robust risk aggregation and merging p-values · background useful results p-values merging via...
TRANSCRIPT
Background Useful results P-values Merging via averaging Simulation Conclusion
Robust Risk Aggregation and Merging P-values
Ruodu Wang
http://sas.uwaterloo.ca/~wang
Department of Statistics and Actuarial Science
University of Waterloo
4th European Actuarial Journal Conference
KU Leuven, Belgium September 11, 2018
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 1/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Agenda
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 2/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Fundamental problem in Finance/Insurance
Basic setup.
I A vector of risk factors: X = (X1, . . . ,Xd)
I A financial position Ψ(X)
I A risk measure ρ
Calculate ρ(Ψ(X))
Most relevant choices:
I ρ = VaRp or ρ = ESp (TVaRp)
I Ψ(X) =∑d
i=1 Xi
Challenge: We need a joint model for the random vector X
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 3/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Frechet problem
Model assumption: Xi ∼ Fi , Fi known, i = 1, . . . , d .
Let Sd = Sd(F1, . . . ,Fd) =
d∑
i=1
Xi : Xi ∼ Fi , i = 1, . . . , d
.
I Every element in Sd is a possible risk position
I Determination (distributions-wise) of Sd : very challenging.
I Example:
• F = U[0, 1]
• Characterization of S2(F ,F ) is an open question
• S3(F ,F ,F ) characterized analytically recently
S ∈ S3(F ,F ,F ) ⇔ S ≤cx U[0, 3]
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 4/59
Background Useful results P-values Merging via averaging Simulation Conclusion
VaR and ES
Two regulatory risk measures in Basel III & IV and Solvency II
Value-at Risk (VaR)
For p ∈ (0, 1], X ∈ L0,
VaRp(X ) = F−1X (p) = infx ∈ R : FX (x) ≥ p.
Expected Shortfall (ES, or TVaR, CVaR, CTE, AVaR, ...)
For p ∈ (0, 1), X ∈ L1,
ESp(X ) =1
1− p
∫ 1
pVaRq(X )dq =
(F cont.)E [X |X > VaRp(X )] .
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 5/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Worst- and best-values of VaR and ES
The Frechet problems
I For p ∈ (0, 1),
VaRp(Sd) = supVaRp(S) : S ∈ Sd(F1, . . . ,Fd),
VaRp(Sd) = infVaRp(S) : S ∈ Sd(F1, . . . ,Fd).
I Same notation for ESp
I ES is subadditive: ESp(Sd) =∑d
i=1 ESp(Xi ).
I VaRp(Sd), VaRp(Sd), and ESp(Sd): generally open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 6/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Basel III & IV ES calculation
In the Basel FRTB internal model approach, for market risk:
Capital Charge = λESp
(d∑
i=1
Xi
)︸ ︷︷ ︸
internal model
+(1− λ)d∑
i=1
ESp(Xi )︸ ︷︷ ︸ESp(Sd )
,
where
I Xi is the total random loss from a risk class, i = 1, . . . , d
• commodity, equity, credit spread, interest rate, exchange
I T = 10-day, p = 0.975, λ = 0.5
I ESp is calculated under a stressed scenario
Dependence uncertainty!
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 7/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Solvency II SCR calculation
Copied from Solvency II, 2009
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 8/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Robust risk aggregation
Calculation of VaRp(Sd), VaRp(Sd), and ESp(Sd)
I Dependence modeling, (co/counter)-monotonicity, ...
• Dhaene, Denuit, Goovaert, Vanduffel, ...
I Risk measures
• Artzner-Delbaen-Eber-Heath, Wang, ...
I Mass transportation
• Ruschendorf, Embrechts, Puccetti, ...
I Joint mixability
I Computation, neural network methods, ...
I Uncertainty, ambiguity, decision making, ...
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 9/59
Background Useful results P-values Merging via averaging Simulation Conclusion
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 10/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Some results in robust VaR aggregation
Some results
I Embrechts-Puccetti’06, F&S
I W.-Peng-Yang’13, F&S
I Bernard-Jiang-W.’14, IME
I Wang-W.’15, JMVA
I Embrechts-Wang-W.’15 F&S
I Jakobsons-Han-W.’16, SAJ
I Bignozzi-Mao-Wang-W.’16, Extremes
I Wang-W.’16, MOR
This talk - work with Vladimir Vovk (Royal Holloway, London)
I Vovk-W.’18
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 11/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simple VaR-ES relation
I For p ∈ (0, 1), define
ES←p (·) =1
p
∫ p
0VaRq(·)dq.
