Hybrid Atlas Models of Financial Equity Market

1. Hybrid Atlas Model of ﬁnancial equity market Tomoyuki Ichiba 1 Ioannis Karatzas 2,3 Adrian Banner 3 Vassilios Papathanakos 3 Robert Fernholz 3 1 University of California, Santa Barbara 2 Columbia University, New York 3 INTECH, Princeton November 2009 1

2. Outline Introduction Hybrid Atlas model Martingale Problem Stability Effective dimension Rankings Long-term growth relations Portfolio analysis Stochastic Portfolio Theory Target portfolio Universal portfolio Conclusion

3. Flow of Capital Figure: Capital Distribution Curves (Percentage) for the S&P 500 Index of 1997 (Solid Line) and 1999 (Broken Line). 3

4. Log-Log Capital Distribution Curves Figure: Capital distribution curves for 1929 (shortest curve) - 1999 (longest curve), every ten years. Source Fernholz(’02). What kind of models can describe this long-term stability? 4

5. A Model of Rankings [Hybrid Atlas model] Capital process X := {(X1 (t), . . . , Xn (t)) , 0 ≤ t < ∞} . Order Statistics: X(1) (t) ≥ · · · ≥ X(n) (t) ; 0 ≤ t < ∞ . Log capital Y := log X : Y(1) (t) ≥ · · · ≥ Y(n) (t) ; 0 ≤ t < ∞ . Dynamics of log capital: d Y(k ) (t) = (γ + γi + gk ) d t + σk d Wi (t) if Y(k ) (t) = Yi (t) ; for 1 ≤ i , k ≤ n , 0 ≤ t < ∞ , where W (·) is n−dim. B. M. company name i k th ranked company ∗ Drift (“mean”) γi gk Diffusion (“variance”) σk > 0 ∗ Banner, Fernholz & Karatzas (’05), Chatterjee & Pal (’07, ’09), Pal & Pitman (’08). 5

8. Illustration (n = 3) of interactions through rank Time X3(t)>X1(t)>X2(t) X1(t)>X3(t)>X2(t) x1 > x3 > x2 x3 > x1 > x2 x1 > x2 > x3 X1(t)>X2(t)>X3(t) x3 > x2 > x1 x2 > x1 > x3 x2 > x3 > x1 X3 X2 X1 Paths in R+ × Time . A path in different wedges of Rn . Symmetric group Σn of permutations of {1, . . . , n} . For n = 3 , {(1, 2, 3), (1, 3, 2), (2, 1, 3), (2, 3, 1), (3, 1, 2), (3, 2, 1)} . 6

9. Illustration (n = 3) of interactions through rank Time X3(t)>X1(t)>X2(t) X1(t)>X3(t)>X2(t) x1 > x3 > x2 x3 > x1 > x2 x1 > x2 > x3 X1(t)>X2(t)>X3(t) x3 > x2 > x1 x2 > x1 > x3 x2 > x3 > x1 X3 X2 X1 Paths in R+ × Time . A path in different wedges of Rn . Symmetric group Σn of permutations of {1, . . . , n} . For n = 3 , {(1, 2, 3), (1, 3, 2), (2, 1, 3), (2, 3, 1), (3, 1, 2), (3, 2, 1)} . 6

10. Vector Representation d Y (t) = G(Y (t))d t + S(Y (t))dW (t) ; 0≤t <∞ Σn : symmetric group of permutations of {1, 2, . . . , n}. For p ∈ Σn deﬁne wedges (chambers) Rp := {x ∈ Rn : xp(1) ≥ xp(2) ≥ · · · ≥ xp(n) }, Rn = ∪p∈Σn Rp , (the inner points of Rp and Rp are disjoint for p = p ∈ Σn ), (i) Qk : = {x ∈ Rn : xi is ranked k th among (x1 , . . . , xn )} = ∪{p : p(k )=i} Rp ; 1 ≤ i, k ≤ n , (j) (i) (p(k )) ∪n Qk = Rn = ∪n=1 Q j=1 and Rp = ∩n =1 Qk k . G(y ) = + γ1 + γ, . . . , gp−1 (n) + γn + γ) · 1Rp (y ) , p∈Σn (gp−1 (1) S(y ) = p∈Σn diag(σp−1 (1) , . . . , σp−1 (n) ) ·1Rp (y ) ; y ∈ Rn . sp

