Mathematical Finance Seminar @ U of Southern California Los Angeles, CA – September 12, 2016
Longevity Risk: Methods, Models, and Management
Daniel Bauer Georgia State University
This presentation is primarily based on a collaborations with Enrico Biffis, Fausto Gozzi, and Nan Zhu. Financial support from the Society of Actuaries under a CAE Grant is gratefully acknowledged. Page 2 Longevity Risk | September 12, 2016 | Daniel Bauer
Introduction What is "mortality/longevity risk"? Is it a big deal? For whom?
Mortality Surface Models Motivation Factor Analysis of Mortality Forecasts Mortality Term Structure Models Self-consistency Condition Applications
Conclusion Page 3 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction
Introduction What is "mortality/longevity risk"? Is it a big deal? For whom?
Mortality Surface Models Motivation Factor Analysis of Mortality Forecasts Mortality Term Structure Models Self-consistency Condition Applications
Conclusion I Are we all getting older? If so, is there a limit? Is there risk/uncertainty?
I What does this mean for individuals, corporations, and economies?
Page 4 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Some folklore
I Jeanne Louise Calment
I longest confirmed human lifespan on record
I living to the age of 122 years, 164 days
I died in 1997 (she was born before the telephone was invented!)
I Signed a contingency contract (similar to a reverse mortgage) at age 90 that paid 2,500 francs ($500) a month for her apartment at death. Altogether, more than 900,000 francs ($180,000)
I Smoked till 119, loved chocolate ^¨ Page 4 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Some folklore
I Jeanne Louise Calment
I longest confirmed human lifespan on record
I living to the age of 122 years, 164 days
I died in 1997 (she was born before the telephone was invented!)
I Signed a contingency contract (similar to a reverse mortgage) at age 90 that paid 2,500 francs ($500) a month for her apartment at death. Altogether, more than 900,000 francs ($180,000)
I Smoked till 119, loved chocolate ^¨
I Are we all getting older? If so, is there a limit? Is there risk/uncertainty?
I What does this mean for individuals, corporations, and economies? Page 5 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Vaupel et al. – We are getting older, and the trend has been fairly stable...
More information: Oeppen & Vaupel (2002) "Broken Limits to Life Expectancy." Science, 296: 1029-1031. Page 6 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Olshansky et al. – ...but there may be a limit (?)
How did we get here? Why did life expectancy rise by 30 years in the last century? U.S. FEMALES Maximum Lifespan Potential = 120
Maximum Observed Age at Death = 105, 113 *
Period Life Expectancy at Birth = 49, 80 *
Modal Age at Death = 73, 88 *
2000
1900
More information: Olshansky, Carnes, & Désquelles (2001) "Demography: Prospects for Human Longevity." Science, 291: 1491-1492. Graph from presentation by J. Olshansky: http://www.aaas.org/spp/rd/Olshansky.pdf I Mitigation/Management (?)
I Transaction design / transfer risk to individuals
I Shift from DB to DC pensions, increase in retirement age, self-annuitization schemes, tontines, etc. ? But individuals not very well equipped to take on risk post retirement...1
I Securitization / mortality derivatives
I Various transactions to transfer mortality risk to investors, at a premium (Buy-Outs, Buy-Ins, Longevity/Survivor Swaps, Longevity Bonds, etc.) ? Longevity market has not lived up to expectations (reminiscent of mortgage-backed securities?) although picked up a little in recent past...
I Are these risks appropriately understood? Do existing models capture all the risks in an appropriate way?
Page 7 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction So there is risk/uncertainty... But how grave is it?
I E.g., how will the life expectancy change over time? → Individual: level of retirement savings → Insurance company: (risk of) liability of deferred annuity → Economy: Social security, retirement age
1Turner (2006) "Pensions, risks, and capital markets." JRI 73: 559-574. I Are these risks appropriately understood? Do existing models capture all the risks in an appropriate way?
Page 7 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction So there is risk/uncertainty... But how grave is it?
I E.g., how will the life expectancy change over time? → Individual: level of retirement savings → Insurance company: (risk of) liability of deferred annuity → Economy: Social security, retirement age
I Mitigation/Management (?)