I By definition
ES←p ≤ VaRp ≤ ESp.
I By subaddivity of ESp and superadditivity of ES←p ,
d∑i=1
ES←p (Xi ) ≤ VaRp
(d∑
i=1
Xi
)≤
d∑i=1
ESp(Xi ).
I Asymptotically sharp as d →∞ (Embrechts-Wang-W’15).
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 12/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Dual bound of Embrechts-Puccetti’06
Assume a homogeneous model F1 = · · · = Fd = F .
Derived from Theorem 4.2 of Embrechts-Puccetti’06
For a continuous cdf F and X1, . . . ,Xd ∼ F ,
P(X1 + · · ·+ Xd ≥ ds) ≤ infr∈[0,s)
∫ ds−(d−1)rr (1− F (x))dx
s − r.
I Based on Ruschendorf’82
I probability bounds ⇔ VaR bounds
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 13/59
Background Useful results P-values Merging via averaging Simulation Conclusion
VaRp formula of W.-Peng-Yang’13
Derived from Theorem 3.4 of W.-Peng-Yang’13
Suppose that the density of F is decreasing on its support. For
p ∈ (0, 1) and X ∼ F ,
VaRp(Sd) = dE[X |F−1(p + (d − 1)c) ≤ X ≤ F−1(1− c)],
where c is the smallest number in [0, 1d (1− p)] such that
∫ 1−c
p+(d−1)cF−1(t)dt ≥ 1− p − dc
d((d − 1)F−1(p + (d − 1)c) + F−1(1− c)).
I only decreasing density in the tail part is required
I dual form in Puccetti-Ruschendorf’13
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 14/59
Background Useful results P-values Merging via averaging Simulation Conclusion
VaRp formula of W.-Peng-Yang’13
Derived from Theorem 3.4 of W.-Peng-Yang’13
Suppose that the density of F is decreasing on its support. For
p ∈ (0, 1) and X ∼ F ,
VaRp(Sd) = max(d − 1)F−1(0) + F−1(p), dES←p (X ).
I An asymmetry between upper and lower bounds ...
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 15/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Asymptotic equivalence of Wang-W.’15
Corollary 3.7 of Wang-W.’15
For p ∈ (0, 1) and X ∼ F ,
limd→∞
1
dVaRp(Sd) = ESp(X ).
I VaRp(Sd)/ESp(Sd)→ 1.
I First results obtained by Puccetti-Ruschendorf’14
I Inhomogeneous portfolios: Embrechts-Wang-W.’15
I Other risk measures, rank dependent utilities:
W.-Bignozzi-Tsanakas’15, Cai-Liu-W.’18, W.-Xu-Zhou’18
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 16/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Aggregation ratio of Bignozzi-Mao-Wang-W.’16
Derived from Proposition 3.5 of Bignozzi-Mao-Wang-W.’16
For a Pareto distribution F with index β ∈ (0, 1),
limd→∞
VaRp(Sd)
d1/βF−1(p)=
(1
1− β
) 1β
.
I If β ≥ 1, asymptotic equivalence implies
limd→∞
VaRp(Sd)
dF−1(p)=
ESp(X )
VaRp(X )≈ β
β − 1.
I For MRV models (e.g. Embrechts-Lambrigger-Wuthrich’09),
limp↑1
VaRp(X1 + · · ·+ Xd)
d1/βF−1(p)≤ 1.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 17/59
Background Useful results P-values Merging via averaging Simulation Conclusion
VaRp formula of Jakobsons-Han-W.’16
Assume inhomogeneous models
Derived from Corollary 4.7 of Jakobsons-Han-W.’16
Suppose that the density of each F1, . . . ,Fd is decreasing on its
support. For p ∈ (0, 1),
VaRp(Sd) = max
maxi=1,...,d
F−1i (p) +
∑j 6=i
F−1j (0)
, µp
,
where µp =∑d
i=1 ES←p (Xi ), Xi ∼ Fi , i = 1, . . . , d .
