introduction to bayesian mcmc models - home (en) · introduction to bayesian mcmc models glenn...
TRANSCRIPT
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Introduction to Bayesian MCMC Models
Glenn Meyers
Presentation to 2017 ASTIN ColloquiumPanama City, Panama
August 23, 2017
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Outline of Workshop
1 Theory behind Bayesian Markov Chain Monte Carlo(MCMC) models
2 Bayesian MCMC in practice (Software)Introductory Examples
3 Stochastic Loss Reserve Models
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Objectives of Workshop
Show what Bayesian MCMC can doWatch it in actionProvide templates for software that attendees can use ormodify for their own purposes.Course materials (including R scripts) are available byemailing me at [email protected].
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Bayesians vs. Frequentists
Given the model X ∼ f (X |θ)Given the set of observations x .
Frequentists test the hypothesis θ = θ0.Bayesians calculate the posterior distribution f (θ|x).
f (θ|x) = f (x |θ) · π(θ)∫ϑ
f (x |ϑ) · π(ϑ) · dϑ
The issue — What is the prior distribution, π(θ)?
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Philosophical Issue
Bayesians select π “subjectively” according to prioropinion.Frequentists respond by saying that conclusions should bedictated solely by looking at “the data.”Bayesians respond with “noninformative” priors.
Is there such a thing?
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Practical Issue — Can we do the calculations?
For most of the 20th century, the frequentists werewinning.
Calculations were easy with quadradic forms needed for thenormal distributions.The General Linear Model (PROC GLM in SAS).As computers and numerical analysis progressed we gotthe Generalized Linear Model (PROC GENMOD in SAS).
Now the Bayesians are winning - with MCMC.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Problem with Bayesian Analysis
Let θ be an n-parameter vector — e.g. developmentfactors.Let X be a set of observations — e.g. a loss triangle.
f (θ|x) = f (x |θ) · π(θ)∫ϑ1
· · ·∫ϑn
f (x |ϑ) · π(ϑ) · dϑ
f (X |θ) is the likelihood of X given θ.π(θ) is the prior distribution of θ.f (θ|X ) is the posterior distribution of θ.
Calculating the n-dimensional integral is intractable.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Problem with Bayesian Analysis
Let θ be an n-parameter vector — e.g. developmentfactors.Let X be a set of observations — e.g. a loss triangle.
f (θ|x) = f (x |θ) · π(θ)∫ϑ1
· · ·∫ϑn
f (x |ϑ) · π(ϑ) · dϑ
f (X |θ) is the likelihood of X given θ.π(θ) is the prior distribution of θ.f (θ|X ) is the posterior distribution of θ.Calculating the n-dimensional integral is intractable.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
A New World Order
This impasse came to an end in 1990 when asimulation-based approach to estimating posteriorprobabilities was introduced.Sampling Based Approach to Calculating MarginalDensities
Alan E. Gelfand and Adrian F.M. SmithJournal of the American Statistical Association, June 1990
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Markov Chains
Let Ω be a finite state with random events
X1,X2, . . . ,Xt , . . .
A Markov chain P satisfies
Pr(Xt = y |Xt−1 = xt−1, . . . ,X1 = x1) = Pr(Xt = y |Xt−1 = xt−1)
The probability of an event in the chain depends only onthe immediate previous event.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Markov Convergence Theorem
There is a branch of probability theory, called ErgodicTheory, that gives conditions for which there exists aunique stationary distribution, π, such that
Pr(y |Xt−1) −→ π(y)
as t −→∞
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Metropolis Hastings AlgorithmA Very Important Markov Chain.
1 Time t = 1: select a random initial position θ1 inparameter space.
2 Select a proposal distribution p(θ|θt−1) that we will use toselect proposed random steps away from our currentposition in parameter space.
3 Starting at time t = 2: repeat the following until you getconvergence:
At step t, generate a proposal θ∗ ∼ p(θ|θt−1).Generate U ∼ uniform(0,1)Calculate
R = f (θ∗|x)f (θt−1|x) ·
p(θt−1|θ∗)p(θ∗|θt−1)
If U < R then θt = θ∗. Else, θt = θt−1.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Dodging the Intractable Integral
R = f (θ∗|x)f (θt−1|x) ·
p(θt−1|θ∗)p(θ∗|θt−1)
R =
f (x |θ∗)·π(θ∗)∫ϑ1
···∫
ϑn
f (x |ϑ)·π(ϑ)·dϑ
f (x |θt−1)·π(θt−1)∫ϑ1
···∫
ϑn
f (x |ϑ)·π(ϑ)·dϑ
· p(θt−1|θ∗)p(θ∗|θt−1)
The integral∫ϑ1
· · ·∫ϑn
f (x |ϑ) · π(ϑ) · dϑ cancels out!
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Metropolis Hastings Algorithm Restated
1 Time t = 1: select a random initial position θ1 inparameter space.
2 Select a proposal distribution p(θ|θt−1) that we will use toselect proposed random steps away from our currentposition in parameter space.
