You are on page 1of 92

# Thesis for the Degree of Doctor of Philosophy

Portfolio Optimization and
Statistics in Stochastic Volatility
Markets
Carl Lindberg

Department of Mathematical Sciences
Division of Mathematical Statistics
Chalmers University of Technology and Göteborg University
Göteborg, Sweden 2005

Portfolio optimization and statistics in
stochastic volatility markets
Carl Lindberg
Department of Mathematical Sciences,
Chalmers University of Technology
and Göteborg University, Sweden
Large …nancial portfolios often contain hundreds of stocks. The aim of this
thesis is to …nd explicit optimal trading strategies that can be applied to portfolios of that size for di¤erent n-stock extensions of the model by Barndor¤-Nielsen
and Shephard [3]. A main ambition is that the number of parameters in our
models do not grow too fast as the number of stocks n grows. This is necessary
to obtain stable parameter estimates when we …t the models to data, and n
is relatively large. Stability over the parameter estimates is needed to obtain
accurate estimates of the optimal strategies. Statistical methods for …tting the
models to data are also given.
The thesis consists of three papers. Paper I presents an n-stock extension to
the model in [3] where the dependence between di¤erent stocks lies strictly in
the volatility. The model is primarily intended for stocks that are dependent,
but not too dependent, such as stocks from di¤erent branches of industry. We
develop optimal portfolio theory for the model, and indicate how to do the statistical analysis. In Paper II we extend the model in Paper I further, to model
stronger dependence. This is done by assuming that the di¤usion components
of the stocks contain one Brownian motion that is unique for each stock, and
a few Brownian motions that all stocks share. We then develop portfolio optimization theory for this extended model. Paper III presents statistical methods
to estimate the model in [3] from data. The model in Paper II is also considered.
It is shown that we can divide the centered returns by a constant times the daily
number of trades to get normalized returns that are i:i:d: and N (0; 1) : It is a
key feature of the Barndor¤-Nielsen and Shephard model that the centered returns divided by the volatility are also i:i:d: and N (0; 1) : This suggests that we
identify the daily number of trades with the volatility, and model the number of
trades within the framework of Barndor¤-Nielsen and Shephard. Our approach
is easier to implement than the quadratic variation method, requires much less
data, and gives stable parameter estimates. A statistical analysis is done which
shows that the model …ts the data well.
Key words: Stochastic control, portfolio optimization, veri…cation theorem, Feynman-Kac formula, stochastic volatility, non-Gaussian Ornstein-Uhlenbeck
iii

.

submitted. v . Paper II: Portfolio optimization and a factor model in a stochastic volatility market. submitted. Paper III: The estimation of a stochastic volatility model based on the number of trades.This thesis consists of the following papers: Paper I: News-generated dependence and optimal portfolios for n stocks in a market of Barndor¤-Nielsen and Shephard type. to appear in Mathematical Finance.

.

Contents 1 Introduction 1 2 Portfolio optimization 2 3 "Stylized" features of stock returns 4 4 Stochastic volatility models in …nance 4 5 Summary of papers 5.2 Paper II 5.3 Paper III 7 7 8 8 Paper I Paper II Paper III vii .1 Paper I 5.

.

constructive criticism. They have both. and encouragement. My beloved sons Adam and Bo. These years at the Department of Mathematical Sciences would not have been as fun and productive without my good friend and o¢ ce roommate Erik Brodin. for their unconditional love and support. ix . I have bene…tted a lot from being able to discuss matters of research and life with him. for constantly reminding me of what is important in life. I love you. and mistress Cia. who make my life great. Especially My wonderful parents Ewa and Lars-Håkan. I dedicate this thesis to my friends and family. been of great help to me with their advice. and sisters Kristina and Anna. in di¤erent ways. My best friend. …ancée.Acknowledgements First of all I would like to thank my supervisors Holger Rootzén and Fred Espen Benth. I am grateful to my friends and colleagues at the Department of Mathematical Sciences for providing a pleasant working environment.

so that it can model stronger dependence between di¤erent stocks. and a few Brownian motions that all stocks share.Paper I presents an n-stock extension to the Barndor¤-Nielsen and Shephard model where the dependence between di¤erent stocks lies in that they partly share the Ornstein-Uhlenbeck processes of the volatility. 2 Portfolio optimization The …rst papers on continuous time portfolio optimization are due to Merton ([20] and [21]). In Section 2 of this summary we recapitulate some results from classical continuous time portfolio optimization. S2 takes the form S1 (t) = S1 (0) exp ( 1 S2 (t) = S2 (0) exp ( 2 1 2 2 11 1 2 2 21 1 2 2 12 )t 1 2 2 22 )t + 11 W1 (t) + 12 W2 (t) . in Section 5 we present the three papers that constitute this thesis. This is done by introducing a factor structure in the di¤usion components. Further. We present in this section a version of Merton’s problem in its classical setting. are independent Brownian motions. The model is mainly intended for stocks that are dependent. Finally. and model the number of trades within the framework of Barndor¤-Nielsen and Shephard. which for two stocks S1 . Our approach gives more stable parameter estimates than if we analyzed only the marginal distribution of the returns directly with the standard maximum likelihood approach. Wi . It is shown that we can divide the centered returns by a constant times the daily number of trades to get normalized returns that are i:i:d: and N (0. such as stocks that are not in the same branch of industry. i = 1. and indicate how to do the statistical analysis for the model. we indicate why the classical models lack all these characteristics. We develop portfolio optimization portfolio theory. We also consider the model from Paper II. + 21 W1 (t) + 22 W2 (t) : (2. The matrix gives the dependence between the two stocks. and is a volatility matrix. 2 . i = 1. In Paper II we extend the model in Paper I further. Merton modelled the stock prices as multi-variate geometric Brownian motions. The idea of a factor structure is that the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. We then develop optimal portfolio theory for this extended model. 1) : It is an important theoretical feature of the stochastic volatility framework of Barndor¤-Nielsen and Shephard that the centered returns divided by the volatility are also i:i:d: and N (0. but not too dependent. Further. are constants. A statistical analysis is done which shows that the model …ts the data well. 2.1) Here i . Section 4 introduces the stochastic volatility model of [3]. Paper III presents statistical methods to estimate the model in [3] from data. and requires much less data. it is easier to implement than the quadratic variation method. 1) : This suggests that we identify the daily number of trades with the volatility. Section 3 discusses "stylized" facts of stock price data. 2. and the ideas from stochastic control used to derive them.