I Formula for VaRp(Sd) involves solving an implicit ODE
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 18/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Summary on VaR aggregation
d = 2
I solved analytically (Makarov’81, Ruschendorf’82)
I based on counter-monotonicity
d ≥ 3
I solved analytically for decreasing densities
• also solved for increasing densities by putting a negative sign
• relies on joint mixability for monotone densities (Wang-W’16)
I generalization to other distributions is limited
I homogeneous relaxation in Bernard-Jiang-W.’14
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 19/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Summary on VaR aggregation
Remarks.
I Efficient numerical algorithm: the Rearrangement Algorithm• Puccetti-Ruschendorf’12, Embrechts-Puccetti-Ruschendorf’13,
Hofert-Memartoluie-Saunders-Wirjanto’17,
Bernard-Bondarenko-Vanduffel’18, ...
I Risk aggregation with partial dependence information• Bernard-Vanduffel’15, Puccetti-Ruschendorf-Manko’16,
Lux-Papapantoleon’17, Bernard-Ruschendorf-Vanduffel’17,
Bernard-Ruschendorf-Vanduffel-W.’17, ...
I Risk aggregation with marginal and dependence uncertainty• Li-Shao-W.-Yang’18, Shao’18, ...
I Connection to distributionally robust optimization
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 20/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Robust VaR aggregation
Some disadvantages:
I often too high risk values for practical use
I analytical results require monotone density assumptions
I assumed no dependence information
I assumed full marginal information
Are there real-world problems with these features?
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 21/59
Background Useful results P-values Merging via averaging Simulation Conclusion
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 22/59
Background Useful results P-values Merging via averaging Simulation Conclusion
P-values
STAT 101
A (true) p-value p for testing a hypothesis H0:
I p is uniform on [0, 1] under H0 ⇔ PH0(p ≤ ε) = ε for ε ∈ [0, 1]
I Specify a significant level α, typically α = 0.05, 0.01, 0.005 ...
I Rejects H0 if (realized) p ≤ α• cannot reject H0 if p > α
I Probability of type I error = PH0(reject H0) = α
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 23/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging p-values
Suppose we are testing the same hypothesis using K ≥ 2 different
statistical tests and obtain p-values p1, . . . , pK . How can we
combine them into a single p-value?
Examples.
I backtesting credit risk ratings: typically 17 binomial tests
I backtesting market risk models: several quantile level tests
I meta-analysis
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 24/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Meta-analysis
A typical example from meta-analysis
The sex differences dataset, from p.35 of Hedges-Olkin’85
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 25/59
Background Useful results P-values Merging via averaging Simulation Conclusion
The Bonferroni method
A question of a long history
I Fisher’48: assumes p1, . . . , pK are independent
Without any assumptions on the p-values p1, . . . , pK ...
I The Bonferroni method (Dunn’58):
F (p1, . . . , pK ) = K min(p1, . . . , pK ).
I Ruger’78:
F (p1, . . . , pK ) =K
kp(k).
In particular, 2 times the median or the maximum.
I Hommel’83:
F (p1, . . . , pK ) =
(1 +
1
2+ · · ·+ 1
K
)min
k=1,...,K
K
kp(k).
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 26/59
Background Useful results P-values Merging via averaging Simulation Conclusion
The Bonferroni method
The Bonferroni method
I overly conservative ... if tests are similar?
I dictated by a single experiment?
I what if some p-values are more important (e.g. bigger
experiments)?
Particular interest: the case of heavily dependent tests.
I is it more natural to look at some average of p-values?
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 27/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging functions
In any atomless probability space (Ω,A,P) ...
Definition 1
(i) A (conservative) p-value is a random variable P that satisfies
P(P ≤ ε) ≤ ε, ε ∈ (0, 1).
(ii) A merging function is an increasing Borel function
F : [0, 1]K → [0,∞) such that F (P1, . . . ,PK ) is a p-value for
all p-values P1, . . . ,PK .
I Controlled type I error
I Merging functions may be applied iteratively in multiple layers
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 28/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging functions
For an increasing Borel function F : [0, 1]K → [0,∞), equivalent
are:
I F is a merging function;
I F (U1, . . . ,UK ) is a p-value for all U1, . . . ,UK ∈ U ;
I for all ε ∈ (0, 1), P(F ≤ ε) ≤ ε, where
P(F ≤ ε) = sup P(F (U1, . . . ,UK ) ≤ ε) | U1, . . . ,UK ∈ U .
U : the set of all uniform [0, 1] random variables
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 29/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Precise merging functions
Definition 2
A merging function F is precise if, for all ε ∈ (0, 1), P(F ≤ ε) = ε.