3 Starting at time t = 2: repeat the following until you getconvergence:
At step t, generate a proposal θ∗ ∼ p(θ|θt−1).Generate U ∼ uniform(0,1)Calculate
R = f (x |θ∗) · π(θ∗)f (x |θt−1) · π(θt−1) ·
p(θt−1|θ∗)p(θ∗|θt−1)
If U < R then θt = θ∗. Else, θt = θt−1.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Relevance of theMetropolis Hastings Algorithm
Defined in terms of the conditional distribution
f (X |θ)
and the prior distribution
π(θ)The limiting distribution is the posterior distribution!
Code f (X |θ) and π(θ) into a Markov chain and let it runfor a while, and you have a large sample from the posteriordistribution.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Relevance of theMetropolis Hastings Algorithm
Defined in terms of the conditional distribution
f (X |θ)
and the prior distribution
π(θ)The limiting distribution is the posterior distribution!Code f (X |θ) and π(θ) into a Markov chain and let it runfor a while, and you have a large sample from the posteriordistribution.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Relevance of theMetropolis Hastings Algorithm
The theoretical limiting distribution is the same, no matterwhat proposal distribution, p(θ|θt−1), is used.There is no fundamental limit on the number ofparameters in you model!The practical limit is within range of stochastic lossreserve models.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
A Short History of MCMC
Originated with the study of nuclear fission.Enrico Fermi, John von Neumann, Nicolas Metropolis andStanislaw Ulam.Developed the Metropolis algorithm.
Keith Hastings (1970) recognized the potential of theMetropolis algorithm to solve statistical problems.Simulations were not readily accepted by the statisticalcommunity at that time.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
A Short History of MCMC
Gelfand and Smith (1990) pulled together the relevantideas at a time when simulation was deemed OK.
Seized upon by scientists in other fields.Used the Gibbs sampler (A one parameter at a time specialcase of Metropolis Hastings algorithm).
Statisticians in the UK started the BUGS project toproduce software for MCMC.
Bayesian inference Using the Gibbs Sampler
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Evolution of MCMC Software
WinBUGS (Original — now discontinued)OpenBUGS (Continuation of WinBUGS)
Designed mainly for the Windows operating system.JAGS — Just Another Gibbs Sampler
Originated by Martyn Plummer.Runs on multiple operating systems.Callable from R (“runjags” package.)
Stan (in honor of Stanislaw Ulam)Stan team led by Andrew Gelman at Columbia University.Runs on multiple operation systems.Callable from R (“rstan” package) and other languages,e.g. Python and Matlab.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Illustrate Adaptation and Convergence withLognormal Example
Given the data below, estimate the cost of a 15,000 xs of10,000 layer.Find predictive distribution of losses in that layer.Fit a lognormal distribution with
log(mean) = µlog(standard deviation) = 1
484 603 631 1189 12291407 1565 1894 2140 22442262 2654 2672 4019 43185015 5354 5464 5598 60606500 6747 9143 12782 18349
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Lessons from the Example
Open “MH Intro with Lognormal.R”Adaptation — Scaling the proposal distribution with σpropto get a representative sample in as few iterations aspossible.
µi ∼ normal(µi−1, σprop)
Convergence — Running multiple chains and comparingresults.Thinning — When adaption does not work well, takeevery nth iteration for sufficiently large n. Takes time.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Lessons from the Example
Open “MH Intro with Lognormal.R”Adaptation — Scaling the proposal distribution with σpropto get a representative sample in as few iterations aspossible.
µi ∼ normal(µi−1, σprop)
Convergence — Running multiple chains and comparingresults.
Thinning — When adaption does not work well, takeevery nth iteration for sufficiently large n. Takes time.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Lessons from the Example
Open “MH Intro with Lognormal.R”Adaptation — Scaling the proposal distribution with σpropto get a representative sample in as few iterations aspossible.
µi ∼ normal(µi−1, σprop)
Convergence — Running multiple chains and comparingresults.Thinning — When adaption does not work well, takeevery nth iteration for sufficiently large n. Takes time.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Adaptation in One Dimension is Easy
Trial and error works.If there are many dimensions, say 30 or 1000, some mathis required.The software (e.g. JAGS or Stan) does the math.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Computing Environment for Workshop Examples
Downloadable SoftwareR at http://www.r-project.org/RStudio at http://www.rstudio.com/rstan at http://mc-stan.org
R packages“rstan” and “loo”
Stan is a language separate from R.Stan language put in a character string in R.The command ”stan” from the ”rstan” package calls thestring.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
An Introductory Example Using Stan
Simulate 500 observations from a gamma distribution.
f (y |α, β) = βα · yα−1 · eβ·yΓ(α)
Introduce a covariate, x .
