Business Principles, Tools, and Techniques in Participating in Various Types...
Abstract, Classic, and Explicit Turnpikes
1. Problem Abstract Diffusions
Portfolio Turnpikes for Incomplete Markets
Paolo Guasoni1,2
Kostas Kardaras1 Scott Robertson3 Hao Xing4
1 Boston University
2 Dublin City University
3 Carnegie Mellon University
4 London School of Economics
Princeton ORFE Seminar
September 22nd , 2010
2. Problem Abstract Diffusions
Outline
• Turnpike Theorems:
for Long Horizons, use Constant Relative Risk Aversion.
3. Problem Abstract Diffusions
Outline
• Turnpike Theorems:
for Long Horizons, use Constant Relative Risk Aversion.
• Results:
Abstract, Classic, and Explicit Turnpikes.
4. Problem Abstract Diffusions
Outline
• Turnpike Theorems:
for Long Horizons, use Constant Relative Risk Aversion.
• Results:
Abstract, Classic, and Explicit Turnpikes.
• Consequences:
Risk Sensitive Control and Intertemporal Hedging.
5. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
6. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
7. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
8. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
9. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
• Turnpike theorems: (under some conditions)
as T increases, the optimal portfolio for U is close to the optimal
portfolio for either power or log utility (CRRA).
10. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
• Turnpike theorems: (under some conditions)
as T increases, the optimal portfolio for U is close to the optimal
portfolio for either power or log utility (CRRA).
• The power depends on the properties of U at large wealth levels.
11. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
• Turnpike theorems: (under some conditions)
as T increases, the optimal portfolio for U is close to the optimal
portfolio for either power or log utility (CRRA).
• The power depends on the properties of U at large wealth levels.
• Different papers find different conditions.
12. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
• Turnpike theorems: (under some conditions)
as T increases, the optimal portfolio for U is close to the optimal
portfolio for either power or log utility (CRRA).
• The power depends on the properties of U at large wealth levels.
• Different papers find different conditions.
• Conditions involve preferences and market structure.
13. Problem Abstract Diffusions
Portfolio Turnpikes
• An investor with utility U...
• ...invests optimally for a terminal wealth at horizon T .
• As the horizon increases, today’s optimal portfolio...
• ...converges? To what?
• Turnpike theorems: (under some conditions)
as T increases, the optimal portfolio for U is close to the optimal
portfolio for either power or log utility (CRRA).
• The power depends on the properties of U at large wealth levels.
• Different papers find different conditions.
• Conditions involve preferences and market structure.
• Literature:
conditions neither more nor less general that others.
14. Problem Abstract Diffusions
Literature
Mossin (1968) JB IID Disc −U /U = ax + b
Leland (1972) Proc IID Disc −U /U = ax + f (x)
Ross (1974) JFE IID Disc U sum of powers
(x−a)p p
Hakansson (1974) JFE IID Disc p
−A(p)<U(x)< (x+a) +A(p)
p
Huberman Ross (1983) EC IID Disc p>0, bounded below, U’ reg. var
Cox Huang (1992) JEDC IID Compl Cont |U −1 − A1 y −1/b | ≤ A2 y −a
Jin (1997) JEDC IID Compl Cont |U −1 − A1 y −1/b | ≤ A2 y −a
U0 (x)
Dybvig et al. (1999) RFS Compl Cont U1 (x)
→K
U0 (x)
Huang Zariph. (1999) FS IID Compl Cont x p−1
→ K , U(0) = 0
• Either IID returns, or market completeness, or both.