One of the most widely used optimal value functions is V (t. one assumes that sup lim t#0 E [V (t. for 0 < < 1: The trading strategies are recipes for how we are going to allocate our wealth between di¤erent assets. The last and …nal step is then to actually …nd the solution to the HJB equation. where T is a future date in time. Hence. So far. which is independent of the rest of the model. W (T ) is our wealth at time T . where 1 is a vector of ones. The concavity means that the more money an investor gets. This formulation of the portfolio optimization problem means that we seek the trading strategies such that we obtain the maximum expected utility from wealth on a future day T: We outline now the stochastic control approach to …nding this optimal value function V: First. di¤erent portfolio 3 . w) = sup E [U (W (T )) jt. we have only found an equation whose solution we guess is the optimal value function. However. The condition that the utility function should be increasing implies that the investor always prefers more to less. [23]. The utility function is a measure of how much we want to risk to obtain more wealth. and [26]. It can be evaluated using Itô’s formula which gives an equation called the Hamilton-JacobiBellman (HJB) equation.In portfolio optimization one has to choose a value function to optimize. and is a trading strategy. W (t))] t V (0. [6]. W (t) = w ] . Merton suggested the utility function U (w) = 1 w . w) = 0: This "derivative" serves as a necessary condition for optimality. In the papers [13]. In [5]. and the optimal trading strategies turn out to be =( 0 ) 1 ( r1) ¯ 1 1 . since the HJB-equation is nonlinear. The next step is to prove a veri…cation theorem. In [7]. the less interested she will be in obtaining a little more. ¯ Portfolio optimization with more general stock price models than the geometric Brownian motion has been treated in a number of recent articles. it can be done in the setting of this section. This theorem says that a solution to the HJB-equation is in fact equal to the optimal value function. U ( ) is our utility function. [14]. The paper [9] model the volatility as a continuous-time Markov chain with …nite state-space. we have veri…ed that our guess was correct. It is typically assumed to be concave and increasing. the stochastic volatility depends on a Brownian motion which is correlated to the di¤usion process of the risky asset. a one-stock portfolio problem in the model in [3] is solved. This is typically quite hard. and [12].

among other things. which is used in the classical portfolio optimization problem above. R (k ) R ((k 1) ) . where is one day. 4 Stochastic volatility models in …nance There has been published some di¤erent models that include stochastic volatility in stock price dynamics. they are peaked around zero. [11].3 is positive.1). The most obvious example of a volatility cluster is the latter part of 2002.problems are treated when the stocks are driven by general Lévy processes. [15] derive explicit solutions for log-optimal portfolios in complete markets in terms of the semimartingale characteristics of the price process. The stock price model in Equation (2. the autocorrelation function of the absolute returns in Figure 3. It shows a clear non-normality. skew. and [18] show that there exists a unique solution to the optimal investment problem for any arbitrage-free model if and only if the utility function has asymptotic elasticity strictly less than one. the volatility of the returns is evidently not constant. It is widely agreed that the returns of …nancial data have. In Figure 3. The autocorrelation function for absolute returns is clearly positive even for long lags. The volatility of the returns changes stochastically over time. see for example [3].1 seems consistent with the …rst listed feature. skew. A common approach to improve the geometric Brownian motion as a stock price model is to assume that the volatility is stochastic. does not capture any of the features listed above: The returns in this model are i:i:d: and normally distributed. Instead. and more heavy-tailed than the normal distribution. and the volatility is constant. and appears to be clustered. the following characteristic features: The returns are not normally distributed. and [10] look at portfolio optimization in a market with Markov-modulated drift process. The empirical density function of the returns in Figure 3. 3 "Stylized" features of stock returns The standard approach to analyze …nancial data is to look at the increments of the returns process R (t) := log (S (t) =S (0)) for the stock S: We assume that we are observing returns R ( ) . and so appears compatible with the last condition. and [17]. and k+1 is the number of consecutive trading days in our period of observation. Further. We now give a brief indication that these empirical facts hold.2. and have heavier tails than the normal distribution. there seems to be a random succession of periods with high return variance and periods with low return variance. R (2 ) R ( ) . That is. This thesis 4 . and appears to be both peaked around zero. :::. [16]. Further. Hence these data give no reason to doubt the second "stylized" fact.

02 0.05 0 -0.04 -0.25 1999-08-16 2002-02-06 2004-08-16 Figure 3.02 0 0.06 -0.1: Stars indicate the empirical density function for daily returns for Volvo B during 1999-08-16 to 2004-08-16.1 -0.1 0.05 -0. The solid line is the estimated normal density function to the same data set.06 0.08 Figure 3.2: Returns for Ericsson from 1999-08-16 to 2004-08-16.08 -0.04 0.2 -0. 5 .2 0. 0.25 20 15 10 5 0 -0.15 0.15 -0.

and [25]. which can be interpreted as the release of unexpected information.1 0 0 10 20 Figure 3. A further advantage of the Barndor¤-Nielsen and Shephard model is that it is analytically tractable.8 0.9 0. that is." This models volatility clustering. for example the normal inverse Gaussian (N IG) distribution. In this model the volatility 2 ( ) of a stock is de…ned as a linear combination of non-Gaussian Ornstein-Uhlenbeck processes of the form Rt (4. The GH family is quite general. The use of subordinators allows for sudden increases in the volatility 2 ( ). builds upon extensions of the model in [3]. t 0: Here yj := Yj (0) .1) Yj (t) = yj e j t + 0 e j (t u) dZj ( j u).1 0. 2 0 0 for some Brownian motion W: It can be shown that the returns in this model can get marginal distributions from the Generalized Hyperbolic (GH) distribution.5 0. Further. In essence. t 0: The process Zj is a subordinator. the model captures all "stylized" facts of …nancial data listed above.6 0.3: Empirical autocorrelation function for absolute returns for SKF from 1999-08-16 to 2004-08-16. a Lévy process with positive increments. the e¤ect of large jumps in the volatility 2 ( ) "lingers. and yj has the stationary marginal distribution of the process and is independent of Zj (t) Zj (0) .2 0. Option pricing is treated 6 .3 0. [3].4 0. since the Ornstein-Uhlenbeck processes Yj decrease exponentially. and includes many distributions that have been used to model …nancial return data.7 0. see [1]. The stock price process S then takes the form Z t Z t 2 1 S (t) = S (0) exp + (u) du + (u) dW (u) .

