Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Stochastic Optimization
and
Indifference Valuation
Thaleia Zariphopoulou
UT, Austin
1
Contents
• Historical perspective
• Beyond derivatives
• Investments and stochastic optimization
• Optimal behavior and valuation
• Integrated pricing systems
2
Replication
3
Concept of value
• The option price corresponds to the cost of perfect replication
of the option payoff.
• Perfect replication requires strong assumptions about the model.
• In 1973, Black, Scholes and Merton developed such a model.
• Another example was proposed in 1979 by Cox, Ross and Rubinstein.
4
Does it work in practice?
• In reality no risk can be perfectly replicated.
• The discrepancies depend a lot on the type of risk we look at and
on the variables that affect it.
• So why is this idea so fundamental?
• It laid conceptual foundations for the development of a multibillion
derivatives industry.
5
Milestones
6
Stochastic Calculus and Finance
• In 1979, Harrison and Kreps created a link between pricing by perfect repli-
cation and the general theory of martingales.
• The price was shown to correspond to the expected value, calculated under
the risk neutral probability, of the discounted option payoff.
• In 1981, Harrison and Pliska presented this stochastic calculus based ap-
proach to a large community of researchers from that field.
7
Term structure models
• Black-Scholes model assumes constant interest rates.
• In reality, the interest rate is not constant and the entire curve of interest
rates of different maturities is traded.
• In 1992, Heath, Jarrow and Morton developed a framework for the derivatives
pricing under stochastic interest rates.
• At this stage, infinite dimensional analysis and the stochastic PDEs started
to play a role in Mathematical Finance.
8
Volatility models
• Black-Scholes model assumes constant volatility.
• In reality the volatility is not constant and is traded according to a market
specific convention.
• It depends on the option strike and maturity.
• This dependence is known as the volatility smile or skew.
• In 1993, Dupire constructed a martingale diffusion that is consistent with
the market smile.
9
Models consistent with smile
• In general, asset evolution is represented by a martingale and the option
price may not be defined uniquely by replication.
• The marginals are given by the market smile/skew — there are many ways
to construct such a martingale.
• Solutions differ by the nature of the forward volatilities they generate.
• The notion of value drifts away from the concept of payoff replication. The
price is defined as the expected value under a martingale measure of the
discounted option payoff.
10
Credit derivatives
• The Black-Scholes model did not have a major impact on the development
of this market.
• Is there anything quantitative that the credit markets embraced on their
side to the same extent as the other markets accepted the Black-Scholes
formula?
• A credit default swap provides information about the distribution of the
default time for the underlying name.
• The pricing of tranches of synthetic indices requires the knowledge of the
joint distribution of all default times.
11
Credit derivatives
• The price is defined by an expectation of the payoff, calculated with respect
to the joint distribution of the default times.
• The marginals are given by the relevant CDSs.
• The joint distribution is constructed by choosing a copula function.
• The notion of price has nothing to do with the concept of payoff replication.
12
Beyond derivatives
13
Funds management
• Funds management industry uses relatively simple mathematics when com-
pared with the derivatives side.
• Mathematical models of investment and consumption are very complex, to
an extent more complex than the models used in the derivatives pricing.
• The challenge is to bridge this gap.
• One needs to connect the concept of replication used in the derivatives
pricing with the investment optimization.
14
From replication towards
optimal investment
• In a sense one needs to move from: ‘I can do it for you’ – replication,
towards: ‘I will do my best’ – optimal investment.
• First steps in the academic literature are being undertaken now.
• Pricing of derivatives has been linked with optimal investment.
15
Market uncertainty, preferences
and investments
16
Investments and stochastic optimization
• Maximal expected utility models
• Methods
Primal problem (HJB eqn)
Dual problem (BS eqn)
• Optimal policies: consumption and portfolios
17
Maximal expected utility models
• Market uncertainty
(Ω,F , P), W = (W 1, . . . , Wd)∗ d-dim Brownian motion
Trading horizon: [0, T ], (0, +∞)
Asset returns: dRt = µt dt + σt dWt
µ ∈ L1(Rm), σ ∈ L2(R
d×m)
riskless asset
Wealth process: dXt = πt dRt − Ct dt
Control processes: consumption rate Ct, asset allocation πt
18
Maximal expected utility models
• Preferences: U : R → R
increasing, concave, asymptotically elastic....