I Transaction design / transfer risk to individuals
I Shift from DB to DC pensions, increase in retirement age, self-annuitization schemes, tontines, etc. ? But individuals not very well equipped to take on risk post retirement...1
I Securitization / mortality derivatives
I Various transactions to transfer mortality risk to investors, at a premium (Buy-Outs, Buy-Ins, Longevity/Survivor Swaps, Longevity Bonds, etc.) ? Longevity market has not lived up to expectations (reminiscent of mortgage-backed securities?) although picked up a little in recent past...
1Turner (2006) "Pensions, risks, and capital markets." JRI 73: 559-574. Page 7 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction So there is risk/uncertainty... But how grave is it?
I E.g., how will the life expectancy change over time? → Individual: level of retirement savings → Insurance company: (risk of) liability of deferred annuity → Economy: Social security, retirement age
I Mitigation/Management (?)
I Transaction design / transfer risk to individuals
I Shift from DB to DC pensions, increase in retirement age, self-annuitization schemes, tontines, etc. ? But individuals not very well equipped to take on risk post retirement...1
I Securitization / mortality derivatives
I Various transactions to transfer mortality risk to investors, at a premium (Buy-Outs, Buy-Ins, Longevity/Survivor Swaps, Longevity Bonds, etc.) ? Longevity market has not lived up to expectations (reminiscent of mortgage-backed securities?) although picked up a little in recent past...
I Are these risks appropriately understood? Do existing models capture all the risks in an appropriate way?
1Turner (2006) "Pensions, risks, and capital markets." JRI 73: 559-574. I (We need) Data:
I Popular: Human Mortality Database. University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany). Available at www.mortality.org or www.humanmortality.de.
I Other sources: CDC – NVSS or Wonder Berkeley Mortality database
Page 8 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction (Insurance / Financial) Mathematics
2 I (We need a) Model: Lee-Carter model (by far the most popular model, applied by policy makers, governmental agencies, and the broader scholarly literatures)
log{mx,t } = αx + βx κt + εx,t ; κt = θ + κt−1 + et
(mx,t is the central death rate for age group x in year t)
I Estimation typically in two step procedure using a singular value decomposition on {mx,t }x,t to identify κt .
2Lee & Carter (1992) "Modeling and forecasting US mortality." JASA, 87: 659-671. Page 8 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction (Insurance / Financial) Mathematics
2 I (We need a) Model: Lee-Carter model (by far the most popular model, applied by policy makers, governmental agencies, and the broader scholarly literatures)
log{mx,t } = αx + βx κt + εx,t ; κt = θ + κt−1 + et
(mx,t is the central death rate for age group x in year t)
I Estimation typically in two step procedure using a singular value decomposition on {mx,t }x,t to identify κt .
I (We need) Data:
I Popular: Human Mortality Database. University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany). Available at www.mortality.org or www.humanmortality.de.
I Other sources: CDC – NVSS or Wonder Berkeley Mortality database
2Lee & Carter (1992) "Modeling and forecasting US mortality." JASA, 87: 659-671. Page 9 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Lee-Carter In Matlab:
l_mx_female = log(mx_female); alpha = mean(l_mx_female,2); for t=1:T l_mx_female(:,t) = l_mx_female(:,t) - alpha * ones(X,1); end; Sigma = cov(transpose(l_mx_female)); [eigvec,eigval] = eig(Sigma); beta_first = eigvec(:,X); kappa = transpose(beta_first) * l_mx_female; deltakappa = zeros(T-1,1); for t=1:T-1 deltakappa(t) = kappa(t+1) - kappa(t); end; theta_kap = mean(deltakappa); sigma_kap = std(deltakappa);
−1 0.18 2
1.5 0.16 −2 1
0.14 0.5 −3
0 0.12
−4 −0.5
0.1 −1
−5 0.08 −1.5
−2 −6 0.06 −2.5
−7 0.04 −3 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60
β x αx → Gompertz (1825): mx = α e βx κt Page 10 Longevity Risk | September 12, 2016 | Daniel Bauer Introduction Mortality probability and cohort life expectancy for 50-year old US female
-3 x 10 38 5.5 37 5 36
4.5 35
4 34
33 3.5 32 3 31
2.5 life expectancy at age 50 in years 30 one-year mortality rate at age 50 in years
2 29 1960 1970 1980 1990 2000 2010 2020 2030 1960 1970 1980 1990 2000 2010 2020 2030 year year
I One year mortality probability (q50) seems relatively straightforward
I Cohort life expectancy (Forward looking! Not period life expectancy as for Vaupel line!) increasing, although good bit of variation. The 95% confidence region (red dashed) seems too narrow... Page 11 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
Introduction What is "mortality/longevity risk"? Is it a big deal? For whom?