11. Vector Representation d Y (t) = G(Y (t))d t + S(Y (t))dW (t) ; 0≤t <∞ Σn : symmetric group of permutations of {1, 2, . . . , n}. For p ∈ Σn deﬁne wedges (chambers) Rp := {x ∈ Rn : xp(1) ≥ xp(2) ≥ · · · ≥ xp(n) }, Rn = ∪p∈Σn Rp , (the inner points of Rp and Rp are disjoint for p = p ∈ Σn ), (i) Qk : = {x ∈ Rn : xi is ranked k th among (x1 , . . . , xn )} = ∪{p : p(k )=i} Rp ; 1 ≤ i, k ≤ n , (j) (i) (p(k )) ∪n Qk = Rn = ∪n=1 Q j=1 and Rp = ∩n =1 Qk k . G(y ) = + γ1 + γ, . . . , gp−1 (n) + γn + γ) · 1Rp (y ) , p∈Σn (gp−1 (1) S(y ) = p∈Σn diag(σp−1 (1) , . . . , σp−1 (n) ) ·1Rp (y ) ; y ∈ Rn . sp

12. Martingale Problem Theorem [Krylov(’71), Stroock & Varadhan(’79) , Bass & Pardoux(’87)] Suppose that the coefficients G(·) and a(·) := SS (·) are bounded and measurable, and that a(·) is uniformly positive-definite and piece-wise constant in each wedge. For each y0 ∈ Rn there is a unique one probability measure P on (Ω = C([0, ∞), Rn ), F) such that P(Y0 = y0 ) = 1 and t f (Yt ) − f (Y0 ) − L f (Ys ) d s ; 0≤t <∞ 0 is a P local martingale for every f ∈ C 2 (Rn ) where 1 n n L f (x) = 2 i,j=1 aij (x)Dij f (x) + i=1 Gi (x)Di f (x) ; x ∈ Rn . This implies that the hybrid Atlas model is well-defined. 8

13. Model Market capitalization X follows Hybrid Atlas model: the log capitalization Yi = log Xi of company i has drift γ + gk + γi and volatility σk , (i) when company i is k th ranked, i.e., Y ∈ Qk for 1 ≤ k , i ≤ n . n d Yi (t) = γ + gk 1Q (i) (Y (t)) + γi d t k k =1 n + σk 1Q (i) (Y (t))d Wi (t) ; 0 ≤ t < ∞. k k =1 9

14. Model assumptions Market capitalization X follows Hybrid Atlas model: the log capitalization Yi = log Xi of company i has drift γ + gk + γi and volatility σk , (i) when company i is k th ranked, i.e., Y ∈ Qk for 1 ≤ k , i ≤ n . Assume σk > 0 , (gk , 1 ≤ k ≤ n) , (γi , 1 ≤ i ≤ n) and γ are real constants with stability conditions n n k gk + γi = 0 , (g +γp( ) ) < 0 , k = 1, . . . , n−1, p ∈ Σn . k =1 i=1 =1 γi = 0 , 1 ≤ i ≤ n, g1 = · · · = gn−1 = −g < 0, gn = (n − 1)g > 0 . γi = 1 − (2i)/(n + 1) , 1 ≤ i ≤ n , gk = −1, k = 1, . . . , n − 1 , gn = n − 1 . 10

17. Model Summary The log-capitalization Y = log X follows n d Yi (t) = γ + gk 1Q (i) (Y (t)) + γi d t k k =1 n + σk 1Q (i) (Y (t))d Wi (t) ; 0≤t <∞ k k =1 where σk > 0 , (gk , 1 ≤ k ≤ n) , (γi , 1 ≤ i ≤ n) and γ are real constants with stability conditions n n k gk + γi = 0 , (g +γp( ) ) < 0 , k = 1, . . . , n−1, p ∈ Σn . k =1 i=1 =1

18. Stochastic stability n The average Y (·) := i=1 Yi (·) / n of log-capitalization: n 1 n dY (t) = γ d t + n k =1 σk 1Q (i) (Y (t))d Wi (t) k i=1 d Bk (t) is a Brownian motion with variance rate n =1 σk /n2 drift γ by k 2 the Dambis-Dubins-Schwartz Theorem. Proposition Under the assumptions the deviations Y (·) := (Y1 (·) − Y (·), . . . , Yn (·) − Y (·)) from the average are stable in distribution, i.e., there is a unique invariant probability measure µ(·) such that for every bounded, measurable function f we have the Strong Law of Large Numbers T 1 lim f (Y (t)) d t = f (y )µ(d y ) , a.s. T →∞ T 0 Π where Π := {y ∈ Rn : y1 + · · · + yn = 0} .