I A precise merging function cannot be improved
Examples.
I The Bonferroni method F (p1, . . . , pK ) = K min(p1, . . . , pK )
I F (p1, . . . , pK ) = max(p1, . . . , pK )
I F (p1, . . . , pK ) = p1 (trivial)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 30/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Precise merging functions
The Bonferroni method F (p1, . . . , pK ) = K min(p1, . . . , pK )
P(K min(p1, . . . , pK ) ≤ ε) = P
(K⋃i=1
Kpi ≤ ε
)
≤K∑i=1
P (Kpi ≤ ε)
=K∑i=1
ε
K= ε.
The inequality is an equality if Kpi ≤ ε, i = 1, . . . ,K are
mutually exclusive.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 31/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Asymptotically precise merging functions
Definition 3
A family of merging functions FK on [0, 1]K , K = 2, 3, . . ., is called
asymptotically precise if, for any a ∈ (0, 1), aFK is not a merging
function for a large enough K .
I This family of merging functions cannot be improved by a
constant multiplier
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 32/59
Background Useful results P-values Merging via averaging Simulation Conclusion
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 33/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging p-values via averaging
A general notion of averaging
I Axiomatized by Kolmogorov’30,
Mφ,K (p1, . . . , pK ) = φ−1
(φ(p1) + · · ·+ φ(pK )
K
),
where φ : [0, 1]→ [−∞,∞] is continuous and strictly
monotonic.
I Most common forms, for r ∈ R \ 0,
Mr ,K (p1, . . . , pK ) =
(pr1 + · · ·+ prK
K
)1/r
.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 34/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging p-values via averaging
Special cases:
I Arithmetic: M1,K (p1, . . . , pK ) = 1K
∑Kk=1 pk
I Harmonic: M−1,K (p1, . . . , pK ) =(
1K
∑Kk=1
1pk
)−1
I Quadratic: M2,K (p1, . . . , pK ) =√
1K
∑Kk=1 p
2k
Limiting cases:
I Geometric: M0,K (p1, . . . , pK ) =(∏K
k=1 pk
)1/K
I Maximum: M∞,K (p1, . . . , pK ) = max(p1, . . . , pK )
I Minimum: M−∞,K (p1, . . . , pK ) = min(p1, . . . , pK )
The cases r ∈ −1, 0, 1 are known as Platonic means.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 35/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Merging p-values via averaging
The arithmetic average M1,K (p1, . . . , pK ) = 1K
∑Kk=1 pk is not a
merging function (Ruschendorf’82, Meng’93):
P(M1,K ≤ ε) = min(2ε, 1).
I 2M1,K is a precise merging function
Task. Find ar ,K > 0 such that ar ,KMr ,K is a merging function
I Mr ,K increases in r
• The constants ar ,K should decrease in r .
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 36/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Translation to a risk aggregation problem
For α ∈ (0, 1] and a random variable X , define
qα(X ) = infx ∈ R : P(X ≤ x) ≥ α = VaRα(X ).
and for a function F : [0, 1]K → [0,∞), define
qα(F ) = inf qα(F (U1, . . . ,UK )) | U1, . . . ,UK ∈ U .
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 37/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Translation to a risk aggregation problem
Lemma 1
For a > 0, r ∈ [−∞,∞], and F = aMr ,K , equivalent are:
(i) F is a merging function, i.e. P(F ≤ ε) ≤ ε for all ε ∈ (0, 1);
(ii) qε(F ) ≥ ε for all ε ∈ (0, 1);
(iii) P(F ≤ ε) ≤ ε for some ε ∈ (0, 1);
(iv) qε(F ) ≥ ε for some ε ∈ (0, 1).
The same conclusion holds if all ≤ and ≥ are replaced by =.
I In statistical practice one only needs to have P(F ≤ ε) ≤ ε for
a specific ε, e.g. 0.05, 0.01, ...