Mortality Surface Models Motivation Factor Analysis of Mortality Forecasts Mortality Term Structure Models Self-consistency Condition Applications
Conclusion 2. Currently expected to live 25.1 more years. How uncertain is this number? What are the chances that it changes to 23.7 or 24.5 next year? ⇒ Risk in mortality projections
I While the two questions are related, and the distinction does not matter in theory, it is relevant for the statistical approach
I For personal financial planning/household finance, insurers’ liability risk evaluation, and government/public economics, question 2 may be more suitable
Page 12 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Consider a 60 year old individual who faces aggregate mortality risk: 1. Currently expected to die with probability of 0.7% next year (mortality rate). How uncertain is this appraisal? What are the chances that the rate is 0.65% or 0.75%? ⇒ Risk in mortality rates I While the two questions are related, and the distinction does not matter in theory, it is relevant for the statistical approach
I For personal financial planning/household finance, insurers’ liability risk evaluation, and government/public economics, question 2 may be more suitable
Page 12 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Consider a 60 year old individual who faces aggregate mortality risk: 1. Currently expected to die with probability of 0.7% next year (mortality rate). How uncertain is this appraisal? What are the chances that the rate is 0.65% or 0.75%? ⇒ Risk in mortality rates
2. Currently expected to live 25.1 more years. How uncertain is this number? What are the chances that it changes to 23.7 or 24.5 next year? ⇒ Risk in mortality projections Page 12 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Consider a 60 year old individual who faces aggregate mortality risk: 1. Currently expected to die with probability of 0.7% next year (mortality rate). How uncertain is this appraisal? What are the chances that the rate is 0.65% or 0.75%? ⇒ Risk in mortality rates
2. Currently expected to live 25.1 more years. How uncertain is this number? What are the chances that it changes to 23.7 or 24.5 next year? ⇒ Risk in mortality projections
I While the two questions are related, and the distinction does not matter in theory, it is relevant for the statistical approach
I For personal financial planning/household finance, insurers’ liability risk evaluation, and government/public economics, question 2 may be more suitable ⇒ We want to find & estimate tractable, coherent mortality models that adequately reflect risk in mortality projections
I Analogy in the motivation: Term structure models
I Forecast yield curve p(T + 1, τ) based on p(t, τ) (q. 2!) Q R t+τ I p(t, τ) = Et [exp(− t rs ds)], so (in theory) modeling rt suffices (q. 1!) I Necessity to consider cross-sectional data (persistence vs. transience, etc.) I However, there are key differences (age, P vs. Q, etc.)
Page 13 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
!! ! !!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!! "!!"!!"!
!
! ! "#$%! !!!!! &! ! ! '! ! ! ! !!!!!(! ! ! ! ! ! !!!!!()'!
I Current stochastic mortality models focus on stochastically forecasting mortality rates (question 1) → red in graphic
I This paper considers the risk in mortality projections (question 2) → blue in graphic I Analogy in the motivation: Term structure models
I Forecast yield curve p(T + 1, τ) based on p(t, τ) (q. 2!) Q R t+τ I p(t, τ) = Et [exp(− t rs ds)], so (in theory) modeling rt suffices (q. 1!) I Necessity to consider cross-sectional data (persistence vs. transience, etc.) I However, there are key differences (age, P vs. Q, etc.)
Page 13 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
!! ! !!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!! "!!"!!"!
!
! ! "#$%! !!!!! &! ! ! '! ! ! ! !!!!!(! ! ! ! ! ! !!!!!()'!
I Current stochastic mortality models focus on stochastically forecasting mortality rates (question 1) → red in graphic
I This paper considers the risk in mortality projections (question 2) → blue in graphic
⇒ We want to find & estimate tractable, coherent mortality models that adequately reflect risk in mortality projections Page 13 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
!! ! !!!!!!!!!!!!!!!!!
!!!!!!!!!!!!!! "!!"!!"!
!
! ! "#$%! !!!!! &! ! ! '! ! ! ! !!!!!(! ! ! ! ! ! !!!!!()'!