U (x) =1
γxγ, log x, −e−γx
• Objective: maximize intermediate utility of consumption and
utility of terminal wealth
V (x, t) = sup(C,π)
EP
(∫ T
tU1(Cs) ds + U2(XT )/Xt = x
)
• Generalizations: infinite horizon, long-term average, ergodic criteria...
19
Primal maximal expected utility problem
• V solves the Hamilton-Jacobi-Bellman eqn⎧⎪⎨⎪⎩Vt + F (x, Vx, Vxx; U1) = 0
V (x, T ) = U2(x)
• Viscosity theory (Crandall-Lions)
Z., Soner, Touzi, Duffie-Z., Elliott, Davis-Z., Bouchard
• Optimal policies in feedback form
C∗s = C((V −1
x )′(X∗s , s)) , π∗s = π(Vx(X∗
s , s), Vxx(X∗s , s))
• Degeneracies, discontinuities, state and control constraints
20
Dual maximal expected utility problem
• Dual utility functional
U∗(y) = maxx
(U (x) − xy)
• Dual problem becomes linear – direct consequence of market completeness
and representation, via risk neutrality, of replicable contingent claims
• Problem reduces to an optimal choice of measure – intuitive connection with
the so-called state prices
Cox-Huang, Karatzas, Shreve, Cvitanic, Schachermayer, Zitkovic,
Kramkov, Delbaen et al, Kabanov, Kallsen, ...
21
Extensions
• Recursive utilities and Backward Stochasticc Differential Equations (BSDEs)
Kreps-Porteus, Duffie-Epstein, Duffie-Skiadas, Schroder-Skiadas, Skiadas,
El Karoui-Peng-Quenez, Lazrak and Quenez, Hamadene, Ma-Yong,
Kobylanski
• Ambiguity and robust optimization
Ellsberg, Chen-Epstein, Epstein-Schneider, Anderson et al.,
Hansen et al, Maenhout, Uppal-Wang, Skiadas
22
• Mental accounting and prospect theory
Discontinuous risk curvatureHuang-Barberis, Barberis et al., Thaler et al., Gneezy et al.
• Large trader models
Feedback effectsKyle, Platen-Schweizer, Bank-Baum, Frey-Stremme, Back, Cuoco-Cvitanic
• Social interactions
Continuous of agents – Propagation of frontsMalinvaud, Schelling, Glaesser-Scheinkman, Horst-Scheinkman, Foellmer
• Fund management
Non-zero sum stochastic differential gamesHuggonier-Kaniel
23
Bridging the gap between
optimal behavior and valuation
24
Pricing elements for contingent claims
“Extreme cases”
• Arbitrage free valuation theory: ν(CT ) = EQ(CT )
• Actuarial principles for insurance: ν(CT ) = U−1(EQ(U (CT )))
P historical measure, Q risk neutral measure,
U utility function
“General case”
• Pricing via equilibrium arguments and investment optimality
25
Arbitrary risks: CT = C(YT, ST )
Pricing from a perspective of optimal investment
• Use the market to assess both types of risk
• Formulate optimality criteria in terms of individual preferences
• Derive the concept of value from indifference to the various
investment opportunities
26
Issues
• There are hedgeable and unhedgeable risks to specify
• The hedgeable risks could be priced by arbitrage, and the hedgeable ones by
certainty equivalent
• Two probability measures, the nested risk neutral measure and the historical
measure, are then involved
27
Indifference valuation
• Optimal behavior with and without the claim
V 0(Xt, t) = supπ
EP (U (XT )/Ft)
V C(Xt, t) = supπ
EP (U (XT − CT )/Ft)
• Partial equilibrium
V 0(Xt, t) = V C(Xt + νt, t)
νt : indifference price process
Hodges-Neuberger, Constantinides-Z., Davis, Rouge-El Karoui, Hugonnieret al, Delbaen et all, Musiela-Z., Kramkov-Sirbu, Henderson,...
28
Structural result
Duality techniques yield
ν = supQ
(EQ(CT ) − ϑ(CT )),
ϑ(CT ) =1
γH(Q/P ) − inf
Q
(1
γH(Q/P )
)
Rouge–El Karoui, Frittelli, Kabanov and Stricker, Delbaen et al.