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 38/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Translation to a risk aggregation problem
It boils down to calculate qε(Mr ,K ), or equivalently:
(i) for r > 0, aggregation of Beta risks
(qε(Mr ,K ))r = infU1,...,UK∈U
qε
(1
K(U r
1 + · · ·+ U rK )
)(ii) for r = 0, aggregation of exponential risks
ln(qε(Mr ,K )) = infU1,...,UK∈U
qε
(1
K(lnU1 + · · ·+ lnUK )
)(iii) for r < 0, aggregation of Pareto risks
(qε(Mr ,K ))r = supU1,...,UK∈U
q1−ε
(1
K(U r
1 + · · ·+ U rK )
)Ruodu Wang ([email protected]) Risk aggregation and merging p-values 39/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Translation to a risk aggregation problem
Breakdown of U r (or lnU) for r ∈ R
super-heavy
Pareto
integrable
Pareto
negative
exponential
increasing
Beta density
decreasing
Beta density
Pareto(1) uniform
r = −1
harmonic
r = 0
geometric
r = 1
arithmetic
r
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 40/59
Background Useful results P-values Merging via averaging Simulation Conclusion
The integrable case: r > −1
Proposition 1
For r ∈ (−1,∞], (r + 1)1/rMr ,K , K ∈ 2, 3, . . ., is a family of
merging functions and it is asymptotically precise.
Proof.
I r > 0, qε(∑K
k=1 Urk) ≥
∑Kk=1 ES
←ε (U r
k) = K 1r+1ε
r
I r = 0, qε(∑K
k=1 lnUk) ≥∑K
k=1 ES←ε (lnUk) = K (ln ε+ 1)
I r < 0, q1−ε(∑K
k=1 Urk) ≤
∑Kk=1 ES1−ε(U
rk) = K 1
r+1εr
I In all cases, qε((r + 1)1/rMr ,K ) ≥ ε
I Use the VaR/ES asymptotic equivalence of Wang-W.’15.
(1 + r)1/r |r=0 = e.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 41/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Main results summary
Constant multiplier ar ,K in front of Mr ,K
rr+1
K1+1/r (1 + r)1/r
e
(1 + r)1/r K1/r
K 1e lnK 2
r = −1
harmonic
r = 0
geometric
r = 1
arithmetic
r = −∞minimum
r =∞maximum
r = 1K−1 r = K − 1
blue: precise; green: asymptotically precise; red: rough
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 42/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Main results summary
Methodology breakdown details
super-heavyaggregation
asymptoticequivalence
VaR formula
joint mixability VaR formula
trivial trivialVaR formula uniform
r = −1
harmonic
r = 0
geometric
r = 1
arithmetic
r = −∞minimum
r =∞maximum
r = 1K−1 r = K − 1
purple: Rushcendorf’82; blue: W.-Peng-Yang’13; brown: Wang-W’11
green: Wang-W.’15; red: Bignozzi-Mao-Wang-W.’16
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 43/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Weighted averaging
Consider weighted averaging functions
Mφ,w(p1, . . . , pK ) = φ−1 (w1φ(p1) + · · ·+ wKφ(pK )) ,
and in particular,
Mr ,w(p1, . . . , pK ) = (w1pr1 + · · ·+ wKp
rK )1/r ,
where w = (w1, . . . ,wK ) ∈ ∆K .
I Intuitively, the weights reflect the prior importance of the
p-values.
∆K =
(w1, . . . ,wK ) ∈ [0, 1]K∣∣w1 + · · ·+ wK = 1
is the standard K -simplex.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 44/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Weighted averaging
Proposition 2
For w = (w1, . . . ,wK ) ∈ ∆K , w = max(w) and r ∈ (−1,∞),
(i) (r + 1)1/rMr ,w is a merging function;
(ii) (r + 1)1/rMr ,w is precise ⇔ w ≤ 1/2 and r ∈ [ w1−w ,
1−ww ];
(iii) if r ∈ [1,∞), min(r + 1, 1w )1/rMr ,w is a precise merging
function.
Proof.
I (r + 1)1/rMr ,w is precise ⇔ VaR = ES ⇔ joint mixability
I For r ∈ (0, 1), use the characterization of joint mixability for
monotone densities of Wang-W.’16.
I For r ≥ 1, use the VaRp formula of Jakobsons-Han-W.’16.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 45/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Weighted averaging
The special case r = 1 (weighted arithmetic mean):
I w ≤ 1/2
• no single experiment outweighs the total of all the others
• the optimal multiplier is 2 (same as uniform weights)
I w > 1/2
• a single experiment outweighs the total of all the others
• assuming w1 = max(w),
1
wM1,w(p1, . . . , pK ) = p1 +
K∑k=2
wk
wpk
• weighted adjustments are added to the p-value obtained from
the most important experiment
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 46/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Weighted averaging
Conjecture
For a > 0 and any r and K , if aMr ,K is a merging function, then
aMr ,w is also a merging function for all w ∈ ∆K .