15. Problem Abstract Diffusions
Literature
Mossin (1968) JB IID Disc −U /U = ax + b
Leland (1972) Proc IID Disc −U /U = ax + f (x)
Ross (1974) JFE IID Disc U sum of powers
(x−a)p p
Hakansson (1974) JFE IID Disc p
−A(p)<U(x)< (x+a) +A(p)
p
Huberman Ross (1983) EC IID Disc p>0, bounded below, U’ reg. var
Cox Huang (1992) JEDC IID Compl Cont |U −1 − A1 y −1/b | ≤ A2 y −a
Jin (1997) JEDC IID Compl Cont |U −1 − A1 y −1/b | ≤ A2 y −a
U0 (x)
Dybvig et al. (1999) RFS Compl Cont U1 (x)
→K
U0 (x)
Huang Zariph. (1999) FS IID Compl Cont x p−1
→ K , U(0) = 0
• Either IID returns, or market completeness, or both.
• Disparate conditions on utility functions.
16. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
17. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
18. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
19. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
20. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
• Classic turnpike:
convergence of portfolios under physical probability.
21. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
• Classic turnpike:
convergence of portfolios under physical probability.
• Abstract turnpike implies classic turnpike if myopic IID optimum.
22. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
• Classic turnpike:
convergence of portfolios under physical probability.
• Abstract turnpike implies classic turnpike if myopic IID optimum.
• More results for diffusion model with many assets but one state.
23. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
• Classic turnpike:
convergence of portfolios under physical probability.
• Abstract turnpike implies classic turnpike if myopic IID optimum.
• More results for diffusion model with many assets but one state.
• Classic turnpike for diffusions.
24. Problem Abstract Diffusions
This Paper
• Relax assumptions on market completeness and IID returns.
• Use condition on marginal utility ratio for U.
• Abstract turnpike:
convergence of portfolios under myopic probabilities PT .
• Holds under minimal conditions on market structure.
• Classic turnpike:
convergence of portfolios under physical probability.
• Abstract turnpike implies classic turnpike if myopic IID optimum.
• More results for diffusion model with many assets but one state.
• Classic turnpike for diffusions.
• Explicit turnpike:
limit portfolio is solution to ergodic HJB equation.
25. Problem Abstract Diffusions
Preferences
• Two investors. One with utility U, the other with CRRA 1 − p.
26. Problem Abstract Diffusions
Preferences
• Two investors. One with utility U, the other with CRRA 1 − p.
• Marginal Utility Ratio measures how close they are:
U (x)
R(x) := , x >0
x p−1
27. Problem Abstract Diffusions
Preferences
• Two investors. One with utility U, the other with CRRA 1 − p.
• Marginal Utility Ratio measures how close they are:
U (x)
R(x) := , x >0
x p−1
Assumption
U : R+ → R continuously differentiable, strictly increasing, strictly
concave, satisfies Inada conditions U (0) = ∞ and U (∞) = 0.
Marginal utility ratio satisfies:
lim R(x) = 1, (CONV)
x↑∞
0 < lim inf R(x), 0 = p < 1, (LB-0)
x↓0
lim sup R(x) < ∞, p < 1. (UB-0)
x↓0
28. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
29. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
30. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
Assumption
For T > 0, X T is a set of nonnegative semimartingales such that:
31. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
Assumption
For T > 0, X T is a set of nonnegative semimartingales such that:
i) X0 = 1 for all X ∈ X T ;
32. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
Assumption
For T > 0, X T is a set of nonnegative semimartingales such that:
i) X0 = 1 for all X ∈ X T ;
ii) X T contains a strictly positive X (Xt > 0 a.s. for all t ∈ [0, T ]);
33. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
Assumption
For T > 0, X T is a set of nonnegative semimartingales such that:
i) X0 = 1 for all X ∈ X T ;
ii) X T contains a strictly positive X (Xt > 0 a.s. for all t ∈ [0, T ]);
iii) X T is convex: ((1 − α)X + αX ) ∈ X T for X , X ∈ X T , α ∈ [0, 1];
34. Problem Abstract Diffusions
Market Structure
• Investors choose from a common set X T of wealth processes.
• (Ω, (Ft )t∈[0,T ] , F T , P) filtered probability space. Usual conditions.