Considerations
• Price respresented via a non-intuitive optimization problem
• Pricing measure depends on the payoff
• Certain pricing elements are lost
29
Towards a constitutive analogue of
the Black and Scholes theory
30
Fundamental elements of an indifference pricing system
• Monotonicity, scaling and concavity with respect to payoffs
• Monotonicity, robustness and regularity with respect to risk aversion
• Consistency with the no-arbitrage principle
• Translation invariance with respect to replicable risks
• Risk quantification and monitoring
• Numeraire independence
• Additivity with respect to incremental risks
• Risk transfering across parties
31
Price representation and payoff decomposition
(Musiela-Z.)
• Complete markets
ν(CT , t) = EQ(CT/Ft)
CT = ν(CT , t) +∫ T
t
“∂ν(CT , s)
∂S
” dSs
Ss
• Incomplete markets
ν(CT , t) = EQ(CT/Ft)
CT = ν(CT , t) +∫ T
t
“∂ν(CT , s)
∂S
” dSs
Ss+ Rt,T
Rt,T : residual risk
32
Joint work with Marek Musiela (BNP Paribas, London)
References
• “A valuation algorithm for indifference prices in incomplete
Finance and Stochastics (2004)
• “An example of indifference prices under exponential preferences”,
Finance and Stochastics (2004)
• “Spot and forward dynamic utilities and their associated pricing systems:
Case study of the binomial model”
Indifference Pricing, PUP (2005)
33
A new probabilistic representation for indifference prices
The static case
• Arbitrage free prices ν(CT ) = EQ∗(CT )
E(.) : linear pricing functionalQ∗ : the (unique) risk neutral martingale measure
• Indifference prices ν(CT ) = EQ(CT )
E : pricing functional Q : pricing measure
nonlinear payoff independent
payoff independent preference independent
wealth independent
preference dependent
34
The indifference price
ν(CT ) = EQ
(1
γlog(EQ(eγC(ST ,YT ) | ST ))
)= EQ(CT )
Q(YT | ST ) = P(YT | ST )
35
Choice of pricing measure Q
• Q needs to be a martingale measure
• Q needs to preserve the conditional distribution of the unhedgeablerisks, given the hedgeable ones, from their historical values
Q(YT | ST ) = P(YT | ST )
Indifference price components
CT =1
γlog
(EP(eγC(ST ,YT ) |ST )
)
=1
γlog
(EQ(eγC(ST ,YT ) |ST )
)ν(CT ) = EQ(CT )
36
Valuation Procedure
ν(CT ) = EQ
(1γ log(EQ(eγC(ST ,YT ) | ST )
)= EQ(CT )
Q(YT | ST ) = P(YT | ST )
• Step 1: Specification, isolation and pricing of unhedgeable risks
The original payoff CT is altered to the preference adjusted payoff
CT =1
γlog
(EQ(eγC(ST ,YT ) |ST )
)Observe that CT = 1
γlog(EQ∗(eγC(ST ,YT ))) = 1
γlog(EP(eγC(ST ,YT )))
but CT =1
γlog(EQ(eγC(ST ,YT ) | ST )) =
1
γlog(EP(eγC(ST ,YT ) | ST ))
• Step 2: Pricing by arbitrage of the remaining hedgeable risks
37
Payoff decomposition
The static case
Risk quantification and monitoring for unhedgeable risks
• Optimal investments in the presence of the claim
αCT ,∗ = α0,∗ +∂ν(CT )
∂S0,
where
α0,∗ = −1
γ
∂H (Q |P)
∂S0
• Optimal wealths with and without the claim
XCT,∗T = x + ν(CT ) + αCT ,∗(ST − S0) and X
CT,∗0 = x + ν(CT )
X0,∗T = x + α0,∗(ST − S0) and X
0,∗T = x
38
• The optimal residual wealth
Lt = XCT,∗t − X
0,∗t for t = 0, T
Indifference hedge and replicability
L0 = ν(CT )
and
LT = ν(CT ) +∂ν(CT )
∂S0(ST − S0)
Conditional certainty equivalent and extraction of unhedgeable risks
LT = CT
Martingale property
EQ(LT ) = L0 = ν(CT ) for Q ∈ Qe
39
• The residual risk Rt
Rt = Ct − Lt for t = 0, T
Nonreplicable components of the claim and of the residual risk
R0 = 0 and RT = CT − CT ; RT = 0
The indifference price of the residual risk
ν(RT ) = 0
Supermartingale property under the pricing measure Q
EQ(RT ) ≤ Rt = 0
Actuarial certainty equivalent
1
γlog EP(eγRT ) = 0
40
Residual risk decomposition
The supermartingale Rt, for t = 0, T , admits the decomposition
Rt = Rmt + Rd
t
where
Rm0 = 0 and Rm
T = RT − EQ(RT ),
and
Rd0 = 0 and Rd
T = EQ(RT ).