I A deeper conjecture: Sd(F1, . . . ,Fd) ⊂ Sd(F , . . . , F ) where
F−1 = 1d
∑di=1 F
−1i .
I Bernard-Jiang-W.’14: Sd(F1, . . . ,Fd) ⊂ Sd(F , . . . , F ) where
F = 1d
∑di=1 Fi .
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 47/59
Background Useful results P-values Merging via averaging Simulation Conclusion
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 48/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
A simulation study for the performance of ar ,KMr ,K
I 6 merging functions: r ∈ −∞,−1, 0, 1, 2,∞• Bonferroni (minimum), harmonic, geometric, arithmetic,
quadratic, maximum
I a−1,K = 1.83 lnK for K = 20 and a−1,K = 1.7 lnK for K ≥ 50
I a0,K = e
I others are precise
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 49/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
Simulation setup
I K tests for whether a given set of data has zero mean
• n iid normal observations with mean d/√n and known
variance 1
• two-sided z-tests
I o = %-overlap of observations between any two tests
I c = correlation of non-overlapping observations between tests
I α = 0.05, significance level
I the procedure is repeated N times
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 50/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
Power, K=50, n=50, o=0, c=0, d=2, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.0045 0.0102 0.0842 0.2917 0.4469 0.8699(0.0086) (0.0136) (0.0294) (0.0608) (0.0797) (0.1166)
Power, K=50, n=50, o=0.7, c=0.7, d=2, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
1.3213 0.6079 0.3049 0.2676 0.2687 0.3516(2.9538) (1.0054) (0.4524) (0.3594) (0.3311) (0.3286)
Power, K=50, n=50, o=0, c=0, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0 1e−04 0.0052 0.0666 0.1553 0.4622(0.0001) (0.0003) (0.0026) (0.0266) (0.0681) (0.2280)
Power, K=50, n=50, o=0.7, c=0.7, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.2748 0.1498 0.0818 0.0777 0.0834 0.1316(1.2512) (0.4745) (0.2215) (0.1826) (0.1744) (0.2087)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 51/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
Power, K=50, n=50, o=0, c=0, d=4, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0 0 1e−04 0.0097 0.0331 0.1175(0.0000) (0.0000) (0.0001) (0.0074) (0.0325) (0.1293)
Power, K=50, n=50, o=0.7, c=0.7, d=4, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.0181 0.0117 0.0077 0.0088 0.0109 0.0221(0.1596) (0.0690) (0.0368) (0.0352) (0.0381) (0.0647)
Power, K=50, n=50, o=0.5, c=0.5, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.0315 0.0379 0.0402 0.0632 0.0905 0.2017(0.1304) (0.1343) (0.1010) (0.1156) (0.1359) (0.2464)
Power, K=50, n=50, o=0.9, c=0.9, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.9101 0.2093 0.0885 0.0673 0.0602 0.056(2.8428) (0.5801) (0.2412) (0.1805) (0.1589) (0.1302)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 52/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
Power, K=50, n=50, o=0.95, c=0, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.6755 0.1921 0.0845 0.0668 0.0619 0.0693(2.4028) (0.5412) (0.2284) (0.1735) (0.1549) (0.1423)
Power, K=50, n=50, o=0, c=0.95, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.309 0.1431 0.0712 0.0627 0.0638 0.0929(1.1628) (0.4210) (0.1906) (0.1537) (0.1446) (0.1709)
Power, K=20, n=50, o=0.7, c=0.7, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.1404 0.1091 0.0704 0.0669 0.0718 0.097(0.5987) (0.3567) (0.2007) (0.1657) (0.1590) (0.1760)
Power, K=500, n=50, o=0.7, c=0.7, d=3, N=1000
0.0
0.2
0.4
0.6
0.8
1.0
Bonferroni harmonic geometric arithmetic quadratic maximum
0.8772 0.1713 0.0622 0.0613 0.0682 0.17(4.2020) (0.5478) (0.1688) (0.1444) (0.1420) (0.2347)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 53/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Simulation
Observations.