19. Stochastic stability n The average Y (·) := i=1 Yi (·) / n of log-capitalization: n 1 n dY (t) = γ d t + n k =1 σk 1Q (i) (Y (t))d Wi (t) k i=1 d Bk (t) is a Brownian motion with variance rate n =1 σk /n2 drift γ by k 2 the Dambis-Dubins-Schwartz Theorem. Proposition Under the assumptions the deviations Y (·) := (Y1 (·) − Y (·), . . . , Yn (·) − Y (·)) from the average are stable in distribution, i.e., there is a unique invariant probability measure µ(·) such that for every bounded, measurable function f we have the Strong Law of Large Numbers T 1 lim f (Y (t)) d t = f (y )µ(d y ) , a.s. T →∞ T 0 Π where Π := {y ∈ Rn : y1 + · · · + yn = 0} .

20. Average occupation times Especially taking f (·) = 1Rp (·) or 1Q (i) (·) , we deﬁne from µ k (i) the average occupation time of X in Rp or Qk : T 1 θp := µ(Rp ) = lim 1Rp (X (t))d t T →∞ T 0 T (i) 1 θk ,i := µ(Qk ) = lim 1Q (i) (X (t))d t , 1 ≤ k, i ≤ n , T →∞ T 0 k since 1Rp (Y (·)) = 1Rp (X (·)) and 1Q (i) (X (·)) = 1Q (i) (Y (·)) . By k k deﬁnition 0 ≤ θk ,i = {p∈Σn :p(k )=i} θp ≤ 1 for 1 ≤ k , i ≤ n, n n =1 θ ,i = j=1 θk ,j = 1 for 1 ≤ k , i ≤ n . What is the invariant distribution µ ? 13

21. Attainability One-dimensional Brownian motion attains the origin inﬁnitely often. Two-dimensional Brownian motion does not attain the origin. Does the process X (·) attain the origin? t t X (t) = X (0) + b(X (s)) d s + σ(X (s))d W (s) 0 0 where b and σ are bounded measurable functions. Friedman(’74), Bass & Pardoux(’87). 14

22. Effective Dimension Let us deﬁne effective dimension ED(·) by trace(A(x)) x 2 ED(x) = ; x ∈ Rn {0} , x A(x)x where A(·) = σ(·)σ(·) . Proposition Suppose X (0) = 0 . If infx∈Rn {0} ED(x) ≥ 2 , then X (·) does not attain the origin. If supx∈Rn {0} ED(x) < 2 and if there is no drift, i.e., b(·) ≡ 0 , then X (·) attains the origin. Exterior Dirichlet Problem by Meyers and Serrin(’60). Removal of drift by Girsanov’s theorem. [trace(A(x))+x b(x)]· x 2 If there is drift, take x A(x)x . 15

23. Triple collision Now consider triple collision: Xi (t) = Xj (t) = Xk (t) for some t > 0 , 1 ≤ i < j < k ≤ n . What is the probability of triple collision? Fix i = 1 , j = 2 , k = 3 . Let us deﬁne the sum of squared distances: s2 (x) := (x1 − x2 )2 + (x2 − x3 )2 + (x3 − x1 )2 = x DD x ; x ∈ Rn , where (n × 3) matrix D is deﬁned by D := (d1 , d2 , d3 ) with d1 := (1, −1, 0, . . . , 0) , d2 := (0, 1, −1, 0, . . . , 0) , d3 := (−1, 0, 1, . . . , 0) . Z := {x ∈ Rn : s(x) = 0 }. 16