The component Rmt is an F (S,Y )
T -martingale under Q while Rdt is decreasing
and adapted to the trivial filtration F (S,Y )0 .
41
Payoff decomposition
Let CT and RT be, respectively, the conditional certainty equivalent and the
residual risk associated with the claim CT . Let also Rmt and Rd
t be the Doob
decomposition components of the residual risk supermartingale.
Define the process MCt , for t = 0, T , by
MC0 = ν (CT ) and MC
T = ν (CT ) +∂ν (CT )
∂S0(ST − S0)
The claim CT admits the unique, under Q, payoff decomposition
CT = CT + RT
= ν(CT ) +∂ν(CT )
∂S0(ST − S0) + RT = MC
T + RmT + Rd
T
42
Convex risk measures ρ : FT → R
Positions: C1T , C2
T ∈ FT
Properties
Convexity: ρ(αC1T + (1 − α)C2
T ) ≤ αρ(C1T ) + (1 − α)ρ(C2
T )
Monotonicity: if C1T ≤ C2
T then ρ(C1T ) ≥ ρ(C2
T )
Translation invariance w.r.t. replicable risks: ρ(C1T + m) ≤ ρ(C1
T ) − m
Artzner et al., Foellmer and Shied, Frittelli and Gianini, Delbaen, ...
Convex measures are represented as expectations with penalty terms
(entropic terms-special case)
43
Claim: CT ∈ FT
ρ(CT ) = ν (−CT ) = EQ
(1γ log EQ
(e−γCT |ST
))ρ(CT ) : Capital requirement in order to accept the position CT
ν (CT ) : Indifference value of the claim CT
The concept of indifference price, deduced from the desire to behaveoptimally as an investor, is consistent with an axiomatic approachused to determine the capital amount, for a position to be acceptedby a supervising body.
• Convex risk measures yield the capital requirement but not the relevant riskmonitoring implementation strategy
• Convex risk measures are traslation invariant w.r.t. to constant risks
• Convex risk measures are static
44
Dynamic models and the term
structure of risk preferences
45
The dynamic case
• Specification of the non linear price functional that yields the indifference
price representation
νt(CT ) = E(t,T )Q (CT/Ft)
• Semigroup property
νt(CT ) = E(t,t′)Q (νt′(CT )/Ft) = E(t,t′)
Q (E(t′,T )Q (CT )/Ft)
• Payoff decomposition
• Residual risk and error quantification
• Proper specification of preferences
46
A reduced binomial model
47
The multiperiod model
• Traded asset: St, t = 0, 1, ..., T (St > 0, ∀t)
ξt+1 =St+1
St, ξt+1 = ξd
t+1, ξut+1 with 0 < ξd
t+1 < 1 < ξut+1
Second traded asset is riskless yielding zero interest rate
• Non-traded asset: Yt, t = 0, 1, ..., T
ηt+1 =Yt+1
Yt, ηt+1 = ηd
t+1, ηut+1 with ηd
t < ηut
St, Yt : t = 0, 1, ..., T : a two-dimensional stochastic process
• Probability space (Ω, (Ft) , P)
Filtrations FSt and FY
t : generated by the random variables Ss (ξs)
and Ys (ηs), for s = 0, 1, . . . , t.