β = 1 + x , x ∈ (0, 1)
Fit a model f (y |α, β0 + β1 · x).Prior distribution
α ∼ f (α|1, 1), β0 ∼ f (β0|1, 1), β1 ∼ f (β1|1, 1)
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
An Introductory Example Using Stan
Simulate 500 observations from a gamma distribution.
f (y |α, β) = βα · yα−1 · eβ·yΓ(α)
Introduce a covariate, x .
β = 1 + x , x ∈ (0, 1)
Fit a model f (y |α, β0 + β1 · x).Prior distribution
α ∼ f (α|1, 1), β0 ∼ f (β0|1, 1), β1 ∼ f (β1|1, 1)
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Structure of stan Script
R script file is “StanIntro.R”Required block - “data”Required block - “parameters”Optional block - “transformed parameters”Required block - “model”Optional block - “generated quantities”
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Calling stan in R
Run four chainsFaster if run in parallel. Requires some setup toward thetop of the script.Test convergence. Trace plots of chains should beindistinguishable. More on that below.
Need to set a list for input data.Need an initialization function.Specify number of iterations, etc, etc.
Now run the “StanIntro.R” script.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Calling stan in R
Run four chainsFaster if run in parallel. Requires some setup toward thetop of the script.Test convergence. Trace plots of chains should beindistinguishable. More on that below.
Need to set a list for input data.Need an initialization function.Specify number of iterations, etc, etc.Now run the “StanIntro.R” script.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Testing for Convergence of the Markov Chain
Trace plots of parameters show rapid convergence.Gelman-Rubin Convergence Diagnostic
Let W = Within Chain VarianceLet B = Between Chain VarianceDefine “Potential Scale Reduction Factor” (PSRF) orR-Hat by: √
R =
√W + B
W−→ 1
Gelman and Rubin suggest that we should acceptconvergence if R-Hat < 1.1
All R-Hat statistics are less than 1.01 for “StanIntro.R”
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Model Selection With the “loo” Package”
Given two different models for the same data, how do youselect the “better” model?
“loo” stands for Leave One Out.Maintained by members of the stan development team.Vehtari, A., Gelman, A., and Gabry, J. (2015). “Efficientimplementation of leave-one-out cross validation andWAIC for evaluating fitted Bayesian models.”See the documentation of the “loo” package for the latestversion of the paper.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Model Selection With the “loo” Package”
Given two different models for the same data, how do youselect the “better” model?“loo” stands for Leave One Out.Maintained by members of the stan development team.Vehtari, A., Gelman, A., and Gabry, J. (2015). “Efficientimplementation of leave-one-out cross validation andWAIC for evaluating fitted Bayesian models.”See the documentation of the “loo” package for the latestversion of the paper.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Selecting Models Fit By Maximum Likelihood
If we fit a model, f (x |θ), by maximum likelihood, define
AIC = 2 · p − 2 · L(x |θ)
Where:p is the number of parameters in the model.L(x |θ) is the maximum log-likelihood of the modelspecified by f .
Lower AIC indicates a better fit.Encourages a larger log-likelihood.Penalizes an increase in the number of parameters.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Selecting Bayesian MCMC Modelswith the LOOIC Statistic
Given an MCMC model with parameters θi10,000i=1 , define
LOOIC = 2 · pLOOIC − 2 · L(x |θi )10,000i=1
WherepLOOIC is the effective number of parameters.
pLOOIC = L(x |θi )10,000i=1 −
N∑n=1L(xn|x(−n), θi )
10,000i=1
x(−n) = x1, . . . , xn−1, xn+1, . . . , xNL(xn|x(−n), θi ) is the log-likelihood of xn from a model fitusing all data except xn.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Selecting Bayesian MCMC Modelswith the LOOIC Statistic
After some algebra we can see that
LOOIC = −2 ·N∑
n=1L(xn|x(−n), θi )
10,000i=1
which we like as it favors the model with the largestlikelihood, and the smallest LOOIC, on the “holdout” data.
Some “loo” package features.∑Nn=1 L(xn|x(−n), θi )
10,000i=1 ≡ ELPDLOO
“loo” does not calculate each summand in ELPDLOO byMCMC. Instead it approximates the sum using a10,000 x N matrix of log-likelihoods.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Selecting Bayesian MCMC Modelswith the LOOIC Statistic
After some algebra we can see that
LOOIC = −2 ·N∑
n=1L(xn|x(−n), θi )
10,000i=1
which we like as it favors the model with the largestlikelihood, and the smallest LOOIC, on the “holdout” data.Some “loo” package features.∑N
n=1 L(xn|x(−n), θi )10,000i=1 ≡ ELPDLOO
“loo” does not calculate each summand in ELPDLOO byMCMC. Instead it approximates the sum using a10,000 x N matrix of log-likelihoods.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Examples Illustrating the LOOIC
“StanIntro.R” model - y ∼ gamma(α, β0 + β1 · x)
Estimate SEELPDLOO -294.59 22.90pLOO 2.82 0.29LOOIC 589.19 45.79
“StanIntro 2.R” model - y ∼ gamma(α, β)
Estimate SEELPDLOO -307.41 23.27pLOO 2.12 0.25LOOIC 614.83 46.54
The model y ∼ gamma(α, β0 + β1 · x) is “better.”