24. Deﬁne the local effective dimension: trace(D A(x)D) x DD x R(x) := ; x ∈ Rn Z . x DD A(x)DD x Proposition Suppose s(X (0)) = 0 . If infx∈Rn Z R(x) ≥ 2 , then P(X1 (t) = X2 (t) = X3 (t) for some t ≥ 0) = 0 . If supx∈Rn Z R(x) < 2 and if there is no drift, i.e., b(·) ≡ 0 , then P(X1 (t) = X2 (t) = X3 (t) for some t ≥ 0) = 1 . R(·) ≡ 2 for n−dim. BM, i.e., A(·) ≡ I . [trace(D A(x)D)+x DD b(x)]·x DD x If there is drift, take x DD A(x)DD x . Idea of Proof: a comparison with Bessel process with dimension two. 17

27. Rankings Recall Y(1) (·) ≥ Y(2) (·) ≥ · · · ≥ Y(n) (·) . Let us denote by Λk ,j (t) the local time accumulated at the origin by the nonnegative semimartingale Y(k ) (·) − Y(j) (·) up to time t for 1 ≤ k < j ≤ n. Theorem[Banner & Ghomrasni (’07)] For a general class of semimartingale Y (·) , the rankings satisfy n d Y(k ) (t) = 1Q (i) (Y (t))d Yi (t) k i=1 n k −1 −1 k, ,k + (Nk (t)) dΛ (t) − dΛ (t) =k +1 =1 where Nk (t) is the cardinality |{i : Yi (t) = Y(k ) (t)}| . 18

28. Rankings Recall Y(1) (·) ≥ Y(2) (·) ≥ · · · ≥ Y(n) (·) . Let us denote by Λk ,j (t) the local time accumulated at the origin by the nonnegative semimartingale Y(k ) (·) − Y(j) (·) up to time t for 1 ≤ k < j ≤ n. Lemma Under the non-degeneracy condition σk > 0 for k = 1, . . . , n , n dY(k ) (t) = γ + gk + γi 1Q (i) (Y (t)) d t + σk d Bk (t) k i=1 1 + d Λk ,k +1 (t) − d Λk −1,k (t) . 2 for k = 1, . . . , n , 0 ≤ t ≤ T . Idea of Proof: a comparison with a Bessel process with dimension one to show Λk , (·) ≡ 0 , |k − | ≥ 2. 19

29. Rankings Recall Y(1) (·) ≥ Y(2) (·) ≥ · · · ≥ Y(n) (·) . Let us denote by Λk ,j (t) the local time accumulated at the origin by the nonnegative semimartingale Y(k ) (·) − Y(j) (·) up to time t for 1 ≤ k < j ≤ n. Lemma Under the non-degeneracy condition σk > 0 for k = 1, . . . , n , n dY(k ) (t) = γ + gk + γi 1Q (i) (Y (t)) d t + σk d Bk (t) k i=1 1 + d Λk ,k +1 (t) − d Λk −1,k (t) . 2 for k = 1, . . . , n , 0 ≤ t ≤ T . Idea of Proof: a comparison with a Bessel process with dimension one to show Λk , (·) ≡ 0 , |k − | ≥ 2. 19

30. Long-term growth relations Proposition Under the assumptions we obtain the following long-term growth relations: Yi (T ) log Xi (T ) lim = lim =γ T →∞ T T →∞ T n log i=1 Xi (T ) = lim a.s. T →∞ T Thus the model is coherent: 1 lim log µi (T ) = 0 a.s.; i = 1, . . . , n T →∞ T where µi (·) = Xi (·)/(X1 (·) + · · · + Xn (·)) . Moreover, n gk θk ,i + γi = 0 ; i = 1, . . . , n . k =1 20

31. n gk θk ,i + γi = 0 ; i = 1, . . . , n . k =1 The log-capitalization Y grows with rate γ and follows n d Yi (t) = γ + gk 1Q (i) (Y (t)) + γi d t k k =1 n + σk 1Q (i) (Y (t))d Wi (t) ; 0≤t <∞ k k =1 for i = 1, . . . , n . The ranking (Y(1) (·), . . . , Y(n) (·)) follows n dY(k ) (t) = γ + gk + γi 1Q (i) (Y (t)) d t + σk d Bk (t) k i=1 1 + d Λk ,k +1 (t) − d Λk −1,k (t) . 2 for k = 1, . . . , n , 0 ≤ t < ∞ .