we then compute the solution to this equation via a Feynman-Kac representation. It is necessary to have many more observations than parameters to obtain stable parameter estimates. The model is primarily intended for assets which are dependent. All results are derived under exponential integrability assumptions on the Lévy measures 7 . We want to develop explicit optimal trading strategies that can be applied to portfolios of this size for di¤erent n-stock extensions of the model in [3].1 Paper I In this paper we consider Merton’s portfolio optimization problem in a Barndor¤Nielsen and Shephard market. such as stocks from di¤erent branches of industry. The model retains all the features of the univariate model in [3]. large …nancial portfolios often contain hundreds of stocks. For power utility. The case with only one stock was solved in [7]. We want to capture the essence of the dependence between di¤erent stocks. and wants to maximize her expected utility from wealth at the terminal date T . An investor is allowed to trade in n stocks and a risk-free bond. and obtain explicit optimal allocation strategies. but still be able to estimate the model accurately from data. A main advantage with the model is that the optimal strategies are functions of only 2n model parameters and the volatility of each stock. 5. for example. This requires careful modeling of the stock price and volatility dynamics. we identify the optimal expected utility from terminal wealth as the solution of a second-order integro-di¤erential equation.in [22]. However. The dependence between stocks is assumed to be that they partly share the Ornstein-Uhlenbeck processes of the volatility. The reason is that the number of parameters in such a model would grow very fast as the number of stocks grows. but not too dependent. We refer to these as news processes. in [4] and [24]. and n Brownian motions in the di¤usion components of all n stocks. The investor is allowed to have restrictions on the fractions of wealth held in each stock. but also borrowing and short-selling constraints on the entire portfolio. By use of a veri…cation theorem. and inference techniques are developed. This is a desirable feature which allows us to obtain good estimates of the optimal strategies even when n is large. We show that this dependence generates covariance between the returns of di¤erent stocks. portfolio optimization for one stock and a bond in [7]. we can not use the standard approach: An explicit stochastic volatility matrix. This gives the interpretation that dependence between stocks lies solely in their reactions to the same news. 5 Summary of papers Most published papers on portfolio optimization with more general stock price models than the geometric Brownian motion consider only the case of one stock and a bond. and give statistical methods for both the …tting and veri…cation of the model to data. Therefore. The stochastic optimization problem is solved via dynamic programming and the associated HJB integro-di¤erential equation.

5. This makes it in theory possible to recover the volatility process from observed stock prices. we try to deal with this weakness. We allow for the investor to have restrictions on the fractions of wealth held in each stock. and obtain explicit optimal allocation strategies. All results are derived under exponential integrability assumptions on the Lévy measures of the subordinators. Instead.of the subordinators. and a few Brownian motions that all stocks share. This might not …t data. Accurate parameter estimates is needed to obtain good estimates of the optimal strategies. In addition. the marginal distributions have to be very skew. it is hard to implement in a statistically sound way due to peculiarities in intraday data. The stochastic optimization problem is solved via dynamic programming and the associated HJB integro-di¤erential equation. In the …rst part of Paper II. Therefore.3 Paper III A drawback with the Barndor¤-Nielsen and Shephard model has been the dif…culty to estimate the parameters of the model from data. 5. A factor model has fewer parameters than a standard model. The reason is that the number of factors can be chosen a lot smaller than the number of stocks. and n is relatively large. We then compute the solution to this equation via a Feynman-Kac representation for power utility.2 Paper II The model in Paper I has a weak point: To obtain strong correlations between the returns of di¤erent stocks. We introduce in Paper II a more general n-stock extension of the model in Paper I. Hence. as well as borrowing and short-selling restrictions on the entire portfolio. in reality the model does not hold on the microscale. The idea of a factor structure is that the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. we de…ne the stochastic volatility matrix implicitly by a factor structure. Perhaps the most intuitive approach to do this is to analyze the quadratic variation of the stock price process. This is necessary to obtain accurate parameter estimates when we …t the model to data. see [4]. and even if is only regarded as an approximation this approach still requires very much data. the dependence between stocks lies both in the stochastic volatility. In the second part we consider an investor who wants to maximize her utility from terminal wealth by investing in n stocks and a bond. we do not use the standard approach with n Brownian motions in the di¤usion components of all n stocks. and in the Brownian motions. We use a veri…cation theorem to identify the optimal expected utility from terminal wealth as the solution of a second-order integro-di¤erential equation. For example. However. It is a primary focus that the number of parameters does not grow too fast as the number of stocks grows. 8 . The latter are called factors. We show that this model can obtain strong correlations between the returns of the stocks without a¤ecting their marginal distributions.