Assumption
For T > 0, X T is a set of nonnegative semimartingales such that:
i) X0 = 1 for all X ∈ X T ;
ii) X T contains a strictly positive X (Xt > 0 a.s. for all t ∈ [0, T ]);
iii) X T is convex: ((1 − α)X + αX ) ∈ X T for X , X ∈ X T , α ∈ [0, 1];
iv) X T stable under compounding: if X , X ∈ X T with X strictly positive
and τ is a [0, T ]-valued stopping time, then X T contains the
process X that compounds X with X at τ :
Xτ Xt (ω), if t ∈ [0, τ (ω)[
X = X I[[0,τ [[ +X I =
Xτ [[τ,T ]] (Xτ (ω)/Xτ (ω)) Xt (ω), if t ∈ [τ (ω), T ]
35. Problem Abstract Diffusions
Well Posedness and Growth
• Use index 0 for the CRRA investor, and index 1 for investor with U.
36. Problem Abstract Diffusions
Well Posedness and Growth
• Use index 0 for the CRRA investor, and index 1 for investor with U.
• Maximization problems:
u 0,T = sup EP [X p /p] , u 1,T = sup EP [U (X )] .
X ∈X T X ∈X T
37. Problem Abstract Diffusions
Well Posedness and Growth
• Use index 0 for the CRRA investor, and index 1 for investor with U.
• Maximization problems:
u 0,T = sup EP [X p /p] , u 1,T = sup EP [U (X )] .
X ∈X T X ∈X T
• Well posedness:
38. Problem Abstract Diffusions
Well Posedness and Growth
• Use index 0 for the CRRA investor, and index 1 for investor with U.
• Maximization problems:
u 0,T = sup EP [X p /p] , u 1,T = sup EP [U (X )] .
X ∈X T X ∈X T
• Well posedness:
Assumption
−∞ < u i,T < ∞ and optimal payoffs X i,T exist for all T > 0 and i = 0, 1.
39. Problem Abstract Diffusions
Central Objects
• Ratio of optimal wealth processes and its stochastic logarithm:
1,T u T
T Xu drv
ru := 0,T
, ΠT :=
u T
, for u ∈ [0, T ].
Xu 0 rv −
40. Problem Abstract Diffusions
Central Objects
• Ratio of optimal wealth processes and its stochastic logarithm:
1,T u T
T Xu drv
ru := 0,T
, ΠT :=
u T
, for u ∈ [0, T ].
Xu 0 rv −
T
• r0 = 1 (investors have same initial capital).
41. Problem Abstract Diffusions
Central Objects
• Ratio of optimal wealth processes and its stochastic logarithm:
1,T u T
T Xu drv
ru := 0,T
, ΠT :=
u T
, for u ∈ [0, T ].
Xu 0 rv −
T
• r0 = 1 (investors have same initial capital).
• myopic probabilities PT T ≥0 :
p
0,T
dPT XT
= p
.
dP 0,T
EP XT
42. Problem Abstract Diffusions
Central Objects
• Ratio of optimal wealth processes and its stochastic logarithm:
1,T u T
T Xu drv
ru := 0,T
, ΠT :=
u T
, for u ∈ [0, T ].
Xu 0 rv −
T
• r0 = 1 (investors have same initial capital).
• myopic probabilities PT T ≥0 :
p
0,T
dPT XT
= p
.
dP 0,T
EP XT
• Myopic probabilities PT boil down to P for log utility.
43. Problem Abstract Diffusions
Central Objects
• Ratio of optimal wealth processes and its stochastic logarithm:
1,T u T
T Xu drv
ru := 0,T
, ΠT :=
u T
, for u ∈ [0, T ].
Xu 0 rv −
T
• r0 = 1 (investors have same initial capital).
• myopic probabilities PT T ≥0 :
p
0,T
dPT XT
= p
.
dP 0,T
EP XT
• Myopic probabilities PT boil down to P for log utility.
• Optimal payoff for x p /p under P equal to log optimal under P.