32. Semimartingale reﬂected Brownian motions The adjacent differences (gaps) Ξ(·) := (Ξ1 (·), . . . , Ξn (·)) where Ξk (·) := Y(k ) (·) − Y(k +1) (·) for k = 1, . . . , n − 1 can be seen as a semimartingale reﬂected Brownian motion (SRBM): Ξ(t) = Ξ(0) + ζ(t) + (In−1 − Q)Λ(t) where ζ(·) := (ζ1 (·), . . . , ζn (·)) , Λ(·) := (Λ1,2 (·), . . . , Λn−1,n (·)) , n · n · ζk (·) := 1Q (i) (Y (s)) d Y (s) − 1Q (i) (Y (s)) d Y (s) 0 k 0 k +1 i=1 i=1 for k = 1, . . . , n − 1 , and Q is an (n − 1) × (n − 1) matrix with elements

33.   0 1/2  1/2 0 1/2     .. ..  Q :=   1/2 . . .   ..   . 0 1/2  1/2 0 Thus the gaps Ξk := Y(k ) (·) − Y(k +1) (·) follow Ξ(t) = Ξ(0) + ζ(t) + (In − Q)Λ(t) semimartingale reﬂection part In order to study the invariant measure µ , we apply the theory of semimartingale reﬂected Brownian motions developed by M. Harrison, M. Reiman, R. Williams and others. In addition to the model assumptions, we assume linearly growing variances: 2 2 2 2 2 2 σ2 − σ1 = σ3 − σ2 = · · · = σn − σn−1 . 23

34. Invariant distribution of gaps and index Let us deﬁne the indicator map Rn x → px ∈ Σn such that xpx (1) ≥ xpx (2) ≥ · · · ≥ xp(n) , and the index process Pt := pY (t) . Proposition Under the stability and the linearly growing variance conditions the invariant distribution ν(·) of (Ξ(·), P· ) is n−1 −1 ν(A × B) = λ−1 q,k exp(− λp , z )d z q∈Σn k =1 p∈Σn A for every measurable set A × B where λp := (λp,1 , . . . , λp,n−1 ) is the vector of components k −4( =1 g + γp( ) ) λp,k := 2 + σ2 > 0; p ∈ Σn , 1 ≤ k ≤ n − 1 . σk k +1 Proof: an extension from M. Harrison and R. Williams (’87).

35. Average occupation time Corollary The average occupation times are n−1 n−1 −1 θp = λ−1 q,k λ−1 p,j and θk ,i = θp q∈Σn k =1 j=1 {p∈Σn :p(k )=i} for p ∈ Σn and 1 ≤ k , i ≤ n . (1,10) (10,10) 0. 15 2 2 If all γi = 0 and σ1 = · · · σn , then 1 θk ,i = n for 1 ≤ k , i ≤ n . 2 Heat map of θk ,i when n = 10 , σk = 1 + k , 5 0.0 (k,i) gk = −1 for k = 1, . . . , 9 , g10 = 9 , and γi = 1 − (2i)/(n + 1) for i = 1, . . . , n . 0. 1 0.15 0.05 0.2 (1,1) 0.1 (10,1) 25

36. Market weights come from Pareto type Corollary The joint invariant distribution of market shares µ(i) (·) := X(i) (·)/(X1 (·) + · · · + Xn (·)) ; i = 1, . . . , n has the density ℘(m1 , . . . , mn−1 ) λp,1 · · · λp,n−1 = θp λ +1 λp,2 −λp,1 +1 λp,n−1 −λp,n−2 +1 −λp,n−1 +1 , p∈Πn m1 p,1 · m2 · · · mn−1 mn 0 < mn ≤ mn−1 ≤ . . . ≤ m1 < 1, mn = 1 − m1 − · · · − mn−1 . This is a distribution of ratios of Pareto type disribution. 26