I The Bonferroni method
• performs very well in cases of no or moderate correlation
• often has a large standard deviation
• is able to utilize information from independent tests
I The arithmetic method
• performs very well in cases of heavy overlap or correlation
• has a relatively stable average p-value
I The geometric method performs quite nicely in most cases
I The harmonic method is between Bonferroni and geometric
I The maximum and quadratic methods are not remarkable
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 54/59
Background Useful results P-values Merging via averaging Simulation Conclusion
1 Robust risk aggregation problems
2 Some results in robust VaR aggregation
3 P-values and hypothesis testing
4 Merging p-values via averaging
5 Simulation study
6 Concluding remarks and open questions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 55/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Concluding remarks
Further directions
I Power analysis for classic statistical models
I Adaptive selection of the merging function
I Relation between dependence and the choice of merging
functions
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 56/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Open questions on risk aggregation
Mathematical questions on robust risk aggregation:
I Characterization of Sd and joint mixability
I Analytical formulas for VaRp, VaRp and ESp
I Aggregation of random vectors
I Partial information on dependence
I RDU and CPT risk aggregation
I Other aggregation functionals
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 57/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Open questions on risk aggregation
A few concrete questions:
I For a given F , determine whether F ∈ S(U[0, 1],U[0, 1])?
I For given correlation matrix Σ and F1, . . . ,Fd , is the set
VΣ = X : Corr(X) = Σ, Xi ∼ Fi , i = 1, . . . , d
empty?
I If VΣ 6= ∅, what are the values of
supVaRp(S) : X ∈ VΣ and supESp(S) : X ∈ VΣ?
Here S = X1 + · · ·+ Xd .
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 58/59
Background Useful results P-values Merging via averaging Simulation Conclusion
Thank you
Thank you for your kind attention
The manuscript is available at
I SSRN: https://ssrn.com/abstract=3166304
I arXiv: https://arxiv.org/abs/1212.4966
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 59/59
Appendix
References I
Bernard, C., Jiang, X. and Wang, R. (2014). Risk aggregation with dependence
uncertainty. Insurance: Mathematics and Economics, 54, 93–108.
Bignozzi, V., Mao, T., Wang, B. and Wang, R. (2016). Diversification limit of
quantiles under dependence uncertainty. Extremes, 19(2), 143–170.
Embrechts, P. and Puccetti, G. (2006). Bounds for functions of dependent risks.
Finance and Stochastics, 10, 341–352.
Jakobsons, E., Han, X. and Wang, R. (2016). General convex order on risk
aggregation. Scandinavian Actuarial Journal, 2016(8), 713–740.
Wang, B. and Wang, R. (2015). Extreme negative dependence and risk
aggregation. Journal of Multivariate Analysis, 136, 12–25.
Wang, B. and Wang, R. (2016). Joint mixability. Mathematics of Operations
Research, 41(3), 808–826.
Wang, R., Peng, L. and Yang, J. (2013). Bounds for the sum of dependent risks
and worst Value-at-Risk with monotone marginal densities. Finance and
Stochastics, 17(2), 395–417.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 60/59
Appendix
References II
Dunn, O. J. (1958). Estimation of the means for dependent variables. Annals of
Mathematical Statistics. 29(4), 1095–1111.
Fisher, R. A. (1948). Combining independent tests of significance. American
Statistician, 2, 30.
Hedges, L. V. and Olkin, I. (1985). Statistical Methods for Meta-Analysis.
Orlando, FL: Academic Press.
Holm, S. (1979). A simple sequentially rejective multiple test procedure.
Scandinavian Journal of Statistics, 6, 65–70.
Mattner, L. (2012). Combining individually valid and arbitrarily dependent
P-variables. In Abstract Book of the Tenth German Probability and Statistics
Days, p. 104. Institut fur Mathematik, Johannes Gutenberg-Universitat Mainz.
Meng, X.-L. (1993). Posterior predictive p-values. Annals of Statistics, 22,
1142–1160.
Ruschendorf, L. (1982). Random variables with maximum sums. Advances in
Applied Probability, 14(3), 623–632.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 61/59
Appendix
Analysis for the sex differences data
For the sex differences dataset, the combined p-values are
(compared with 0.05 significance level; weighted by sample size)
I Bonferroni: 0.029 (significant)
I harmonic: 0.045 (weighted 0.041) (significant)
I geometric: 0.157 (weighted 0.198) (not significant)
I arithmetic: 0.613 (weighted 0.793) (not significant)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 62/59
Appendix
Analysis for the passive smoking data
For the passive smoking dataset (Hartung-Knapp-Sinha’08,
Table 3.1, p.31, K = 19), the combined p-values are (compared
with 0.05 significance level)
I Bonferroni: 0.051 (not significant)
I harmonic: 0.126 (not significant)
I geometric: 0.254 (not significant)
I arithmetic: 0.449 (not significant)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 63/59
Appendix
The non-integrable Pareto case: r < −1
No VaR/ES asymptotic equivalence for r ≤ −1.