44. Problem Abstract Diffusions
Growth
• Growth. As horizon increases, increasingly large payoffs available:
45. Problem Abstract Diffusions
Growth
• Growth. As horizon increases, increasingly large payoffs available:
Assumption
ˆ ˆ
There exists a family (X T )T ≥0 such that X T ∈ X T and:
ˆ
lim PT (X T ≥ N) = 1 for any N > 0. (GROWTH)
T →∞
46. Problem Abstract Diffusions
Growth
• Growth. As horizon increases, increasingly large payoffs available:
Assumption
ˆ ˆ
There exists a family (X T )T ≥0 such that X T ∈ X T and:
ˆ
lim PT (X T ≥ N) = 1 for any N > 0. (GROWTH)
T →∞
• Assumption trivially satisfied with a positive safe rate.
47. Problem Abstract Diffusions
Growth
• Growth. As horizon increases, increasingly large payoffs available:
Assumption
ˆ ˆ
There exists a family (X T )T ≥0 such that X T ∈ X T and:
ˆ
lim PT (X T ≥ N) = 1 for any N > 0. (GROWTH)
T →∞
• Assumption trivially satisfied with a positive safe rate.
• Holds in more generality.
48. Problem Abstract Diffusions
Growth
• Growth. As horizon increases, increasingly large payoffs available:
Assumption
ˆ ˆ
There exists a family (X T )T ≥0 such that X T ∈ X T and:
ˆ
lim PT (X T ≥ N) = 1 for any N > 0. (GROWTH)
T →∞
• Assumption trivially satisfied with a positive safe rate.
• Holds in more generality.
• But note PT , not P!
49. Problem Abstract Diffusions
Abstract Turnpike
Theorem (Abstract Turnpike)
Let previous assumptions hold. Then, for any > 0,
50. Problem Abstract Diffusions
Abstract Turnpike
Theorem (Abstract Turnpike)
Let previous assumptions hold. Then, for any > 0,
T
a) limT →∞ PT supu∈[0,T ] ru − 1 ≥ = 0,
51. Problem Abstract Diffusions
Abstract Turnpike
Theorem (Abstract Turnpike)
Let previous assumptions hold. Then, for any > 0,
T
a) limT →∞ PT supu∈[0,T ] ru − 1 ≥ = 0,
b) limT →∞ PT ΠT , Π T T
≥ =0
52. Problem Abstract Diffusions
Abstract Turnpike
Theorem (Abstract Turnpike)
Let previous assumptions hold. Then, for any > 0,
T
a) limT →∞ PT supu∈[0,T ] ru − 1 ≥ = 0,
b) limT →∞ PT ΠT , Π T T
≥ =0
• For log utility PT ≡ P, hence convergence holds under P.
53. Problem Abstract Diffusions
Abstract Turnpike
Theorem (Abstract Turnpike)
Let previous assumptions hold. Then, for any > 0,
T
a) limT →∞ PT supu∈[0,T ] ru − 1 ≥ = 0,
b) limT →∞ PT ΠT , Π T T
≥ =0
• For log utility PT ≡ P, hence convergence holds under P.
• For a familiar diffusion dSu /Su = µu du + σu dWu , [ΠT , ΠT ]
measures distance between portfolios π 1,T and π 0,T :
·
1,T 0,T 1,T 0,T
ΠT , ΠT = πu − πu Σu πu − πu du,
· 0
54. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
then, for any > 0 and t ≥ 0:
55. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
then, for any > 0 and t ≥ 0:
56. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
57. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
T
a) limT →∞ P supu∈[0,t] ru − 1 ≥ = 0,
58. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
T
a) limT →∞ P supu∈[0,t] ru − 1 ≥ = 0,
b) limT →∞ P Π T , ΠT t
≥ = 0.
59. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
T
a) limT →∞ P supu∈[0,t] ru − 1 ≥ = 0,
b) limT →∞ P Π T , ΠT t
≥ = 0.
• If optimal wealth myopic with IID returns, abstract implies classic.
60. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
T
a) limT →∞ P supu∈[0,t] ru − 1 ≥ = 0,
b) limT →∞ P Π T , ΠT t
≥ = 0.
• If optimal wealth myopic with IID returns, abstract implies classic.
• In practice, if assets have IID returns, optimal portfolio myopic.
61. Problem Abstract Diffusions
IID Myopic Turnpike
Corollary (IID Myopic Turnpike)
If, in addition to previous assumptions:
i) XtT = XtS ≡ Xt a.s. for all t ≤ S, T (myopic optimality);
ii) Xt and XT /Xt are independent for all t ≤ T (independent returns).
then, for any > 0 and t ≥ 0:
T
a) limT →∞ P supu∈[0,t] ru − 1 ≥ = 0,
b) limT →∞ P Π T , ΠT t
≥ = 0.
• If optimal wealth myopic with IID returns, abstract implies classic.
• In practice, if assets have IID returns, optimal portfolio myopic.
• For example, Levy processes.
62. Problem Abstract Diffusions
Diffusion Model
• One state variable Y , with values in interval E = (α, β) ⊆ R, with
−∞ ≤ α < β ≤ ∞.
dYt = b(Yt ) dt + a(Yt ) dWt .
63. Problem Abstract Diffusions
Diffusion Model
• One state variable Y , with values in interval E = (α, β) ⊆ R, with
−∞ ≤ α < β ≤ ∞.
dYt = b(Yt ) dt + a(Yt ) dWt .
• Market includes safe rate r (Yt ) and d risky assets with prices:
dSti
= r (Yt ) dt + dRti , 1 ≤ i ≤ d,
Sti
64. Problem Abstract Diffusions
Diffusion Model
• One state variable Y , with values in interval E = (α, β) ⊆ R, with
−∞ ≤ α < β ≤ ∞.
dYt = b(Yt ) dt + a(Yt ) dWt .
• Market includes safe rate r (Yt ) and d risky assets with prices:
dSti
= r (Yt ) dt + dRti , 1 ≤ i ≤ d,
Sti
• Cumulative excess return R = (R 1 , · · · , R d ) follows diffusion:
n
dRti = µi (Yt ) dt + σij (Yt ) dZtj , 1 ≤ i ≤ d,
j=1
65. Problem Abstract Diffusions
Diffusion Model
• One state variable Y , with values in interval E = (α, β) ⊆ R, with
−∞ ≤ α < β ≤ ∞.
dYt = b(Yt ) dt + a(Yt ) dWt .
• Market includes safe rate r (Yt ) and d risky assets with prices:
dSti
= r (Yt ) dt + dRti , 1 ≤ i ≤ d,
Sti
• Cumulative excess return R = (R 1 , · · · , R d ) follows diffusion:
n
dRti = µi (Yt ) dt + σij (Yt ) dZtj , 1 ≤ i ≤ d,
j=1
• W and Z = (Z 1 , · · · , Z n ) are multivariate Wiener processes with
correlation ρ = (ρ1 , · · · , ρn ) , i.e. d Z i , W t = ρi (Yt ) dt for 1 ≤ i ≤ n.
66. Problem Abstract Diffusions
Regularity Conditions
Assumption
Set Σ = σσ , A = a2 , and Υ = σρa. r ∈ C γ (E, R), b ∈ C 1,γ (E, R),
µ ∈ C 1,γ (E, Rd ), A ∈ C 2,γ (E, R), Σ ∈ C 2,γ (E, Rd×d ), and
Υ ∈ C 2,γ (E, Rd ). For all y ∈ E, Σ is positive and A is strictly positive.
67. Problem Abstract Diffusions
Regularity Conditions
Assumption
Set Σ = σσ , A = a2 , and Υ = σρa. r ∈ C γ (E, R), b ∈ C 1,γ (E, R),
µ ∈ C 1,γ (E, Rd ), A ∈ C 2,γ (E, R), Σ ∈ C 2,γ (E, Rd×d ), and
Υ ∈ C 2,γ (E, Rd ). For all y ∈ E, Σ is positive and A is strictly positive.