37. Expected capital distribution curves E (Ξ ) log µ −log µ ν From the expected slopes Eν [ log(k)+1)−log−1) ] = − log(1+kk−1 ) we (k (k k obtain expected capital distribution curves. e−1 e−1 iv e−5 e−5 v Weight Weight iii e−10 e−10 ii i Rank Rank 1 5 10 50 100 500 1000 5000 1 5 10 50 100 500 1000 5000 2 n = 5000 , gn = c∗ (2n − 1) , gk = 0 , 1 ≤ k ≤ n − 1 , γ1 = −c∗ , γi = −2c∗ , 2 ≤ i ≤ n , σk = 0.075 + 6k × 10−5 , 1 ≤ k ≤ n. (i) c∗ = 0.02 , (ii) c∗ = 0.03 , (iii) c∗ = 0.04 . (iv) c∗ = 0.02 , g1 = −0.016 , gk = 0 , 2 ≤ k ≤ n − 1 , gn = (0.02)(2n − 1) + 0.016 , (v) g1 = · · · = g50 = −0.016 , gk = 0 , 51 ≤ k ≤ n − 1 , gn = (0.02)(2n − 1) + 0.8 .

38. Empirical data 0.7 0.6 Variance Rate 0.5 0.4 0.3 0.2 Historical capital distribution 0 1000 2000 Rank 3000 4000 5000 curves. Data 1929-1999. Growing variances. 1990- 1999. Source: Fernholz (’02) 28

39. Capital Stocks and Portfolio Rules Market X = ((X1 (t), . . . , Xn (t)), t ≥ 0) of n companies T T n Xi (T ) log = Gi (t)dt + Si,ν (t)dWν (t) , Xi (0) 0 0 ν=1 with initial capital Xi (0) = xi > 0, i = 1, . . . , n , 0 ≤ T < ∞ . · Deﬁne aij (·) = n Siν (·)Sjν (·) and Aij (·) = 0 aij (t)d t . ν=1 Long only Portfolio rule π and its wealth V π . Choose π ∈ ∆+ := {x ∈ Rn : xi = 1, xi ≥ 0 } n invest πi V π of money to company i for i = 1, . . . , n, i.e., πi V π /Xi share of company i : n π πi (t)V π (t) dV (t) = dXi (t) , 0 ≤ t < ∞, Xi (t) i=1 V π (0) = w . 29

42. Portfolios and Relative Arbitrage Market portfolio: Take π(t) = m(t) = (m1 (t), · · · , mn (t)) ∈ ∆n where + Xi (t) mi (t) = , i = 1, . . . , n , 0 ≤ t < ∞. X1 + · · · Xn (t) Diversity weighted portfolio: Given p ∈ [0, 1] , take (m (t))p πi (t) = Pn i(m (t))p for i = 1, . . . , n , 0 ≤ t < ∞. j=1 j Functionally generated portfolio (Fernholz (’02) & Karatzas (’08)). A portfolio π represents an arbitrage opportunity relative to another portfolio ρ on [0, T ] , if P(V π (T ) ≥ V ρ (T )) = 1, P(V π (T ) > V ρ (T )) > 0. Can we ﬁnd an arbitrage opportunity π relative to m ? 30

43. Portfolios and Relative Arbitrage Market portfolio: Take π(t) = m(t) = (m1 (t), · · · , mn (t)) ∈ ∆n where + Xi (t) mi (t) = , i = 1, . . . , n , 0 ≤ t < ∞. X1 + · · · Xn (t) Diversity weighted portfolio: Given p ∈ [0, 1] , take (m (t))p πi (t) = Pn i(m (t))p for i = 1, . . . , n , 0 ≤ t < ∞. j=1 j Functionally generated portfolio (Fernholz (’02) & Karatzas (’08)). A portfolio π represents an arbitrage opportunity relative to another portfolio ρ on [0, T ] , if P(V π (T ) ≥ V ρ (T )) = 1, P(V π (T ) > V ρ (T )) > 0. Can we ﬁnd an arbitrage opportunity π relative to m ? 30

44. Constant-portfolio For a constant-proportion π(·) ≡ π , n n Aii (t) Xi (t) 1 V π (t) = w ·exp πi · +log − πi Aij (t)πj 2 Xi (0) 2 i=1 i,j=1 for 0 ≤ t < ∞ . · n Here Aij (·) = 0 aij (t)d t and aij (·) = ν=1 Siν (·)Sjν (·) , d Y (t) = G(Y (t))d t + S(Y (t))dW (t) ; 0 ≤ t < ∞, G(y ) = + γ1 + γ, . . . , gp−1 (n) + γn + γ) · 1Rp (y ) , p∈Σn (gp−1 (1) S(y ) = p∈Σn diag(σp−1 (1) , . . . , σp−1 (n) ) ·1Rp (y ) ; y ∈ Rn . sp 31