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Examples Illustrating the LOOIC
“StanIntro.R” model - y ∼ gamma(α, β0 + β1 · x)
Estimate SEELPDLOO -294.59 22.90pLOO 2.82 0.29LOOIC 589.19 45.79
“StanIntro 2.R” model - y ∼ gamma(α, β)
Estimate SEELPDLOO -307.41 23.27pLOO 2.12 0.25LOOIC 614.83 46.54
The model y ∼ gamma(α, β0 + β1 · x) is “better.”
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Overlapping Predictors
Run “StanIntro 3.R” - y ∼ gamma(α, β01 · β02 + β1 · x)
Note the hyperbolic shape of the β01, β02 plot.I was trying to construct a slowly converging example.The stan adaptation algorithm works very well.Good and bad newsThe Folk Theorem of Statistical Computing
Computational problems are often model problems indisguise.
One sign of “computational problems” is divergenttransitions. Stan will warn you when they occur. There isnot enough time to discuss them here.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Overlapping Predictors
Run “StanIntro 3.R” - y ∼ gamma(α, β01 · β02 + β1 · x)Note the hyperbolic shape of the β01, β02 plot.
I was trying to construct a slowly converging example.The stan adaptation algorithm works very well.Good and bad newsThe Folk Theorem of Statistical Computing
Computational problems are often model problems indisguise.
One sign of “computational problems” is divergenttransitions. Stan will warn you when they occur. There isnot enough time to discuss them here.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Overlapping Predictors
Run “StanIntro 3.R” - y ∼ gamma(α, β01 · β02 + β1 · x)Note the hyperbolic shape of the β01, β02 plot.I was trying to construct a slowly converging example.
The stan adaptation algorithm works very well.Good and bad newsThe Folk Theorem of Statistical Computing
Computational problems are often model problems indisguise.
One sign of “computational problems” is divergenttransitions. Stan will warn you when they occur. There isnot enough time to discuss them here.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Overlapping Predictors
Run “StanIntro 3.R” - y ∼ gamma(α, β01 · β02 + β1 · x)Note the hyperbolic shape of the β01, β02 plot.I was trying to construct a slowly converging example.The stan adaptation algorithm works very well.Good and bad news
The Folk Theorem of Statistical ComputingComputational problems are often model problems indisguise.
One sign of “computational problems” is divergenttransitions. Stan will warn you when they occur. There isnot enough time to discuss them here.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Overlapping Predictors
Run “StanIntro 3.R” - y ∼ gamma(α, β01 · β02 + β1 · x)Note the hyperbolic shape of the β01, β02 plot.I was trying to construct a slowly converging example.The stan adaptation algorithm works very well.Good and bad newsThe Folk Theorem of Statistical Computing
Computational problems are often model problems indisguise.
One sign of “computational problems” is divergenttransitions. Stan will warn you when they occur. There isnot enough time to discuss them here.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Stochastic Loss Reserving Examples
My initial paper on the subject of stochastic loss reservingwas “Estimating Predictive Distributions for Loss ReserveModels” Variance 2007http://www.variancejournal.org/issues/?fa=article&abstrID=6417
Two features of that paperIt was very primitive Bayesian.It included (what I call) “aggressive backtesting.”
Like many others (e.g. Scollnik (2001), De Alba (2002)and Verrall (2007)) - I saw Bayesian MCMC as apromising tool for stochastic loss reserving.But it was the availability of good data for backtestingthat drew me into the game.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Stochastic Loss Reserving Examples
My initial paper on the subject of stochastic loss reservingwas “Estimating Predictive Distributions for Loss ReserveModels” Variance 2007http://www.variancejournal.org/issues/?fa=article&abstrID=6417
Two features of that paperIt was very primitive Bayesian.It included (what I call) “aggressive backtesting.”
Like many others (e.g. Scollnik (2001), De Alba (2002)and Verrall (2007)) - I saw Bayesian MCMC as apromising tool for stochastic loss reserving.But it was the availability of good data for backtestingthat drew me into the game.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The CAS Loss Reserve Database
The National Association of Insurance Commissioners(NAIC) gave the University of Wisconsin free access totheir database for academic studies.Jed Frees and I were involved in a joint project with theAustralian Institute of Actuaries to study dependenciesbetween lines of business. One of Jed’s graduate studentsat the time, Peng Shi, assembled the database from theNAIC Schedule P data.I (with the help of the CAS) requested to make the dataavailable on the CAS web site. It was granted by TerriVaughan, the then president of the NAIC and CASmember. The link to the database is below.