E. it is easier to implement than the quadratic variation method. E. E. Finance and Stochastics 2. (2001): Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in …nancial economics. In addition. Our approach gives more stable parameter estimates than if we analyzed only the marginal distribution of the returns directly with the standard maximum likelihood method.the stock market is closed at night. None of these features are present in the mathematical model. in Lévy Processes . [4] Barndor¤-Nielsen. (2002): Econometric analysis of realized volatility and its use in estimating stochastic volatility models.. It gives also an economical interpretation of the discretely observed linear combination of non-Gaussian Ornstein-Uhlenbeck processes that de…ne the stochastic volatility.. T. Mikosch and S. 1) : This implies that we identify the daily number of trades with the volatility. O. t for " 2 N (0.. A statistical analysis is performed on data from the OMX Stockholmsbörsen. We argue that it is inappropriate to estimate the GH-distribution directly from …nancial return data. To overcome this problem. Shephard. we impose some restrictions on the volatility in order to be able to estimate the model from data. Boston: Birkhäuser. (2001a): Modelling by Lévy processes for …nancial econometrics. References [1] Barndor¤-Nielsen. Journal of the Royal Statistical Society: Series B 64. and there is more intense trading on certain hours of the day. The results indicate a good model …t. N. [2] Barndor¤-Nielsen. Shephard. O. 1) : It is an important feature of the stochastic volatility framework in [3] that the centered returns divided by the volatility are also i:i:d: and N (0. O. Further. Journal of the Royal Statistical Society: Series B 63. N. The models are discretized under the assumption that the Wiener integrals in the Barndor¤-Nielsen and Shephard model Z t (s) dB (s) (t) ". The reason is that the GH-distribution is "almost" overparameterized. N. our approach implies that we can view the continuous time volatility as the intensity with which trades arrive. In this paper we develop statistical methods for estimating the models in [3] and Paper II from data. 1) : In addition. 41-68. and requires much less data.E. Barndor¤-Nielsen. O. 9 . Resnick). we verify that we can divide the centered returns by a constant times the number of trades in a trading day to get a sample that is i:i:d: and N (0. (1998): Processes of normal inverse Gaussian type. 167-241 (with discussion). and model the number of trades within the model in [3]. [3] Barndor¤-Nielsen. E. Shephard.Theory and Applications (eds O. 253-280.

(2001b): Optimal portfolio management rules in a non-Gaussian market with durability and intertemporal substitution. H. A. 275-303. [8] Black. [6] Benth.. Madan. [17] Hull. (2004): Portfolio optimization with Markovmodulated stock prices and interest rates. 215-244. 10 . Karlsen. Rieder. [16] Heston. 904-950. R. U. 245-262. White. (1999): The asymptotic elasticity of utility functions and optimal investment in incomplete markets. Annals of Applied Probability 13. H.. Reikvam K. K. D. (1973): The pricing of options and corporate liabilities.. Finance and Stochastics 5.[5] Benth. W. [9] Bäuerle. Klüppelberg. Finance and Stochastics 7.. F.. U. G. (2001a): Optimal portfolio selection with consumption and non-Linear integro-di¤erential equations with gradient constraint: A viscosity solution approach. (1993): A closed-form solution for options with stochastic volatility with applications to bond and currency options. W. Hernández-Hernández. 17-44. S. F. P.. Review of Financial Studies 6. Finance and Stochastics 8. Yor. K. 774-799.. to appear in Journal of Applied Probability. F. M.-P. Geman. 327-343... No. S.. [11] Carr. Journal of Political Economy 81. H. Reikvam K. 345-382. C. Annals of Applied Probability 9. [10] Bäuerle. [13] Fleming. D. [14] Fouque. Mathematical Finance 13(3). Scholes M. K. N. Mathematical Finance 13(2). 3. T. Kallsen. [18] Kramkov. (2003): Stochastic volatility for Lévy processes. Rieder. K. [12] Emmer.. (1987): The pricing of options on assets with stochastic volatilities. Karlsen.. B... Sircar. H. IEEE Transactions on Automatic Control . 637-654. E. Reikvam K. (2000): Derivatives in Financial Markets with Stochastic Volatility. 281-300. (2003): A complete explicit solution to the log-optimal portfolio problem. (2005): Portfolio optimization with unobservable Markov-modulated drift process. Journal of Finance 42. N. F. 442-447. (2004): Optimal portfolios when stock prices follow an exponential Lévy process. (2003): An optimal consumption model with stochastic volatility. H. Karlsen... J. 447-467. E. J. E.. Schachermayer... [7] Benth. (2003): Merton’s portfolio optimization problem in a Black and Scholes market with non-Gaussian stochastic volatility of Ornstein-Uhlenbeck type. D. Finance and Stochastics 5. 49. J. Cambridge: Cambridge University Press. Papanicolaou. [15] Goll. Vol.

R. P. [26] Zariphopoulou. 77-91. 369-393. (1969): Lifetime portfolio selection under uncertainty: The continuous time case. Finance and Stochastics 5.. 213-214.[19] Markowitz H. M-C. E. Review of economics and statistics 51. R.. 887-910. 61-82. Journal of Finance 7. 210-238. (2004): Bayesian inference for non-Gaussian Ornstein-Uhlenbeck stochastic volatility processes. 445466. T.O. Papaspiliopoulos. H. (2001): Optimal portfolio in partially observed stochastic volatility models. Journal of Economic Theory 3. O. [22] Nicolato. (2003): Option pricing in stochastic volatility models of the Ornstein-Uhlenbeck type. H. (1971): Optimum consumption and portfolio rules in a continuous time model. (1952): Portfolio selection. Dellaportas. (2001): A solution approach to valuation with unhedgeable risks. Journal of the Royal Statistical Society: Series B 66 (2). G. Annals of Applied Probability 11.. [23] Pham.. (1997): The normal inverse Gaussian Lévy process: Simulation and approximation. Mathematical Finance 13(4). 11 . Erratum (1973) 6. E. T. Quenez. [21] Merton. [24] Roberts. [20] Merton. [25] Rydberg. Venardos. 247-257. 373-413. Communications in Statistics: Stochastic Models 13(4).

A primary advantage with the model is that the optimal strategies are functions of only 2n model parameters and the volatility of each stock. Sweden Abstract We consider Merton’s portfolio optimization problem in a Black and Scholes market with non-Gaussian stochastic volatility of Ornstein-Uhlenbeck type. Recently. This allows us to obtain accurate estimates of the optimal strategies even when n is large. The model is primarily intended for assets which are dependent. We refer to these as news processes. 1 . and give statistical methods for both the …tting and veri…cation of the model to data. this was solved by Merton in [16] and [17]. In a Black and Scholes market with constant coe¢ cients. as well as for carefully reading through preliminary versions of this paper. Barndor¤-Nielsen and Shephard propose modeling the volatility in asset price dynamics as a weighted sum of non-Gaussian OrnsteinUhlenbeck (OU) processes of the form dy (t) = y (t) dt + dz (t) . We show that this dependence generates covariance. We assume that the dependence between stocks lies in that they partly share the Ornstein-Uhlenbeck processes of the volatility. [6] solved a similar problem for one stock and a bond in the more general market model of [3]. but not too dependent. The investor can trade in n stocks and a risk-free bond. such as stocks from di¤erent branches of industry. 1 Introduction A classical problem in mathematical …nance is the question of how to optimally allocate capital between di¤erent assets. we derive and verify explicit trading strategies and Feynman-Kac representations of the value function for power utility. and interpret this as that dependence between stocks lies solely in their reactions to the same news. In [3]. Using dynamic programming. Chalmers University of Technology and Göteborg University.News-generated dependence and optimal portfolios for n stocks in a market of Barndor¤-Nielsen and Shephard type Carl Lindberg Department of Mathematical Sciences. The author would like to thank Holger Rootzén and Fred Espen Benth for valuable discussions.