48
Indifference pricing mechanism
• State wealth process: Xs, s = t + 1, ..., T
αs, s = t + 1, t + 2, ..., T : the number of shares of the traded asset held
in this portfolio over the time period [s − 1, s]
XT = x +T∑
s=t+1
αs Ss
• Claim CT (Path dependence/early exercise are allowed)
• Value function: V CT (Xt, t; T ) = supαt+1,...,αT
EP
(−e−γ(XT−CT ) | Ft
)• Indifference price: νt(CT )
V 0(Xt, t; T ) = V CT (Xt + νt(CT ), t; T )
49
Fundamental multiperiod pricing blocks
Let Zs, 0 ≤ s ≤ t be an Fs-adapted process and Q a martingale measure
E(t,s)Q (Zs) E(t,s−1)
Q (E(s−1,s)Q (Zs))
where
E(s−1,s)Q (Zs) EQ
(1γ log EQ(eγZs|Fs−1 ∨ FS
s )|Fs−1
)and
E(s,s)Q (Zs) = Zs
• Note that for t < s − 1
E(t,s)Q (Zs) = EQ
(1
γlog EQ(Zs|Ft ∨ FS
s )|Ft
)
50
The valuation algorithm
Let Q be a martingale measure satisfying, for t = 0, 1, . . . , T
Q(ηt+1 | Ft ∨ FSt+1) = P(ηt+1 | Ft ∨ FS
t+1)
• The indifference price νt(CT ) satisfies
νt(CT ) = E(t,t+1)Q (νt+1(CT )) νT (CT ) = CT
• The indifference price process is given by
νt(CT ) = E(t,T )Q (CT )
• The pricing algorithm is consistent across time
νt(CT ) = E(t,s)Q (E(s,T )
Q (CT )) = E(t,s)Q ((νs(CT )) = νt(E(s,T )
Q (CT ))
51
Interpretation of the pricing algorithm
Valuation is done via an iterative pricing schemeapplied backwards in time.
νt(CT ) = E (t,t+1)Q (E (t+1,t+2)
Q (...(E (T−1,T )Q (CT )))
νt(CT ) = E (t,t+1)Q (νt+1(CT ))
– Specification, isolation and pricing of unhedgeable risks
νt+1(CT ) : the end of the period payoff is altered to thepreference adjusted payoff
νt+1(CT ) =1
γlog EQ
(eγνt+1(CT )
∣∣∣Ft ∨ FSt+1
)– Pricing of remaining hedgeable risks by arbitrage
νt(CT ) = EQ(νt+1(CT ))
52
Indifference prices and quasilinear pdes
53
Continuous time reduced models
The diffusion case
• Probabilistic setting
– Traded asset: St, dSs = µ(Ss, s)Ssds + σ(Ss, s)SsdW 1s
– Riskless bond Bt
– Nontraded asset: Yt, dYs = b(Ys, s)ds + a(Ys, s)dWs
– Probability space: (Ω,F , (Fs), P)
Fs = σ(W 1u, Wu; 0 ≤ u ≤ s) d(W 1,W )(s) = ρds
– Claim: CT = C(ST, YT )
54
Indifference price representation
The indifference price process is given by
νs(CT ) = H(Ss, Ys, s)
where H(S, y, t) is the unique viscosity solution of the quasilinear price equation
Ht + L(H(S, y, t)) +1
2γ(1 − ρ2)a2(y, t)H2
y = 0
H(S, y, T ) = C(S, y)
The operator L(.) is given by
L(.) =1
2σ2(S, t)
ϑ2
ϑS2 + ρa(y, t)σ(S, t)ϑ2
ϑSϑy+
1
2a2(S, t)
ϑ2
ϑy2
+
(b(y, t) − ρ
µ(S, t)
σ(S, t)a(y, t)
)ϑ
ϑy
55
• Observation 1
i) H may be directly related to a quadratic stochastic control problem
ii) H may be related to a BSDE
But, there is no direct Feynman-Kac type representation of H available up
to now!
• Observation 2
Solutions of the above equation are directly structurally related
to optimal investment policies in models with stochastic Sharpe
ratio, recursive utilities, and stochastic labor income
Stoikov-Z.
56
Special case
(Musiela-Z., Henderson)
• Lognormal dynamics for the traded asset St :
Ss = µSsds + σSsdW 1s
• Payoff depending exclusively on the nontraded asset Yt
C(ST , YT ) = C(YT )
H(y, t) =1
γ(1 − ρ2)ln(EQ(eγ(1−ρ2)C(YT ) | Yt = y)
)
The above formula is not a naive extension of the classical certainty
equivalent price!