http://www.casact.org/research/index.cfm?fa=loss_reserves_data
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Schedule P Data for Incurred Losses
Table: Illustrative Insurer Net Written PremiumAY 1 2 3 4 5 6 7 8 9 10
Premium 5812 4908 5454 5165 5214 5230 4992 5466 5226 4962
Table: Illustrative Insurer Incurred Losses Net of ReinsuranceAY \ Lag 1 2 3 4 5 6 7 8 9 10 Source
1988 1722 3830 3603 3835 3873 3895 3918 3918 3917 3917 19971989 1581 2192 2528 2533 2528 2530 2534 2541 2538 2532 19981990 1834 3009 3488 4000 4105 4087 4112 4170 4271 4279 19991991 2305 3473 3713 4018 4295 4334 4343 4340 4342 4341 20001992 1832 2625 3086 3493 3521 3563 3542 3541 3541 3587 20011993 2289 3160 3154 3204 3190 3206 3351 3289 3267 3268 20021994 2881 4254 4841 5176 5551 5689 5683 5688 5684 5684 20031995 2489 2956 3382 3755 4148 4123 4126 4127 4128 4128 20041996 2541 3307 3789 3973 4031 4157 4143 4142 4144 4144 20051997 2203 2934 3608 3977 4040 4121 4147 4155 4183 4181 2006
Insurer 353 - Commercial Auto
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Schedule P Data for Paid Losses
Table: Illustrative Insurer Net Written PremiumAY 1 2 3 4 5 6 7 8 9 10
Premium 5812 4908 5454 5165 5214 5230 4992 5466 5226 4962
Table: Illustrative Insurer Paid Losses Net of ReinsuranceAY \ Lag 1 2 3 4 5 6 7 8 9 10 Source
1988 952 1529 2813 3647 3724 3832 3899 3907 3911 3912 19971989 849 1564 2202 2432 2468 2487 2513 2526 2531 2527 19981990 983 2211 2830 3832 4039 4065 4102 4155 4268 4274 19991991 1657 2685 3169 3600 3900 4320 4332 4338 4341 4341 20001992 932 1940 2626 3332 3368 3491 3531 3540 3540 3583 20011993 1162 2402 2799 2996 3034 3042 3230 3238 3241 3268 20021994 1478 2980 3945 4714 5462 5680 5682 5683 5684 5684 20031995 1240 2080 2607 3080 3678 2004 4117 4125 4128 4128 19971996 1326 2412 3367 3843 3965 4127 4133 4141 4142 4144 20051997 1413 2683 3173 3674 3805 4005 4020 4095 4132 4139 2006
Insurer 353 - Commercial Auto
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Made use of the database in my monograph
The url for the monographhttp://www.casact.org/pubs/monographs/index.cfm?fa=meyers-monograph01
Analyzed 50 insurers in each of four lines of insurance.Backtested the Mack model with the R “ChainLadder”package on incurred data and found that it understatedthe variability of the lower triangle holdout data.Backtested the Mack and ODP Bootstrap model with theR “ChainLadder” package on paid data and found that ittended to overstate the losses in the lower triangle holdoutdata.Proposed new models that improve the performance onthe lower triangle holdout data.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Models Presented in this Workshop
A current version of the “Correlated Chain Ladder” (CCL)model on cumulative incurred triangle data.
Allowing for correlations between accident years canincrease the variability of the predictive distribution.
A current version of the “Changing Settlement Rate”(CSR) model on cumulative paid triangle data.
Allows for a gradual shift in the payout pattern to accountfor a change in the claim settlement rate.
A model that simultaneously fits both the paid andincurred data with the CCL and CSR models.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Correlated Chain Ladder (CCL) Model
Let C Iwd be the cumulative incurred loss for a 10 x 10
triangle for accident year w and development year d .Let Pw be the earned premium for accident year w .Let αw ∼ Normal(0,
√10) for w = 2, . . . , 10. Set α1 ≡ 0.
Let βId ∼ Normal(0,
√10) for d = 1, . . . , 9. Set βI
10 ≡ 0.Let logelr ∼ Normal(0,
√10).
Let ρ ∼ β(2, 2) scaled to go between -1 and 1, whereβ(., .) denotes the β distribution.Let aI
i ∼ Uniform(0, 1) for d = 1, . . . , 10. Then set(σI
d )2 =∑10
i=daI
i . This forces σI1 < σI
2 < . . . < σI10.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Continuing with the CCL Model
Set µI1d = log(P1) + logelr + βI
d for d = 1, . . . , 10.Set µI
wd =log(Pw ) + logelr + αw + βI
d + ρ · (log(C Iw−1,d )− µI
w−1,d )for w = 2, . . . , 10 and d = 1, . . . , 11− w .Then C I
wd ∼ lognormal(µIwd , σ
Id ).
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Outline of the “CCL.R” Script
1 Read in data from the CAS Loss Reserve Database.”2 Compile the stan script with dummy data. This saves time
when you are fitting several triangles in the script.3 Create a function that fits the model.