Proposition 3
For r ∈ (−∞,−1), rr+1K
1+1/rMr ,K , K ∈ 2, 3, . . ., is a family of
merging functions and it is asymptotically precise.
Proof.
I To show rr+1K
1+1/rMr ,K is a merging function, directly apply
the dual bound of Embrechts-Puccetti’06.
I To show the asymptotic precision, use the aggregation ratio of
Bignozzi-Mao-Wang-W.’16 for super-heavy Pareto risks.
Letting r → −∞ one recovers the Bonferroni method: KM−∞,K .
back
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 64/59
Appendix
Precise results for the Beta case: r ≥ 1/(K − 1)
Proposition 4
For K ∈ 2, 3, . . . and r ∈ (−1,∞),
(i) (r + 1)1/rMr ,K is a precise merging function ⇔r ∈ [ 1
K−1 ,K − 1].
(ii) If r ≥ K − 1, K 1/rMr ,K is a precise merging function.
Proof.
I r ≥ 1, U r has a decreasing density
I r ∈ [ 1K−1 , 1], U r has an increasing density
I The VaRp and VaRp formulas of W.-Peng-Yang’13 give the
precise value of qε(Mr ,K )
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 65/59
Appendix
Precise results for the Beta case: r ≥ 1/(K − 1)
Examples.
I min(r + 1,K )1/rMr ,K is precise for r ≥ 1/(K − 1).
I The arithmetic average times 2 is precise for K ≥ 2
I The quadratic average times√
3 is precise for K ≥ 3
I Letting r →∞, the maximum M∞,K is precise
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 66/59
Appendix
Geometric averaging
Proposition 5
For each K ∈ 2, 3, . . ., aKM0,K is a precise merging function,
where
aK =1
cKexp (−(K − 1)(1− KcK ))
and cK is the unique solution to the equation
ln(1/c − (K − 1)) = K − K 2c
over c ∈ (0, 1/K ). Moreover, aK ≤ e and aK → e as K →∞.
Proof.
I Obtained from the VaRp formula of W.-Peng-Yang’13.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 67/59
Appendix
Geometric averaging
Table: Numeric values of aK/e for the geometric mean
K aK/e K aK/e K aK/e
2 0.7357589 5 0.9925858 10 0.9999545
3 0.9286392 6 0.9974005 15 0.9999997
4 0.9779033 7 0.9990669 20 1.0000000
I In practice, use aK ≈ e for K ≥ 5
I eM0,K is always a merging function (noted by Mattner’12)
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 68/59
Appendix
Harmonic averaging
Proposition 6
For K > 2, (e lnK )M−1,K is a merging function.
Proof.
I For a given K > 2, e lnK = minr<−1r
r+1K1+1/r
I (e lnK )Mr ,K is a merging function for some r < −1
I M−1,K ≥ Mr ,K for r < −1
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 69/59
Appendix
Harmonic averaging
Proposition 7
Set aK = (yK+K)2
(yK+1)K , K > 2, where yK is the unique solution to the
equation
y2 = K ((y + 1) ln(y + 1)− y), y ∈ (0,∞).
Then aKM−1,K is a precise merging function. Moreover,
aK/ lnK → 1 as K →∞.
Proof.
I Again obtained from the VaRp formula of W.-Peng-Yang’13.
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 70/59
Appendix
Harmonic averaging
Table: Numeric values of aK/ lnK for the harmonic mean
K aK/ lnK K aK/ lnK K aK/ lnK
3 2.499192 10 1.980287 100 1.619631
4 2.321831 20 1.828861 200 1.561359
5 2.214749 50 1.693497 400 1.514096
I The rate of convergence aK/ lnK → 1 is very slow
I Suggestions:
• for K ≥ 3, use (2.5 lnK )M−1,K
• for K ≥ 10, use (2 lnK )M−1,K
• for K ≥ 50, use (1.7 lnK )M−1,K back
Ruodu Wang ([email protected]) Risk aggregation and merging p-values 71/59