57
Valuation algorithm
• Pricing measure Q: a martingale measure satisfying
Q(Yt+dt
∣∣∣Ft ∨ FSt,t+dt) = P(Yt+dt
∣∣∣Ft ∨ FSt,t+dt)
The indifference price νt(Ss, Ys, s) is given by the iterative algorithm
νt(Ss, Ys, s) =
= EQ
(1γ log EQ(eγν(Ss+ds,Ys+ds,s+ds) | Fs ∨ FS
s,s+ds) | Fs
)νT (ST , YT , T ) = C(ST , YT )
58
A novel probabilistic representation of solutionsto quasilinear pdes arises
The solution H(S, y, t) of the quasilinear price equation
Ht + L(H(S, y, t)) +1
2γ(1 − ρ2)a2(y, t)H2
y = 0
with
L(.) = 12σ
2(S, t) ϑ2
ϑS2 + ρa(y, t)σ(S, t) ϑ2
ϑSϑy + 12a
2(S, t) ϑ2
ϑy2
+(b(y, t) − ρ
µ(S,t)σ(S,t)
a(y, t))
ϑϑy
is given byH(S, y, t) = limds→0,s→t νt(Ss+ds, Ys+ds, s + ds)
Proof: Based on convergence results of supermartingales, behavior of nonadditivemeasures and robustness properties of viscosity solutions.
59
Extensions
• Early exercise claims: Cτ = C(Sτ , Yτ ) ; τ ∈ F[0,T ]
νat (Cτ ) = sup
τ∈F[0,T ]
νt(Cτ ) = supτ∈F[0,T ]
(E(t,τ )Q (Cτ ))
(Musiela-Z., Oberman-Z., Sokolova)
• Path dependent claims: C(t,T ) = C(∫ T
tc(Ss, Ys, s)ds; ω)
• Stochastic interest rates (infinite dimensional problems/PSDEs)
(Tehranchi and Zariphopoulou)
• Defaultable claims (Bielecki et al.)
60
Stochastic modelling and indifference prices
• Quasilinear PDEs
• BSDEs
• Stochastic PDEs
• Non-linear expectations
61
Utility-based pricing systems
for general market models
62
Important issues
• Traditional utilities are “fixed” w.r.t. a given horizon but we need to be able
to price claims across all maturities.
• Typically, risk aversion coefficient is chosen in isolation from the market
environment.
• Numeraire independence and semigroup property of prices imposes
endogeneous restrictions on preferences.
63
Towards a utility-based pricing system
Ingredients for a term structure of risk preferences that ultimately
yields consistent prices across units, numeraires, investment horizons
and maturities
• Stochastic risk tolerance
• Time evolution of preferences that captures market incompleteness
This leads to the new notions of
Forward and Spot Dynamic Utilities (Musiela-Z., 2005)
64
The diffusion case
• Market dynamics
dSs = µ(Ys, s)Ss ds + σ(Ys, s)Ss dW 1s
dYs = b(Ys, s) ds + a(Ys, s) dWs
ρ = cor(W 1,W ), (Ω,F , P), λs(Ys, s) =µ(Ys, s)
σ(Ys, s)
• Minimal relative entropy measure: Q
dQ
dP= exp
(−∫ T
0λs dW 1
s −∫ T
0λ1
s dW 1,⊥s − 1
2
∫ T
0(λ2
s + (λ⊥s )2) ds
)
λ⊥s = λ⊥s (Ys, s); λ⊥(y, t) ∼ gradient to the sln of a quasilinear pde
(Hobson, Rheinlander, Stoikov-Z., Benth-Karlsen)
65
Spot and forward dynamic utilities
• Forward dynamic utility
For T = 1, 2.., the processUf (x, t; T ) : t = 0, 1, ..., T
defined,
for x ∈ R, by
Uf (x, t; T ) =
⎧⎪⎨⎪⎩−e−γx if t = T
−e−γx−HT (Q(·|Ft)| P(·|Ft)) if 0 ≤ t ≤ T − 1
with
HT (Q(·|Ft)/P(·|Ft)) = EQ
(∫ T
t
1
2(λ2
s + (λ⊥s )2) ds/Ft
)
is called the forward to time T , or, normalized at time T ,dynamic exponential utility
66
• Spot dynamic utility
For s = 0, 1, .., the process Us (x, t) : t ≥ s, ... defined,for x ∈ R, by
Us (x, t; s) =
⎧⎪⎪⎨⎪⎪⎩−e−γx if t = s
−e−γx+∫ ts
12λ
2u du if t > s
is called the spot, normalized at time s, dynamic exponential utility.