Call the stan function that outputs a sample of theparameters.Turn the parameters into predicted outcomes by takingrandom draws from a lognormal distribution with log-meanµI
w ,10 and log-standard deviation σIw ,10 for w = 1, . . . , 10.
Makes use of the “parallel” and “doparallel” packages.Calculate statistics of interest, e.g. Overall Estimate,LOOIC, percentile of the outcome.
4 Call the model fitting function for a triangle, or a list oftriangles.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CCL.R”
Run Insurer 353 - CA - Explore output.
Does the inclusion of the ρ parameter improve the modelfit?LOOIC(CA 353) for CCL = -138.3LOOIC(CA 353) for CCL(−ρ) = -142.8 (Lower!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CCL.R”
Run Insurer 353 - CA - Explore output.Does the inclusion of the ρ parameter improve the modelfit?LOOIC(CA 353) for CCL = -138.3
LOOIC(CA 353) for CCL(−ρ) = -142.8 (Lower!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CCL.R”
Run Insurer 353 - CA - Explore output.Does the inclusion of the ρ parameter improve the modelfit?LOOIC(CA 353) for CCL = -138.3LOOIC(CA 353) for CCL(−ρ) = -142.8 (Lower!)
Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CCL.R”
Run Insurer 353 - CA - Explore output.Does the inclusion of the ρ parameter improve the modelfit?LOOIC(CA 353) for CCL = -138.3LOOIC(CA 353) for CCL(−ρ) = -142.8 (Lower!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Distribution of the Mean of ρ for the CCL Model
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Distribution of LOOICCCL − LOOICCCL(−ρ)Negative difference favors CCL
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Meta-Testing Percentiles of Outcomes with theLower Triangle Holdout Data.
Use the model to calculate the percentile of the predictivedistribution of the actual outcome.We should expect the percentiles of the outcomes to beuniformly distributed.Uniformity is testable with model fits and outcomes ofseveral insurers.
PP Plots - Plot the sorted values of a uniformly distributedset of numbers, Expected, against the sorted percentiles ofthe outcomes predicted by the model, Predicted.
We expect the plot to lie along a 45o line.Kolmogorov-Smirnov test puts bounds around how far thedifference between the predicted and expected can be.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Meta-Testing Percentiles of Outcomes with theLower Triangle Holdout Data.
Use the model to calculate the percentile of the predictivedistribution of the actual outcome.We should expect the percentiles of the outcomes to beuniformly distributed.Uniformity is testable with model fits and outcomes ofseveral insurers.PP Plots - Plot the sorted values of a uniformly distributedset of numbers, Expected, against the sorted percentiles ofthe outcomes predicted by the model, Predicted.
We expect the plot to lie along a 45o line.Kolmogorov-Smirnov test puts bounds around how far thedifference between the predicted and expected can be.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
PP Plot Characteristics
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
PP Plot Characteristics
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Mack Model on Incurred Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CCL(−ρ) Model on Incurred Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CCL Model on Incurred Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Discussion Points for CCL Model
Posterior means of ρ exhibit a strong positive tendency.In relatively few of the insurers does the CCL improve thefit as measured by the LOOIC.Does the meta-analysis indicate a superior fit?How should the results of the meta-analysis affect anindividual insurer’s choice of a model?
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Changing Settlement Rate (CSR) Model
Let CPwd be the cumulative paid loss for a 10 x 10 triangle
for accident year w and development year d .Let Pw be the earned premium for accident year w .Let αw ∼ Normal(0,
√10) for w = 2, . . . , 10. Set α1 ≡ 0.
Let βPd ∼ Normal(0,
√10) for d = 1, . . . , 9. Set βP
10 ≡ 0.Let logelr ∼ Normal(0,
√10).
Let γ ∼ Normal(0, 0.05).Let aP
i ∼ Uniform(0, 1) for d = 1, . . . , 10. Then set(σP
d )2 =∑10
i=daP
i . This forces σP1 < σP
2 < . . . < σP10.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Continuing with the CSR Model.
Set µPwd = log(Pw ) + logelr + αw + βp
d · (1− γ)w−1 forw = 1, . . . , 10 and d = 1, . . . , 11− w .Then CP
wd ∼ lognormal(µPwd , σ
Pd ).
A positive γ moves the development year componentcloser to zero with each accident year.Hence a speedup of claim settlement.If γ is negative, the claim settlement slows down.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Continuing with the CSR Model.
Set µPwd = log(Pw ) + logelr + αw + βp
d · (1− γ)w−1 forw = 1, . . . , 10 and d = 1, . . . , 11− w .Then CP
wd ∼ lognormal(µPwd , σ
Pd ).
A positive γ moves the development year componentcloser to zero with each accident year.Hence a speedup of claim settlement.If γ is negative, the claim settlement slows down.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CSR.R”
Run Insurer 353 - CA - Explore output.