Assumption
˜ Σ Υ ˜ µ
A= b= . Infinitesimal generator of (R, Y ):
Υ A b
2
L = 2 d+1 Aij (ξ) ∂ξ∂∂ξj + i=1 bi (ξ) ∂ξi
1
i,j=1
˜
i
d+1 ˜ ∂
Martingale problem for L well posed, in that unique solution exists.
68. Problem Abstract Diffusions
Regularity Conditions
Assumption
Set Σ = σσ , A = a2 , and Υ = σρa. r ∈ C γ (E, R), b ∈ C 1,γ (E, R),
µ ∈ C 1,γ (E, Rd ), A ∈ C 2,γ (E, R), Σ ∈ C 2,γ (E, Rd×d ), and
Υ ∈ C 2,γ (E, Rd ). For all y ∈ E, Σ is positive and A is strictly positive.
Assumption
˜ Σ Υ ˜ µ
A= b= . Infinitesimal generator of (R, Y ):
Υ A b
2
L = 2 d+1 Aij (ξ) ∂ξ∂∂ξj + i=1 bi (ξ) ∂ξi
1
i,j=1
˜
i
d+1 ˜ ∂
Martingale problem for L well posed, in that unique solution exists.
Assumption
ρ ρ is constant (does not depend on y ), and supy ∈E c(y ) < ∞,
c(y ) := 1 (pr (y ) − q µ Σ−1 µ(y )) for y ∈ E, q := p−1 , and δ := 1−qρ ρ .
δ 2
p 1
69. Problem Abstract Diffusions
HJB Assumption (finite horizon)
Assumption
There exist (v T (y , t))T >0 and v (y ) such that:
ˆ
70. Problem Abstract Diffusions
HJB Assumption (finite horizon)
Assumption
There exist (v T (y , t))T >0 and v (y ) such that:
ˆ
i) v T > 0, v T ∈ C 1,2 ((0, T ) × E), and solves reduced HJB equation:
∂t v + Lv + c v = 0, (t, y ) ∈ (0, T ) × E,
v (T , y ) = 1, y ∈ E,
where L := 1 A ∂yy + B ∂y and B := b − qΥ Σ−1 µ.
2
2
71. Problem Abstract Diffusions
HJB Assumption (finite horizon)
Assumption
There exist (v T (y , t))T >0 and v (y ) such that:
ˆ
i) v T > 0, v T ∈ C 1,2 ((0, T ) × E), and solves reduced HJB equation:
∂t v + Lv + c v = 0, (t, y ) ∈ (0, T ) × E,
v (T , y ) = 1, y ∈ E,
where L := 1 A ∂yy + B ∂y and B := b − qΥ Σ−1 µ.
2
2
ii) The finite horizon martingale problems (PT )T >0 are well posed:
T
vy (y ,t)
dRt = 1 ˜
dt + σ d Zt
1−p µ + δΥ v T (y ,t)
T
(P ) T
.