We show that this dependence generates covariance. [5]. In this extended model we consider an investor who wants to maximize her utility from terminal wealth by investing in n stocks and a bond. Recently. the general case is completely analogous. we then compute the solution to this equation via a Feynman-Kac representation. We extend the model by introducing a new dependence structure. and [10] treat di¤erent portfolio problems when the risky assets are driven by Lévy processes. [4]. We allow for the investor to have restrictions on the fractions of wealth held in each stock. By use of a veri…cation theorem. This process is assumed to be independent of the di¤usion process. This problem is an n-stock extension of [6]. The paper [19] uses partial observation to solve a portfolio problem with a stochastic volatility process driven by a Brownian motion correlated to the dynamics of the risky asset. we identify the optimal expected utility from terminal wealth as the solution of a second-order integro-di¤erential equation. Further. and [9] look at portfolio optimization in a market with unobservable Markov-modulated drift process. For simplicity of notation. Going beyond the classical geometric Brownian motion. In [11]. and skewness. portfolio optimization under stochastic volatility has been treated in a number of articles. We will refer to the OU processes as news processes. This is a desirable feature which allows us to do portfolio optimization with a large number of stocks. such as semi-heavy tails. For an introduction to the market model of Barndor¤-Nielsen and Shephard we refer to [2] and [3]. in which the dependence between assets lies in that they share some of the OU processes of the volatility. [13]. For option pricing in this context. The paper [8] models the stochastic factor as a continuous-time Markov chain with …nite state-space. In Section 2 we give a rigorous formulation of the market and the portfolio optimization problem. and obtain explicit optimal allocation strategies. Both [8] and [23] use an approach to solve their portfolio optimization problems that is similar to ours. volatility clustering. we have formulated and solved the problem for two stocks and a bond. which implies the interpretation that the dependence between …nancial assets is reactions to the same news. such as stocks from di¤erent branches of industry. and [23]. These strategies are functions of only 2n model parameters and the volatility of each stock. This paper has six sections. and give statistical methods for both the …tting and veri…cation of the model to data. This framework is a powerful modeling tool that allows us to capture several of the observed features in …nancial time series.where z is a subordinator and > 0. see [18]. All results are derived under exponential integrability assumptions on the Lévy measures of the subordinators. For power utility. the stochastic volatility depends on a stochastic factor that is correlated to the di¤usion process of the risky asset. The model is primarily intended for assets which are not too dependent. [14] derive explicit solutions for log-optimal portfolios in terms of the semimartingale characteristics of the price process. We also discuss the market 2 . The stochastic optimization problem is solved via dynamic programming and the associated Hamilton-Jakobi-Bellman (HJB) integro-di¤erential equation. However. as well as borrowing and short-selling restrictions on the entire portfolio.

F. :::. in the general setting. be two Wiener processes independent of all the subordinators. j = 1. where the rate of decay is denoted by j > 0: The unusual timing of Zj is chosen so that the marginal distribution of Yj will be unchanged regardless of the value of j : To make the OU processes and the Wiener processes simultaneously adapted. m. We also set up our optimization problem. and discuss. we use the …ltration f (B1 (t) . which implies that its sample paths are increasing. 3 . The stationary process Yj can be represented as Rt Yj (s) = 1 exp (u) dZj ( j s + u) .1) j Yj (t)dt + dZj ( j t). j = 1. to n stocks. i = 1. P ) with a …ltration fFt g0 t T satisfying the usual conditions. Z1 ( 1 t) . We now introduce our stochastic volatility model. our model is identical to theirs. m in our model as news processes associated to certain events. under a special dependence structure. The next extension of the model. We will discuss the di¤erences as they occur. z)l(dz) < 1: 0+ We assume that we use the cádlág version of Zj : Let Bi . B2 (t) . We prove our veri…cation theorem in Section 4. m: Remember that a subordinator is de…ned to be a Lévy process taking values in [0. In Section 3 we derive some useful results on the stochastic volatility model. and use it in Section 5 to verify the solution we have obtained. Section 6 states our results. and the jump times of Zj . and denote their Lévy measures by lj (dz). The Lévy measure l of a subordinator satis…es the condition Z 1 min(1. s t.1 The market model For 0 t T < 1. m as news or the release of information on the market.model and the implications of the dependence structure. To begin with. the market model. It is an extension of the model proposed by Barndor¤-Nielsen and Shephard in [3] to the case of two stocks. j = 1. j = 1. we assume as given a complete probability space ( . 1) . :::. and on moments of the wealth process. :::. 2 The optimization problem In this section we de…ne. :::. :::. Zm ( m t))g0 t T : From now on we view the processes Yj . 2. is only a matter of notation. without proofs. the OU stochastic processes whose dynamics are governed by dYj (t) = (2. Denote by Yj . 2. Introduce m independent subordinators Zj .