45. Target Portfolio(Cover(’91) & Jamshidian(’92)) n n Aii (t) Xi (·) 1 V π (·) = w · exp πi + log − πi Aij (·)πj 2 Xi (0) 2 i=1 i,j=1 Target Portfolio Π∗ (t) maximizes the wealth V π (t) for t ≥ 0: V∗ (t) := max V π (t), n Π∗ (t) := arg max V π (t) , n π∈∆+ π∈∆+ where by Lagrange method we obtain n n 1 −1 1 Xj (t) Π∗ (t) i = 2Aii (t) 2−n−2 log Ajj (t) Ajj (t) Xj (0) j=1 j=1 1 1 Xi (t) + + log ; 0 ≤ t < ∞. 2 Aii (t) Xi (0) 32

46. Asymptotic Target Portfolio Under the hybrid Atlas model with the assumptions n n 1 1 v (π) := lim log V π (T ) = γ + πi a ∞ − ii π i a ∞ πj ii T →∞ T 2 i=1 i=1 ∞ γπ where (a∞ )1≤i≤n is the (i,i) element of ij T 1 a∞ := lim (aij (t))1≤i,j≤n d t = θp sp sp . T →∞ T 0 p∈Σn ∞ Asymptotic target portfolio maximizes the excess growth γπ : n n π := arg max ¯ n πi a ∞ − ii πi a ∞ πj . ii π∈∆+ i=1 i=1 We obtain n 1 n−2 1 −1 πi = ¯ 1− ∞ = lim Π∗ (t) ; i i = 1, . . . , n . 2 aii a∞ jj t→∞ j=1 33

47. Asymptotic Target Portfolio Under the hybrid Atlas model with the assumptions n n 1 1 v (π) := lim log V π (T ) = γ + πi a ∞ − ii π i a ∞ πj ii T →∞ T 2 i=1 i=1 ∞ γπ where (a∞ )1≤i≤n is the (i,i) element of ij T 1 a∞ := lim (aij (t))1≤i,j≤n d t = θp sp sp . T →∞ T 0 p∈Σn ∞ Asymptotic target portfolio maximizes the excess growth γπ : n n π := arg max ¯ n πi a ∞ − ii πi a ∞ πj . ii π∈∆+ i=1 i=1 We obtain n 1 n−2 1 −1 πi = ¯ 1− ∞ = lim Π∗ (t) ; i i = 1, . . . , n . 2 aii a∞ jj t→∞ j=1 33

48. Universal Portfolio(Cover(’91) & Jamshidian(’92)) Universal portfolio is deﬁned as ∆n πi V π (·)dπ ∆n V π (·)dπ + Π + Πi (·) := , 1 ≤ i ≤ n, V (·) = . b ∆n V π (·)dπ ∆n dπ + + Proposition Under the hybrid Atlas model with the model assumptions, 1 V Π (T ) 1 V Π (T ) b b lim log π = lim log =0 P − a.s. T →∞ T V ¯ (T ) T →∞ T V∗ (T ) 34

49. Universal Portfolio(Cover(’91) & Jamshidian(’92)) Universal portfolio is deﬁned as ∆n πi V π (·)dπ ∆n V π (·)dπ + Π + Πi (·) := , 1 ≤ i ≤ n, V (·) = . b ∆n V π (·)dπ ∆n dπ + + Proposition Under the hybrid Atlas model with the model assumptions, 1 V Π (T ) 1 V Π (T ) b b lim log π = lim log =0 P − a.s. T →∞ T V ¯ (T ) T →∞ T V∗ (T ) 34

50. Conclusion Ergodic properties of Hybrid Atlas model Diversity weighted portfolio, Target portfolio, Universal portfolio. Further topics: short term arbitrage, generalized portfolio generating function, large market (n → ∞), numèraire portfolio, data implementation. References: 1. arXiv: 0909.0065 2. arXiv: 0810.2149 (to appear in Annals of Applied Probability) 35

Hybrid Atlas Models of Financial Equity Market

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Hybrid Atlas Models of Financial Equity Market (20)

Hybrid Atlas Models of Financial Equity Market