dYt = B + A vyT (y ,t) dt + a d Wt
˜
v (y ,t)
72. Problem Abstract Diffusions
HJB Assumption (long run)
Assumption
73. Problem Abstract Diffusions
HJB Assumption (long run)
Assumption
iii) v > 0, v ∈ C 2 (E), and (v , λc ) solves the ergodic HJB equation:
ˆ ˆ ˆ
L v + c v = λ v, y ∈ E, for some λc ∈ R
74. Problem Abstract Diffusions
HJB Assumption (long run)
Assumption
iii) v > 0, v ∈ C 2 (E), and (v , λc ) solves the ergodic HJB equation:
ˆ ˆ ˆ
L v + c v = λ v, y ∈ E, for some λc ∈ R
ˆ
iv) The long run martingale problem (P) is well posed:
ˆ
dRt = 1 µ + δΥ vy (y ) dt + σ d Ztˆ
1−p ˆ
ˆ
(P) v (y )
ˆ
dYt = B + A vy (y ) dt + a d Wt
ˆ
ˆ
v (y )
75. Problem Abstract Diffusions
HJB Assumption (long run)
Assumption
iii) v > 0, v ∈ C 2 (E), and (v , λc ) solves the ergodic HJB equation:
ˆ ˆ ˆ
L v + c v = λ v, y ∈ E, for some λc ∈ R
ˆ
iv) The long run martingale problem (P) is well posed:
ˆ
dRt = 1 µ + δΥ vy (y ) dt + σ d Ztˆ
1−p ˆ
ˆ
(P) v (y )
ˆ
dYt = B + A vy (y ) dt + a d Wt
ˆ
ˆ
v (y )
1 y 2B(z)
v) Setting m(y ) := A(y ) exp y0 A(z) dz , for some y0 ∈ E:
y0 1 β 1 β β
α v 2 Am(y ) dy
ˆ = y0 v 2 Am(y ) dy
ˆ = ∞, α v 2 m(y ) dy ,
ˆ α
ˆ
v m(y ) dy < ∞,
76. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
77. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
78. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
ˆ
• Proposition allows to replace PT with P in abstract turnpike.
79. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
ˆ
• Proposition allows to replace PT with P in abstract turnpike.
ˆ
• Classic turnpike theorem follows from equivalence of P and P.
80. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
ˆ
• Proposition allows to replace PT with P in abstract turnpike.
ˆ
• Classic turnpike theorem follows from equivalence of P and P.
Theorem (Classic Turnpike for Diffusions)
Let previous assumptions hold. Then, for 0 = p < 1 and any , t > 0:
81. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
ˆ
• Proposition allows to replace PT with P in abstract turnpike.
ˆ
• Classic turnpike theorem follows from equivalence of P and P.
Theorem (Classic Turnpike for Diffusions)
Let previous assumptions hold. Then, for 0 = p < 1 and any , t > 0:
T
a) limT →∞ P (supu∈[0,t] ru − 1 ≥ ) = 0,
82. Problem Abstract Diffusions
Myopic Probabilities and Classic Turnpike
• Proposition
Let diffusions assumptions hold. Then, for any t ≥ 0:
dPT ˆ
dP
lim | Ft = |F .
T →∞ dP dP t
ˆ
• Proposition allows to replace PT with P in abstract turnpike.
ˆ
• Classic turnpike theorem follows from equivalence of P and P.
Theorem (Classic Turnpike for Diffusions)
Let previous assumptions hold. Then, for 0 = p < 1 and any , t > 0:
T
a) limT →∞ P (supu∈[0,t] ru − 1 ≥ ) = 0,
b) limT →∞ P ΠT , ΠT t ≥ = 0.
83. Problem Abstract Diffusions
Classic vs. Explicit
• Abstract and Classic turnpikes:
compare portfolios for U and x p /p at finite horizon T .
84. Problem Abstract Diffusions
Classic vs. Explicit
• Abstract and Classic turnpikes:
compare portfolios for U and x p /p at finite horizon T .
• Theorem says they come close for large horizons...
85. Problem Abstract Diffusions
Classic vs. Explicit
• Abstract and Classic turnpikes:
compare portfolios for U and x p /p at finite horizon T .
• Theorem says they come close for large horizons...
• ...but neither one has explicit solution. Portfolio for x p /p is:
T
vy (t, y )
1
π T (t, y ) = Σ−1 µ + δΥ
1−p v T (t, y )
86. Problem Abstract Diffusions
Classic vs. Explicit
• Abstract and Classic turnpikes:
compare portfolios for U and x p /p at finite horizon T .
• Theorem says they come close for large horizons...
• ...but neither one has explicit solution. Portfolio for x p /p is:
T
vy (t, y )
1
π T (t, y ) = Σ−1 µ + δΥ
1−p v T (t, y )
• Explicit turnpike:
compare portfolio for U with horizon T to long run portfolio:
1 ˆ
vy (y )
π (y ) =
ˆ Σ−1 µ + δΥ .