but also has very appealing economical interpretations. where " =L " denotes equality in law.but can also be written as Yj (s) = yj e j (s t) + Rs t e j (s u) dZj ( j u). We will call i + i i (t) the mean rate of return for stock i at time t: For notational simplicity in our portfolio problem we denote the volatility processes by i instead of the more customary i2 : We de…ne i as Pm t.3) dSi (t) = ( i + i i (t)) Si (t) dt + i (t)Si (t) dBi (t) . T ] . if yj = Yj (t) 0.y (2. De…ne the two stocks S1 . These volatility dynamics gives us the stock price processes Z sp Z s 1 (u) du + Si (s) = Si (t) exp + i (u)dBi (u) : (2. The increments of the returns Ri are stationary since Ri (s) Ri (t) = log Si (s) Si (0) log Si (t) Si (0) = log Si (s) Si (t) =L Ri (s t) . where !i. then Yj (s) 0.4) i (s) := i (s) := j=1 !i. to have semi-heavy tails as well as both volatility clustering and skewness. since Zj is non-decreasing. :::. with interest rate r > 0. s t. and yj has the stationary marginal distribution of the process and is independent of Zj (s) Zj (t) . s t: In particular.5) i i i 2 t t This stock price model does not have statistically independent increments and it is non-stationary. :::. j = 1.1 is not only simple from a statistical point of view. ym ) : We assume the usual risk-free bond dynamics dR (t) = rR (t) dt. S2 to have the dynamics p (2.j 0 are weights summing to one for each i: The notation it. m. i = 1. It also allows for the increments of the returns Ri (t) := log (Si (t) =Si (0)) .2 Discussion of the market model This section aims to show that the dependence structure proposed in Section 2.y denotes conditioning on Y (t) : Our model is here not the same as just two separate models of Barndor¤-Nielsen and Shephard type. . The di¤erence is that the volatility processes depend on the same news processes. Here i are the constant mean rates of return. and set y := (y1 . We set Zj (0) = 0. 2. The paper [3] suggests a model with n stocks with dynamics dS (t) = f + (t)g S (t) dt + 4 1 (t) 2 S (t) dB(t). and i are skewness parameters. 2.j Yj (s) . (2.2) where yj := Yj (t) . s 2 [t.

stocks from di¤erent branches of industry. For example. dependence is modelled by covariance. R2 (u)) Cov (R1 (s) . T ] . for s. As we will see. which is the most frequently used measure of dependence in …nance. 1 2 11 2 12 (s and S2 (s) = S2 (t) exp 1 2 21 2 1 2 22 (s t) + p 21 B1 (s) + p 22 B2 (s) .j V ar (Yj (0)) j=1 5 e jt 1+ 2 j jt : .6) 1 2 m X !1. v 2 [0.7) 1 2 m X !1. The model is mainly intended for assets that are dependent. t.j !2. it is su¢ cient to show this result for Ri . for every speci…c value of the volatility i . and are vectors. T ] : As will be shown below. for the purpose of portfolio optimization we do not need to know the weights !i. which for s = t simpli…es to Cov (R1 (t) . 2: Note that we have Cov (R1 (s) R1 (t) . R2 (t)) =2 1 1 2 2 (2. stock prices develop independently beside from reacting to the same news. the model generates a non-diagonal covariance matrix for the increments of the returns over the same time period. In our model. For example. for s. in the classical Black and Scholes market. From an economic viewpoint.j !2. t) 2 j . i. R2 (v)) Cov (R1 (t) . Note that we do not make a distinction between good and bad news. R2 (v)) .where is a time-varying stochastic volatility matrix. Instead we ask that every time the market is ”nervous”to a certain degree. However. we do not require stability over expected rate of return.e. In the case of two stocks this means that for s t.j V ar (Yj (0)) j=1 e js +e jt j js e tj 1+2 j min (s. the mean rate of return i + i i will be the same. t 2 [0.j : More importantly. but not too dependent. one can expect the model parameters to be more stable than in the classical Black and Scholes market. and B1 (t) = B2 (t) = 0. This model includes ours as a special case with being a diagonal matrix. p p 1 1 S1 (s) = S1 (t) exp t) + 11 B1 (s) + 12 B2 (s) . R2 (u)) + Cov (R1 (t) . R2 (t)) = 1 2 1 2 (2. R2 (u) R2 (v)) = Cov (R1 (s) . we have that Cov (R1 (s) . We can interpret this as that we only need stability in how the market reacts to news. i = 1. u. Since the returns have stationary increments. for a volatility matrix . and B is a vector of independent Wiener processes.

by de…nition of i we have that E [R1 (s) R2 (t)] Z s =E 1+ 1 2 1 0 Z 2 + 1 2 2 1 2 st + 1s (u) du + 2 m X 1 2 2 + 2t m X 1 !1. the independence of the increments of Yj gives that Cov (Yj (u) .j=1 Z t Yj (u) du : 0 Similarly.6). E [R1 (t)] = 1t + 1 m X 1 2 !1.i !2. we have that E [Yj (t)] = Yj . R2 (t)) = 1 1 2 1 2 2 m X Z !1. Yj (v)) = e 6 j jv uj V ar (Yj (0)) .j Cov s Yj (u) du. for some constant Yj > 0. 1) : To derive Equation (2.j E t Yj (u) du 0 j=1 1 2 Z 0 t 0 = 1 (u) du + s Yj (u) du 0 Z !1. R2 (t)) in the entire interval ( 1. It turns out that we can get correlations Corr (R1 (t) . 0 j=1 Z t Yj (u) du : 0 By stationarity.j E j=1 + 1 2 1 m X 1 2 2 s Z p 1 Z tp (u)dB1 (u) 2 (u)dB2 (u) 0 Z !2. for all t 2 R: If we assume that u v. 2 Yj . but we do not immediately know which correlations that can be obtained.j E j=1 Z t Yj (u) du : 0 This gives that Cov (R1 (s) .j !2.j E s Yi (u) du 0 i.This result says that the model generates a covariance matrix between returns. Yj (v)) =E Yj (u) =e j (v =e j (v Yj (v) Yj Z v 2 Yj (u) + Yj (u) e j (v u h i 2 u) E Yj (0) e j (v u) 2Yj j (v =E e Yj u) u) s) dZ ( j s) V ar (Yj (0)) : The same calculations for v u shows that Cov (Yj (u) .