I Current stochastic mortality models focus on stochastically forecasting mortality rates (question 1) → red in graphic
I This paper considers the risk in mortality projections (question 2) → blue in graphic
⇒ We want to find & estimate tractable, coherent mortality models that adequately reflect risk in mortality projections
I Analogy in the motivation: Term structure models
I Forecast yield curve p(T + 1, τ) based on p(t, τ) (q. 2!) Q R t+τ I p(t, τ) = Et [exp(− t rs ds)], so (in theory) modeling rt suffices (q. 1!) I Necessity to consider cross-sectional data (persistence vs. transience, etc.) I However, there are key differences (age, P vs. Q, etc.) Page 14 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models Preview of Results: Forecasts of Life Expectancy
38
37
36
35
34
33
32
31
life expectancy at age 50 in years 30
29 1960 1970 1980 1990 2000 2010 2020 2030 year I Assume: Time-homogeneous, Gaussian model
dµt = (A µt + Λ) dt + Σ dWt
I SDE on some Hilbert space H, A = (∂τ − ∂x )
I Markov, time-homogeneous (?) [different / much milder than Markov assumption on hazard rate!]
I Gaussian [W fin-dim. Brownian motion] → Not necessarily positive (???) → Use for now [tractability!] but also discuss non-negative models I Define: τ+l px (tj+1) τ+l+∆px−∆(tj ) Fl (tj , (τ, x)) = − log , tj+1 − tj = ∆ τ px (tj+1) τ+∆px−∆(tj ) ¯ Fl (tj ,(τ,x)) I Proposition: Vectors Fl (tj ) = ω(τ, x) × √ iid Gaussian. t −t j+1 j (τ,x)∈C˜
Page 15 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Given: (Forward) survival probabilities {τ px (t)|(τ, x) ∈ C}
I Easier to model/work with: (Forward) force of mortality / hazard rate ∂ µ (τ, x) = − log{ p (t)} t ∂τ τ x I Define: τ+l px (tj+1) τ+l+∆px−∆(tj ) Fl (tj , (τ, x)) = − log , tj+1 − tj = ∆ τ px (tj+1) τ+∆px−∆(tj ) ¯ Fl (tj ,(τ,x)) I Proposition: Vectors Fl (tj ) = ω(τ, x) × √ iid Gaussian. t −t j+1 j (τ,x)∈C˜
Page 15 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Given: (Forward) survival probabilities {τ px (t)|(τ, x) ∈ C}
I Easier to model/work with: (Forward) force of mortality / hazard rate ∂ µ (τ, x) = − log{ p (t)} t ∂τ τ x
I Assume: Time-homogeneous, Gaussian model
dµt = (A µt + Λ) dt + Σ dWt
I SDE on some Hilbert space H, A = (∂τ − ∂x )
I Markov, time-homogeneous (?) [different / much milder than Markov assumption on hazard rate!]
I Gaussian [W fin-dim. Brownian motion] → Not necessarily positive (???) → Use for now [tractability!] but also discuss non-negative models Page 15 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Given: (Forward) survival probabilities {τ px (t)|(τ, x) ∈ C}
I Easier to model/work with: (Forward) force of mortality / hazard rate ∂ µ (τ, x) = − log{ p (t)} t ∂τ τ x
I Assume: Time-homogeneous, Gaussian model
dµt = (A µt + Λ) dt + Σ dWt
I SDE on some Hilbert space H, A = (∂τ − ∂x )
I Markov, time-homogeneous (?) [different / much milder than Markov assumption on hazard rate!]