Spot and forward dynamic utilities
• Uf aggregates information
• Us follows information along the path
67
Spot and forward dynamic value functions
Forward dynamic value function
Let Uf be the forward dynamic utility process, normalized at T . The associated
forward dynamic value function V f is defined, for x ∈ R, s = 0, 1, .. and
s ≤ t ≤ T , as
V f (x, s, t; T ) = supπ
EP
(Uf (Xt, t; T ) |Ft
)with
dXt = µtπt dt + σtπt dW 1t , t ≥ s
and Xs = x.
68
Spot dynamic value function
Let Us be the spot dynamic utility process, normalized at s ≥ 0.
The associated spot dynamic value function V s is defined, for x ∈ R,
and s ≤ t ≤ T as
V s (x, t, T ) = supπ
EP (Us (XT, T ) |Ft)
with
dXu = µuπu du + σuπu dW 1u
and Xt = x
69
Fundamental properties of the forward dynamic value function
The forward dynamic value function V f has the following properties:
V f (x, T, T ; T ) = −e−γx for x ∈ R
V f (x, s, t1; T ) = V f (x, s, t2; T ) for t1 = t2 and s ≤ min (t1, t2)
V f (x, s, t; T ) = V f (x, s, s; T ) = −e−γx−HT (Q(·|Ft)/P(·|Ft))
The forward dynamic value function V f coincides with the forwarddynamic utility
V f (x, s, t; T ) = Uf (x, s; T ) for x ∈ R and s ≤ t ≤ T
70
Fundamental properties of the spot dynamic value function
The spot dynamic value function V f has the following properties:
V s (x, s, s) = −e−γx for x ∈ R
V s (x, t, T1) = V s (x, t, T2) for T1 = T2 and s ≤ t ≤ min (T1, T2)
V s (x, t, T ) = V s (x, t, t) = −e−γx+∫ ts
12λ
2u du
The spot dynamic value function V s coincides
with the spot dynamic utility
V s (x, t, T ) = Us (x, T ) for x ∈ R and s ≤ t ≤ T
71
Forward indifference price
• Forward dynamic utility Uf , normalized at time T
• Let T ≤ T and consider a claim, written at time t0 ≥ 0, yielding payoff CTat time T , with CT ∈ FT
• Let, also, V f,CT and V f,0 be, respectively, the forward dynamic value func-
tions with and without the claim
For t0 ≤ t ≤ T , the forward indifference price, associated with the forward,
to time T, utility Uf, is defined as the amount νf (CT , t; T ) for which
V f,0(x, t, T ; T
)= V f,CT
(x + νf (CT , t; T ), t, T ; T
)for x ∈ R
72
Spot indifference price
• Spot dynamic utility Us, normalized at time s
• Consider a claim, written at time t0 ≥ s, yielding payoff CT at time T , with
CT ∈ FT
• Let, also, V s,CT and V s,0 be, respectively, the spot dynamic value functions
with and without the claim
For t0 ≤ t ≤ T , the spot indifference price, associated with the spot,
normalized at time s, utility Us, is defined as the amount νs(CT , t; s) for which
V s,0 (x, t, T ) = V s,CT (x + νs(CT , t), t, T ) for x ∈ R
73
Important observation
The spot and forward indifference prices do not in general coincide,
i.e., for
s ≤ t0 ≤ t ≤ T ≤ T
νs(CT , t; s) = νf (CT , t; T )
74
Consistency across maturities
Both V f and V s are independent of the investment horizon
Spot and forward parity
−νs
(∫ T
t
1
2
λ2u
γdu, t
)= νf
(∫ T
t−1
2
λ2u
γdu; T
)
νs
(∫ T
tCT , t
)+ νs
(∫ T
t−1
2
λ2u
γdu, t
)= νf
(CT −
∫ T
t
1
2
λ2u
γdu, t; T
)
Spot and foward prices coincide if Sharpe ratios are replicable.
75
Extensions
• Forward and spot prices in Markovian models
Connection with non-linear expectations
• Forward and spot prices for stochastic risk tolerance
The term structure of utilities, stochastic pdes
• Indifference pricing systems and contract theory
Zero- and Non zero-sum stochastic differential games
• Infinite dimensional models
76