j j e jt 1+ jt 2 j + !2. for s = t.j V ar (Yj (0)) jt e 1+ jt 2 j j=1 + !i.8) 0 t Cov (Yj (u) . R2 (t)) = 1 2 1 2 1 2 1 1 e 1 2 1 2 2 2 js +e jt m X !1. 2 V ar (Y (0)) !1.j + 2 1 Yj s 1 2 2 1 v u uPm t j=1 and.j ar (Yj (0)) 7 e jt 1+ 2 j jt + !1. similarly as above. t) 2 j 1 v u uPm t j=1 js 1+ js 2 j !1. 0 s Z Z t Yj (u) du (2. Yj (v)) dudv 0 0 = V ar (Yj (0)) e js +e jt j js e tj 1+2 j min (s. Corr (R1 (t) .j V ar (Yj (0)) jt e 1+ j=1 v u uPm t j=1 jt 2 j 1 2 V !1.j 2 2 Yj t 1 2 2 ! !.j 2 1 Yj t 1 2 2 ! Yj t ! . R2 (t)) = 1 2 1 1 2 m X 1 2 2 1 2 1 2 !1. t) 2 j : By Itô’s isometry (see [24]) we get.j !2.j V ar (Yj (0)) j=1 e j js tj 1+2 j min (s.and we get Cov = Z Z s Yj (u) du. 2: This gives Corr (R1 (s) .j !2.j j e 2 V ar (Y (0)) !2. V ar (Ri (t)) = m X 2 i 1 2 2 2 !i. .j for i = 1.

and k +1 is the number of consecutive trading days p in our period of observation. We will do this for a Normalized Inverse Gaussian distribution (N IG) . An alternative stock price model would be p 2 2 dSi (t) = i + i1 i1 (t) i (t)Si (t) dBi (t) . [3]. 2. i2 > 0.3 Statistical methodology In this section we describe a methodology for …tting the model to return data. which has been shown to …t …nancial data well.1 v u uPm t j=1 2 V ar (Y (0)) !2. the marginal distributions of the returns will no longer …t in the framework of Barndor¤-Nielsen and Shephard. and 0 j j : A standard result is that if we take to have an Inverse Gaussian distribution p " (IG) . and hence we need less data to obtain good estimates of the model parameters. Ri (2 ) Ri ( ) . We are also required to obtain estimates of the "positive" respectively "negative" volatilities in the statistical estimation of the model. Our choice plays no formal role in the analysis. > 0. p where q (x) = 1 + x2 and K1 denotes the modi…ed Bessel function of the third kind with index 1: The domain of the parameters is 2 R. see e. An obvious advantage of our model is that we do not have to estimate a stochastic volatility matrix. where and are the same as in the N IG-distribution.j j e jt 1+ jt 2 j + !2. The IG-distribution has density function fIG (x. i i (t) Si (t) dt + where i1 . Ri (k ) Ri ((k 1) ) . where is one day. then x = + + will be N IG-distributed. . . and i1 . A drawback is that. and [21]. . :::. 1)-distributed random variable ". . . We assume that we are observing Ri ( ) . Another drawback is that we do not distinguish between good and bad news. The existence and integrability of Lévy measures lj such that the volatility processes i will have 8 . ) p 2 = exp 2 1 x q K1 x q e x. x > 0. and draw a N (0. to obtain high correlations. . ) = p 2 exp ( )x 3 2 exp 1 2 2 x 1 + 2 x . i2 are linear combinations of the news processes such that i1 + i2 = i : We have chosen to not use this model as it would be hard to estimate from data. [1].j 2 2 Yj t 1 2 2 !: There is always a trade-o¤ between accuracy and applicability when designing models. 2 + 2 . we need the model to be very skew. . For example. and : Its density The N IG-distribution has parameters = function is fN IG (x.g. This might not …t observed data.

and that we have found a such that our model has the right autocorrelation function.m exp ( m jhj) . i ) .3 Y3 IG ( 1 . However. a 1 2 X : Furthermore.1 Y1 IG ( 1. (h) = Cov ( i (h) . ) are the parameters of the IG-distribution.IG-distributed marginals is not obvious. 1 ) .7) we can now …t the covariance of the model to the empirical covariance from the return data. for reasons to be explained later. i = 1. see [3]. The method described in [3]. We assume that we have …tted N IG-distributions to the empirical marginal distributions of two stocks. we have that X + Y we can let !1. ”of their own.1 Y2 + !2. we have that aX 1 IG a 2 X. i = 1. 1 ) !1. This can be done by letting the two stocks ”share”the news process Y3 .”In general. We formulate this mathematically as 1 2 = !1. For X IG ( X . we immediately get that i (h) = exp ( jhj) : We proved this simpler result in Subsection 2. and that all 1 = 2 = 3 = : For our model calculations show that. if Y IG ( Y . Observe that since we have assumed the rates of decay j to be equal. Y ) and is independent of X and we assume that IG ( X + Y . this is done for each rate of decay. We denote the IG-parameters of the volatility processes i by ( i . 9 . 2: By Equation (2.1 Y1 + !1. i (0)) =V ar ( i (0)) . and each have one of the news processes Yi . This can be done by empirically calculating the autocorrelation functions i (h) for di¤erent values of h.3 Y3 IG ( 1.1 exp ( 1 jhj) + ::: + !i. 1 ) IG ( 2 .2. uses that the marginal distributions of the volatility processes i are invariant to the rates of decay j : These parameters j are then used to …t the autocorrelation function of the i . which we further extend. ) : Because of this formula X = Y =: . The proof of the general case is analogous. h 2 R.j 0. are the weights from the volatility processes that sum to one. i (h) = !i. where ( . X ). where the !i. we will assume that m = 3. 2. 2 ) : We now state two properties of IG-distributed random variables that we will need below. i For simplicity of exposition we will assume that we only need one to correctly model the autocorrelation function of both stocks. and then …nd a so that the theoretical and empirical autocorrelation functions match.1 . to log-return data.3 .3 Y3 = !2. See [2] and [22] for this theory. The Lévy density l of the subordinator Z of an IG-distributed news process Y is l (x) = (2 ) 1 2 2 x 1 + 2 x 1 2 e 2x 2 . for general m.