Does the inclusion of the γ parameter improve the modelfit?LOOIC(CA 353) for CSR = -99.9LOOIC(CA 353) for CSR(−γ) = -97.4 (Higher!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CSR.R”
Run Insurer 353 - CA - Explore output.Does the inclusion of the γ parameter improve the modelfit?LOOIC(CA 353) for CSR = -99.9
LOOIC(CA 353) for CSR(−γ) = -97.4 (Higher!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Running “CSR.R”
Run Insurer 353 - CA - Explore output.Does the inclusion of the γ parameter improve the modelfit?LOOIC(CA 353) for CSR = -99.9LOOIC(CA 353) for CSR(−γ) = -97.4 (Higher!)Let’s look at other insurers.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Distribution of LOOICCSR − LOOICCSR(−γ)Negative difference favors CSR
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Mack Model on Paid Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CSR(−γ) Model on Paid Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CSR Model on Paid Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Discussion Points for the CSR Model
Posterior means of γ are mixed. Most are positive,indicating a speedup of claim settlement rates.But many speedup rates are either close to zero, ornegative.When the speedup rate is close to zero, we should expectthe CSR(−γ) to have a better fit, as measured by theLOOIC.
Why do we get more occurrences of better fits with theCSR Model?When looking at the PP plots for CSR and CCL models,we see a noticeably larger difference betweenCSR and CSR(−γ) than we do with CCL and CCL(−ρ).
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Discussion Points for the CSR Model
Posterior means of γ are mixed. Most are positive,indicating a speedup of claim settlement rates.But many speedup rates are either close to zero, ornegative.When the speedup rate is close to zero, we should expectthe CSR(−γ) to have a better fit, as measured by theLOOIC.Why do we get more occurrences of better fits with theCSR Model?When looking at the PP plots for CSR and CCL models,we see a noticeably larger difference betweenCSR and CSR(−γ) than we do with CCL and CCL(−ρ).
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Some Intricacies of stan
The model block R/stan script “StanIntro.R” reads:
model alpha ∼ gamma(1,1);beta ∼ gamma(1,1);y ∼ gamma(alpha,beta);
What these sampling statements do is increment (add to)the log-likelihood of the parameters (or data) to a hiddenvariable called “target.”
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Some Intricacies of stan
Here is something you could do.
model alpha ∼ gamma(1,1);beta 1 ∼ gamma(1,1);beta 2 ∼ gamma(1,1);y 1 ∼ gamma(alpha,beta 1);y 2 ∼ gamma(alpha,beta 2);
The log-likelihoods of y 1 and y 2 are added successivelyto “target.”What this does is fit two models to the data y 1 andy 2 with separate rate parameters β 1 and β 2 andcommon shape parameter α.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Apply this idea to the CCL and CSR Models
Look at accident year (level) parameterslogelr - Should be the same if we have a long enoughdevelopment period.The paid and incurred should be close after 10 years.
logelr I ∼ normal(logelrP , 0.05)
Provisionally assume that the αw parameters are the same.Else why bother?
The βd , γ and ρ will differ for paid and incurred losses.
I want to credit Ned Tyrrell, FCAS, who suggested thisintegrated approach to me.Run the “CCL CSR.R” script.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Apply this idea to the CCL and CSR Models
Look at accident year (level) parameterslogelr - Should be the same if we have a long enoughdevelopment period.The paid and incurred should be close after 10 years.
logelr I ∼ normal(logelrP , 0.05)
Provisionally assume that the αw parameters are the same.Else why bother?
The βd , γ and ρ will differ for paid and incurred losses.I want to credit Ned Tyrrell, FCAS, who suggested thisintegrated approach to me.
Run the “CCL CSR.R” script.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Apply this idea to the CCL and CSR Models
Look at accident year (level) parameterslogelr - Should be the same if we have a long enoughdevelopment period.The paid and incurred should be close after 10 years.
logelr I ∼ normal(logelrP , 0.05)
Provisionally assume that the αw parameters are the same.Else why bother?
The βd , γ and ρ will differ for paid and incurred losses.I want to credit Ned Tyrrell, FCAS, who suggested thisintegrated approach to me.Run the “CCL CSR.R” script.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
IP-CCL PP Plots
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Compare with Standalone CCL PP Plots
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
IP-CSR PP Plots
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Compare with Standalone CSR PP Plots
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Comparing Combined IP and Standalone Models
For CA and OL, both the combined and standalone modelare comfortably with the KS bounds. So is PA for the CCLmodel.For the CSR model - PA, the combined model is justwithin the KS bounds for the combined model, andoutside the KS bounds for the standalone model.
What is the difference between WC and the other lines?Run the combined model on Insurer 86 with WC.Big difference between paid and incurred losses at d = 10for WC. Not so for the other lines.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Comparing Combined IP and Standalone Models
For CA and OL, both the combined and standalone modelare comfortably with the KS bounds. So is PA for the CCLmodel.For the CSR model - PA, the combined model is justwithin the KS bounds for the combined model, andoutside the KS bounds for the standalone model.What is the difference between WC and the other lines?Run the combined model on Insurer 86 with WC.
