 1
Claim 2Claim 3
Claim n
:
:
: Residuals
Claim 1
Claim 2Claim 3
Claim n
:
:
: Pseudo-residuals
Re-sample
Generate pseudo-data
β
![Page 38: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/38.jpg)
Adaptive reserving
![Page 39: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/39.jpg)
Static and dynamic models
• Return for a while to models based on aggregate (not individual claim) data
• Model form is still Y=f(β)+ε• Example
– i = accident quarter– j = development quarter– E[Yij] = a jb exp(-cj) = exp [α+βln j - γj]
– (Hoerl curve PPCI for each accident period)
![Page 40: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/40.jpg)
Static and dynamic models (cont’d)• Example
E[Yij] = a jb exp(-cj) = exp [α+βln j - γj]– Parameters are fixed– This is a static model
But parameters α,β,γmay vary (evolve) over time, e.g. with accident period
Then– E[Yij] = exp [α(i) +β(i) ln j – γ(i) j]– This is a dynamic model, or adaptive model
![Page 41: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/41.jpg)
Illustrative example of evolving parameters
Separate curves represent different accident periods
0
10
20
30
40
50
60
70
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
Development period
![Page 42: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/42.jpg)
Formal statement of dynamic model• Suppose parameter evolution takes place over accident
periods• Y(i)=f(β(i)) +ε(i) [observation equation]• β(i) = u(β(i-1)) + ξ(i) [system equation]
• Let (i|s) denote an estimate of β(i) based on only information up to time s
Some function Centred stochastic perturbation
β
![Page 43: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/43.jpg)
Adaptive reservingβ (1|k)
β (2|k)
β (k|k)
k-th diagonal
Forecast at valuation date
k
![Page 44: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/44.jpg)
Adaptive reserving (cont’d)• Reserving by means of an adaptive model is
adaptive reserving• Parameter estimates evolve over time• Fitted model evolves over time• The objective here is “robotic reserving” in
which the fitted model changes to match changes in the data– This would replace the famous actuarial
“judgmental selection” of model
![Page 45: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/45.jpg)
Special case of dynamic model: DGLM
• Y(i)=f(β(i)) +ε(i) [observation equation]• β(i) = u(β(i-1)) + ξ(i) [system equation]• Special case:
– f(β(i)) = h-1(X(i) β(i)) for matrix X(i)– ε(i) has a distribution from the exponential dispersion family
• Each observation equation denotes a GLM– Link function h– Design matrix X(i)
• Whole system called a Dynamic Generalised Linear Model (DGLM)
![Page 46: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/46.jpg)
Special case of DGLM: Kalman filter• Y(i)=f(β(i)) +ε(i) [observation equation]• β(i) = u(β(i-1)) + ξ(i) [system equation]• Special case:
– f(β(i)) = h-1(X(i) β(i)) for matrix X(i)– ε(i) has a distribution from the exponential dispersion family
• Further specialised– h(.) = identity function
• So f(.) is linear– u(.) is linear– ε(i), ξ(i) ~ N(0,.)– This is the model underlying the Kalman filter (see De Jong &
Zehnwirth, 1983)
![Page 47: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/47.jpg)
Form of Kalman filter• Let (i|s) be a fitted value, or forecast, of Y(i) on the
basis of data to time s• Take (i|s) = X(i) (i|s)• Kalman filter estimates
(i|i) = (i|i-1) + K(i) [Y(i) – X(i) (i|i-1)]
β
β ββ
Kalman gain (credibility) matrix
Y
Y
![Page 48: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/48.jpg)
Implementation of DGLMs• The restrictions of the Kalman filter may not always
be convenient– Linear relation between response variate and covariates– Normal distribution of claim observations
• Implementation of a more general DGLM is more difficult
• Can be done using an MCMC (Markov Chain Monte Carlo) approach
• Would be useful to have a simple updating formula similar to that of the Kalman filter (a GLM filter)– See Taylor, 2005
![Page 49: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/49.jpg)
Bootstrapping DGLMs• Recursive nature of the GLM filter creates
correlations between residuals• So conventional bootstrapping is wrong
– It assumes independence between residuals• Necessary to modify the bootstrap to take
account of the correlations– Say how– See Stoffer & Wall (1991)
![Page 50: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/50.jpg)
Adaptive individual claim reserving
![Page 51: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/51.jpg)
We began with…
Data Fitted
Residual
Model f Forecast g
Forecast g
Forecast
Forecast error
![Page 52: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/52.jpg)
moved to GLM modelling…
Data Fitted
Residual
GLM Forecast g
Forecast g
Forecast
Forecast error
![Page 53: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/53.jpg)
changed to adaptive GLM modelling…
Data Fitted
Residual
DGLM Forecast g
Forecast g
Forecast
Forecast error
![Page 54: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/54.jpg)
also considered individual claim modelling…
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Accident period 1
Accident period 2Accident period 3
Accident period m
:
:
β
![Page 55: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/55.jpg)
which can be individual claim GLM modelling…
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
GLM
Accident period 1
Accident period 2Accident period 3
Accident period m
:
:
β
![Page 56: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/56.jpg)
and could be adaptive individual claim GLM modelling…
Forecast g( )
Claim 1
Claim 2Claim 3
Claim n
:
:
:
Claim 1
Claim 2Claim 3
Claim n
:
:
:
DGLM
Accident period 1
Accident period 2Accident period 3
Accident period m
:
:
β
![Page 57: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/57.jpg)
References • De Jong P & Zehnwirth B (1983). Claims reserving state space models
and the Kalman filter. Journal of the Institute of Actuaries, 110, 157-181.
• O’Dowd C, Smith A & Hardy P (2005). A framework for estimating incertainty in insurance claims cost. XVth General Insurance Seminar, Institute of Actuaries of Australia.
• Stoffer D S& Wall K D (1991). Bootstrapping state-space models: Gaussian maximum likelihood estimation and the Kalman filter. Journal of the American Statistical Association, 86, 1024-1033
• Taylor G (2005). Second order Bayesian revision of a generalisedlinear model. Available at www.economics.unimelb.edu.au/actwww/wps2005/No125.pdf
• Taylor G & McGuire G (2004). Loss reserving with GLMs: a case study. Casualty Actuarial Society 2004 Discussion Paper Program, 327-392.
![Page 58: Reserving methods: future - Actuaries Institute · Reserving methods: future trends Greg Taylor Taylor Fry Consulting Actuaries University of Melbourne University of New South Wales](https://reader030.vdocuments.us/reader030/viewer/2022020216/5be810be09d3f2d3638cb4e6/html5/thumbnails/58.jpg)
Reserving methods: future trends
Greg Taylor
Taylor Fry Consulting ActuariesUniversity of Melbourne
University of New South Wales