I Gaussian [W fin-dim. Brownian motion] → Not necessarily positive (???) → Use for now [tractability!] but also discuss non-negative models I Define: τ+l px (tj+1) τ+l+∆px−∆(tj ) Fl (tj , (τ, x)) = − log , tj+1 − tj = ∆ τ px (tj+1) τ+∆px−∆(tj ) ¯ Fl (tj ,(τ,x)) I Proposition: Vectors Fl (tj ) = ω(τ, x) × √ iid Gaussian. t −t j+1 j (τ,x)∈C˜ Raw data: (deterministic) mortality forecasts generated from rolling windows of past mortality experience [from Human Mortality Databases, www.mortality.org]
I Regions: England/Wales (ENW), France (FRA), Japan (JPN), United States (USA), and West Germany (FRG)
I Genders: male and female
I Years: 1956-2006 / as much as we can
I Methods: 1. Lee-Carter
I Weighted-least-squares algorithm I Ages: 0-95 2. CBD-Perks
I Basic model w/o cohort effect I Ages: 25-95 3. P-splines
I Fixing the degree of freedom (df ) at 20 I Ages: 25-95
Page 16 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
"True" mortality forecasts from market place or insurance prices → not abundant/noisy... [key results same for "official" UK life tables] Page 16 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
"True" mortality forecasts from market place or insurance prices → not abundant/noisy... [key results same for "official" UK life tables] Raw data: (deterministic) mortality forecasts generated from rolling windows of past mortality experience [from Human Mortality Databases, www.mortality.org]
I Regions: England/Wales (ENW), France (FRA), Japan (JPN), United States (USA), and West Germany (FRG)
I Genders: male and female
I Years: 1956-2006 / as much as we can
I Methods: 1. Lee-Carter
I Weighted-least-squares algorithm I Ages: 0-95 2. CBD-Perks
I Basic model w/o cohort effect I Ages: 25-95 3. P-splines
I Fixing the degree of freedom (df ) at 20 I Ages: 25-95 Page 17 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models Surfaces of Eigenvectors: PC1
female, US, Lee-Carter male, France, Lee-Carter female, Ger, Lee-Carter
female, US, splines male, France, splines female, Ger, splines
→ Explains 90 − 97% of the variation! Page 18 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
female, US, Lee-Carter Page 19 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I First two factors explain vast majority of variation in most cases (mostly > 90% for first factor, first two factors >> 95%)
I Very similar shapes exhibited across different countries/genders/forecasting methods (at least for the first factor) ⇒ PC1
I Systematic, increasing in age/term
I Forward forces of mortality for high ages in the far future more volatile than in the near future → slope factor
⇒ PC2
I Additional effect for higher ages
I Inverse relationship between high ages in the near and in the far future → twist factor
¯ 0 ¯ I Simple factor models by regressing Fl (tj ) on Yi = bi Fl (tj ) [akin to Diebold & Li (2006, JEconometrics), Duffee (2011), etc.] I Maths:
I Implies martingale property that yields drift condition, i.e. Λ = f (Σ)
I Prop.: µt allows for Gaussian finite-dimensional realization iff
Σ(τ, x) = C(x + τ) × exp{Mτ} × N
× × [M ∈ Rm m, N ∈ Rm d , C ∈ C([0, ∞), Rm) – semi-parametric form] R t ⇒ µt (τ, x) = µ0(τ + t, x − t) + 0 Λ(τ + t − s, x − t + s) ds Z t +C(x + τ) exp {M τ} exp {M (t − s)} N dWs 0 | {z } =Zt (state process)
Page 20 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Interpretation as forecasts:
I Expected values for future survival probs. comprised in the current surface ⇒ Expectation from (stochastic) forecast should coincide with "burnt in" values
I Cross-sectional restriction akin to no-arbitrage condition for interest rate term structure models Page 20 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Interpretation as forecasts:
I Expected values for future survival probs. comprised in the current surface ⇒ Expectation from (stochastic) forecast should coincide with "burnt in" values
I Cross-sectional restriction akin to no-arbitrage condition for interest rate term structure models
I Maths:
I Implies martingale property that yields drift condition, i.e. Λ = f (Σ)
I Prop.: µt allows for Gaussian finite-dimensional realization iff
Σ(τ, x) = C(x + τ) × exp{Mτ} × N
× × [M ∈ Rm m, N ∈ Rm d , C ∈ C([0, ∞), Rm) – semi-parametric form] R t ⇒ µt (τ, x) = µ0(τ + t, x − t) + 0 Λ(τ + t − s, x − t + s) ds Z t +C(x + τ) exp {M τ} exp {M (t − s)} N dWs 0 | {z } =Zt (state process) I Estimation:
I Use Maximum Likelihood Estimation that may or may not enforce self-consistency condition
I In addition, allow for non-systematic deviations ¯ obs ¯ mod I F (tj , tj+1) = F (tj , tj+1) + j , j ∼ N(0, α · diag{Σ}), α small
Page 21 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Identification:
¯ I Relationship to Fl / PCA: Z τ+1 Z ∆ d Mv M(∆−s) Fl (τ, x) = E[Fl (τ, x)] + C(x + v) e dv × e N dWs τ 0 | {z } | {z } =O(τ,x) =Z∆
I Prop.: We can treat factors separately
I Find model by regressing on principal components
I Obtain M and N from regression on factor w/o functional assumptions on C(x) [identification issues in higher dimensions – rely on examples from interest rate modeling to find convenient shapes, in particular Björk and Gombani (1999, FinStoch)]
I Potentially use functional assumptions for capturing C(x) [parametric] Page 21 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models
I Identification:
¯ I Relationship to Fl / PCA: Z τ+1 Z ∆ d Mv M(∆−s) Fl (τ, x) = E[Fl (τ, x)] + C(x + v) e dv × e N dWs τ 0 | {z } | {z } =O(τ,x) =Z∆
I Prop.: We can treat factors separately
I Find model by regressing on principal components
I Obtain M and N from regression on factor w/o functional assumptions on C(x) [identification issues in higher dimensions – rely on examples from interest rate modeling to find convenient shapes, in particular Björk and Gombani (1999, FinStoch)]
I Potentially use functional assumptions for capturing C(x) [parametric]
I Estimation:
I Use Maximum Likelihood Estimation that may or may not enforce self-consistency condition
I In addition, allow for non-systematic deviations ¯ obs ¯ mod I F (tj , tj+1) = F (tj , tj+1) + j , j ∼ N(0, α · diag{Σ}), α small Page 22 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models Associated Hazard Rate Based on our specific one-factor assumption, the aged-dependent hazard rate / (spot) force of mortality is Z t (2) µt (x) = µ0(t, x − t) + Λ(t − s, x − t + s) ds +C(x) × Zt 0 | {z } may use "baseline" if no µ0 e.g. (2) = (A + B exp{c x}) × (ξ2 + Zt ) where ( ) ! Z 1 −2b −b2 1 − ab dZ = d t = Z dt + dW t (2) 1 0 t a t Zt Example path:
8 140 Z2 W 7 Z1 120
6 100 5
80 4
3 60
2 40
1 20 0
0 ï1
ï2 ï20 0 4 8 12 16 20 0 4 8 12 16 20 t t (Non-negative model: OU → Feller square root) Page 23 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models Confidence Intervals for Future LE: USA Female (After 1 yr)
44.4 16.2 44.2 16.1 44 16 43.8 15.9 43.6 15.8 43.4
15.7 43.2
43 15.6
42.8 15.5
42.6 15.4
42.4 15.3 Lee-Carter Non-parametric Factor Lee-Carter Non-parametric Factor
age 40 age 70
17 47
16.8 46
16.6 45 16.4 44 16.2
43 16
42 15.8
41 15.6
40 15.4
15.2 39
15 Factor Mortality Surface Non-negative Factor Mortality Surface Non-negative
age 40 age 70 Page 24 Longevity Risk | September 12, 2016 | Daniel Bauer Mortality Surface Models Cohort Effects
I Key difference to yield curve models: new generations / cohort effects (Renshaw & Haberman, 2006, IME; Cairns et al., 2009, NAAJ)
log{mx,t } = αx + βx κt + γt−x + εx,t
I Deterministic models: µt (τ, x) = µ0(τ + t, x − t), x ≥ t > 0, τ ≥ 0, µt (τ, x) = ψt−x (τ + x), 0 ≤ x < t
I Differential notation on H:
0 (0) µt = A µt − AB ψt 2 µ0 ∈ H, ψ ∈ Lγ (R+)
I Stochastic extension yield SPDEs with boundary noise
I Questions:
I Well-posedness (spaces?)
I Finite-dimensional realizations Page 25 Longevity Risk | September 12, 2016 | Daniel Bauer Conclusion
Introduction What is "mortality/longevity risk"? Is it a big deal? For whom?
Mortality Surface Models Motivation Factor Analysis of Mortality Forecasts Mortality Term Structure Models Self-consistency Condition Applications
Conclusion Page 26 Longevity Risk | September 12, 2016 | Daniel Bauer Conclusion Conclusion
I Having appropriate estimates for the risk in mortality projections is important
I Common approach may not be suitable to appraise risk of changes in life expectancies
I Research Directions:
I Continuous models with cohort effects
I Multiple populations, trade off risk?
I Application in household finance: Annuitization decision in portfolio context/influence of systematic mortality risk
I Asset pricing model with aggregate demographic uncertainty