Big difference between paid and incurred losses at d = 10for WC. Not so for the other lines.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Comparing Combined IP and Standalone Models
For CA and OL, both the combined and standalone modelare comfortably with the KS bounds. So is PA for the CCLmodel.For the CSR model - PA, the combined model is justwithin the KS bounds for the combined model, andoutside the KS bounds for the standalone model.What is the difference between WC and the other lines?Run the combined model on Insurer 86 with WC.Big difference between paid and incurred losses at d = 10for WC. Not so for the other lines.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Common αw parameters are not good for WC
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Compare Results for Combined and StandaloneModels on the Illustrative Insurer
Standalone CCL for CA #353Estimate = 39,158, SE = 1,775
Combined CCLEstimate = 38,806, SE = 1,264
Standalone CSREstimate = 37,483, SE = 2,335
Combined CSREstimate = 38,746, SE = 1,353
Note the significant reduction in the SEs for the CombinedModel!For CA, PA and OL, we get good performance on thelower triangle holdout data for both the combined andstandalone models.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Compare Results for Combined and StandaloneModels on the Illustrative Insurer
Standalone CCL for CA #353Estimate = 39,158, SE = 1,775
Combined CCLEstimate = 38,806, SE = 1,264
Standalone CSREstimate = 37,483, SE = 2,335
Combined CSREstimate = 38,746, SE = 1,353
Note the significant reduction in the SEs for the CombinedModel!For CA, PA and OL, we get good performance on thelower triangle holdout data for both the combined andstandalone models.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CCL log(SE) Comparisons for all 200 Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
CSR log(SE) Comparisons for all 200 Triangles
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Quest for a Minimum Variance Estimator
Ned Tyrrell pulled me into combined models to get moreaccurate, i.e. less variability in the, estimates of the lossreserve liability.Ned’s objective appears to have been met for CA, PA andOL.
My suggestions for using combined models1 Fit standalone CCL and CSR models.2 Compare αw parameters.3 If the above parameters are sufficiently close, fit a
combined CCL and CSR model.Or any other paid and incurred model that you deemappropriate.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Quest for a Minimum Variance Estimator
Ned Tyrrell pulled me into combined models to get moreaccurate, i.e. less variability in the, estimates of the lossreserve liability.Ned’s objective appears to have been met for CA, PA andOL.My suggestions for using combined models
1 Fit standalone CCL and CSR models.2 Compare αw parameters.3 If the above parameters are sufficiently close, fit a
combined CCL and CSR model.
Or any other paid and incurred model that you deemappropriate.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
The Quest for a Minimum Variance Estimator
Ned Tyrrell pulled me into combined models to get moreaccurate, i.e. less variability in the, estimates of the lossreserve liability.Ned’s objective appears to have been met for CA, PA andOL.My suggestions for using combined models
1 Fit standalone CCL and CSR models.2 Compare αw parameters.3 If the above parameters are sufficiently close, fit a
combined CCL and CSR model.Or any other paid and incurred model that you deemappropriate.
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Concluding Remarks
The first part of the workshop was high-level description ofBayesian MCMC, emphasizing one way of “how to do it.”The second part of the workshop was more focused onwhat it can do, and passing over many of the fine pointsof how to do it.
A common theme to all the examples was fitting a lossdistribution with covariates.The predictive distribution is a mixture of fairly simpledistributions over a large sample from the posteriordistribution.Do not expect to master this technique quickly. It willtake a lot of work.
Done!
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Concluding Remarks
The first part of the workshop was high-level description ofBayesian MCMC, emphasizing one way of “how to do it.”The second part of the workshop was more focused onwhat it can do, and passing over many of the fine pointsof how to do it.A common theme to all the examples was fitting a lossdistribution with covariates.The predictive distribution is a mixture of fairly simpledistributions over a large sample from the posteriordistribution.
Do not expect to master this technique quickly. It willtake a lot of work.
Done!
Glenn Meyers Introduction to Bayesian MCMC Models
Introductionto Bayesian
MCMCModels
Glenn Meyers
Introduction
MCMCTheory
MCMCHistory
IntroductoryExample
Using Stan
Loss ReserveModels
CCL Model
CSR Model
CCL ∪ CSR
Remarks
Concluding Remarks
The first part of the workshop was high-level description ofBayesian MCMC, emphasizing one way of “how to do it.”The second part of the workshop was more focused onwhat it can do, and passing over many of the fine pointsof how to do it.A common theme to all the examples was fitting a lossdistribution with covariates.The predictive distribution is a mixture of fairly simpledistributions over a large sample from the posteriordistribution.Do not expect to master this technique quickly. It willtake a lot of work.
Done!Glenn Meyers Introduction to Bayesian MCMC Models