1−p ˆ
v (y )
87. Problem Abstract Diffusions
Classic vs. Explicit
• Abstract and Classic turnpikes:
compare portfolios for U and x p /p at finite horizon T .
• Theorem says they come close for large horizons...
• ...but neither one has explicit solution. Portfolio for x p /p is:
T
vy (t, y )
1
π T (t, y ) = Σ−1 µ + δΥ
1−p v T (t, y )
• Explicit turnpike:
compare portfolio for U with horizon T to long run portfolio:
1 ˆ
vy (y )
π (y ) =
ˆ Σ−1 µ + δΥ .
1−p ˆ
v (y )
• Long run portfolio solve ergodic HJB equation. ODE, not PDE.
88. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
89. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
90. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
Theorem (Explicit Turnpike)
Under the previous assumptions, for any , t > 0 and 0 = p < 1:
91. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
Theorem (Explicit Turnpike)
Under the previous assumptions, for any , t > 0 and 0 = p < 1:
rT
a) limT →∞ P (supu∈[0,t] ˆu − 1 ≥ ) = 0,
92. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
Theorem (Explicit Turnpike)
Under the previous assumptions, for any , t > 0 and 0 = p < 1:
rT
a) limT →∞ P (supu∈[0,t] ˆu − 1 ≥ ) = 0,
ˆ ˆ
b) limT →∞ P ΠT , ΠT ≥ = 0.
t
93. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
Theorem (Explicit Turnpike)
Under the previous assumptions, for any , t > 0 and 0 = p < 1:
rT
a) limT →∞ P (supu∈[0,t] ˆu − 1 ≥ ) = 0,
ˆ ˆ
b) limT →∞ P ΠT , ΠT ≥ = 0.
t
• Explicit turnpike nontrivial even for U(x) = x p /p.
94. Problem Abstract Diffusions
Explicit Turnpike
• Ratio of optimal wealth processes, and stochastic logarithms:
1,T u
Xu ˆu rT
d ˆv
rT
ˆu := , ΠT := , for u ∈ [0, T ],
ˆ
Xu 0 rT
ˆv −
ˆ
• X wealth process of long-run portfolio π .
ˆ
Theorem (Explicit Turnpike)
Under the previous assumptions, for any , t > 0 and 0 = p < 1:
rT
a) limT →∞ P (supu∈[0,t] ˆu − 1 ≥ ) = 0,
ˆ ˆ
b) limT →∞ P ΠT , ΠT ≥ = 0.
t
• Explicit turnpike nontrivial even for U(x) = x p /p.
• Finite horizon portfolios converge to long run portfolio.
95. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
96. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
• Abstract turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the myopic probabilities.
97. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
• Abstract turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the myopic probabilities.
• Classic turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the physical probability P.
98. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
• Abstract turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the myopic probabilities.
• Classic turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the physical probability P.
• Abstract implies classic if optimal wealth myopic with IDD returns.
99. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
• Abstract turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the myopic probabilities.
• Classic turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the physical probability P.
• Abstract implies classic if optimal wealth myopic with IDD returns.
• Class of diffusion models:
classic turnpike without myopic portfolios.
Intertemporal hedging components converge.
100. Problem Abstract Diffusions
Conclusion
• Portfolio turnpikes:
at long horizons, optimal portfolios approach those of CRRA class.
• Abstract turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the myopic probabilities.
• Classic turnpike:
optimal portfolios for U and x p /p at horizon T become close.
Under the physical probability P.
• Abstract implies classic if optimal wealth myopic with IDD returns.
• Class of diffusion models:
classic turnpike without myopic portfolios.
Intertemporal hedging components converge.
• Explicit turnpike:
portfolios for U at horizon T approaches long run portfolio.
Long run portfolio has explicit solutions in several models.
Links risk-sensitive control to expected utility.