3 2. Hence we only have to take care of the covariance of the returns Ri .8).3 3 2 e j 1+ 2 j j =C (2.1 + 2. nor does it matter how many di¤erent rates of decay we use. are t i (u) du. It is now straightforward to check that for reasonably small C there are non-unique choices of !i.. are semimartingales. i = 1.12) must be identical. that the two expressions for the distribution of Y3 .1 . we see that Equation (2. t where convergence is uniformly on compacts in probability. Note that there is nothing crucial in our choice of covariance as measure of dependence.3 Y3 IG ( 2. denoted by [ ] .9). s t: That is.j such that we can obtain both the right autocorrelation function of i and a speci…c covariance for the returns. (2. 2.3 1 1 1 2 2. (2.. their quadratic variations. !1. We begin by …tting a marginal distribution to return data.11) . we have Z s [log (Si =Si (t))] (s) = i (u) du. in which we use that the variance V ar (Y (0)) of a stationary inverse Gaussian process Y is = 3 . for a sequence of random partitions tending to the identity. by the scaling property of the IG-distribution. use that 10 .9) and !2. For each trading day we now empirically calculate the integrated volatility.32 Y3 IG !2.10) We see. We now give a simple approach to determine how well our model captures the true covariance.3 2 . This is a standard result in stochastic calculus. The autocorrelation function parameter is already correct by assumption. 2 ) .3 !2.3 . (2. we calculate the quadratic variation of the observed returns over a trading day and.where 1.. (2. (2.1 + 1.13). and we constructed the news processes Yj so that their marginal distribution would not depend on it. We do this by using Equations (2. that is. 1 2 1 Y3 IG !1. With the aid of Equation (2.3 = 2. thereby obtaining the parameters i and i . by the formula above. 2: Since we have that the return processes R R si .1 Y2 !2. i = 1.13) where C is the covariance that we want the returns to have.7) becomes 2 1 1 2 2 1 2 !1.3 = 1. where 2..32 and 1.3 . 2 ) IG ( 2. !2.3 .

as a constant approximation of the volatility during that day. bi . for all t 1 + 2 s T: The analysis is analogous in this case. i = 1. The utility is measured by a utility function U chosen by the trader. we get approximations of the volatility processes i for that period of time. 1] . We then calculate the covariance-matrix of both the return data set and the simulated alternative returns and compare them statistically.s. are weighted sums of the news processes. and c d. Most of the later sections will be devoted to …nding and verifying solutions to it. we need to adapt their results to our case. such that the constraints would have taken the form i 2 [ai . these restrictions are partly for mathematical convenience. ai < bi . and 1 + 2 1. but more notationally complex. bi ] . We begin by de…ning a value function V as the maximum amount of expected utility that we can obtain from a trading strategy. 2. in a sense. as well as short-selling and borrowing of capital. is dynamic programming and stochastic control. 2 ) : The fraction of wealth held in the risk-free asset is (1 1 2 ). a. for example. as it is. in that it concretizes how much the trader is willing to risk to obtain a certain level of wealth.3). which implies the conditions i 2 [0. 2.4 The control problem A main purpose of this paper is to …nd trading strategies that optimizes the trader’s expected utility from wealth in a deterministic future point in time..5). an optimality condition. We state the main results in this setting in Section 6. Our approach to …nding these trading strategies. i and generated N (0. We will make use of many of the results in [6]. We then set up the associated Hamilton-Jakobi-Bellman equation of the value function V: This equation is a central part of our problem. law enforced restrictions on the fraction of wealth held in a speci…c stock. Denote by i (t) the fraction of wealth invested in stock i at time t. R (s) where i (s) W (s) =Si (s) is the number of shares of stock i which is held at time s: We also assume that the portfolio is self-…nancing in the sense that no capital 11 . Using the …tted parameters i . This general setting allows us to consider. If we do this for a number of trading days. given a certain amount of capital. This utility function U is a measure of the trader’s aversion towards risk. d 2 R. i = 1.s. and the value function V . In this section we set up the control problem under the stock price dynamics of Equation (2. for all t s T: However.4). We allow no short-selling of stocks or bond. a. The wealth process W is de…ned as W (s) = 1 (s) W (s) S1 (s)+ S1 (s) 2 (s) W (s) (1 S2 (s)+ S2 (s) 1 (s) 2 (s)) W (s) R (s) .. However. since most of their ideas are applicable in our setting. we can now simulate ”alternative” returns. Recall that 1 and 2 . c. see Equation (2. 2. We could equally well have chosen constants ai . and set = ( 1 . c < d. 1)-distributed variables in Equation (2.

1 For constants cj > 0 to be speci…ed below. 2. :::. :::.1 holds. An investment strategy = f (s) : t s T g is said to be admissible if 2 At . ym ) 2 Rm 0+ . and assume that y = (y1 . 1] . and a unique solution W of Equation (2. This can be formulated mathematically as 2 Z X W (s) = W (t)+ i=1 s i (u)W (u) Si (u) t Z dSi (u)+ s (1 1 (u) 2 (u)) W (u) R(u) t dR (u) . and 1 + 2 1 a. Hence Condition 2. ) are the parameters of the IG-distribution. T ] : See [15] for a motivating discussion. j = 1.14) existsg. m: 0+ Recall that the Lévy density l of the subordinator Z of an IG-distributed news process Y is l (x) = (2 ) 1 2 2 x 1 + 2 x 1 2 e 2x 2 . w. where ( .w. De…ne the domain D by D := f(w. y. The self-…nancing condition gives the wealth dynamics for t s T as dW (s) = W (s) 1 (s) ( 1 + 1 1 (s) r) ds (2. Later we will need some exponential integrability conditions on the Lévy measures.1 The set At of admissible controls is given by At := f = ( 1 .is entered or withdrawn. for all t s T. with initial wealth W (t) = w: The following de…nition of the set of admissible controls now seems natural.s. 12 . 1) by R0+ . 2 ) : i is progressively measurable. De…nition 2.15) 0+ as long as a cj with cj from Condition 2. We therefore assume that the following holds: Condition 2.1 is satis…ed for cj < 2 =2: We know from the theory of subordinators that we have Z 1 h i E eaZj ( j t) = exp j (eaz 1) lj (dz)t (2. 1) by R+ and [0. i (s) 2 [0. i = 1. Denote (0.14) + W (s) 2 (s) ( 2 + 2 2 (s) r) ds + rW (s)ds p p + 1 (s) 1 (s)W (s)dB1 (s) + 2 (s) 2 (s)W (s)dB2 (s). Z 1 (ecj z 1) lj (dz) < 1. ) = Et. y) 2 R+ Rm 0+ g: We will seek to maximize the value function J(t. for all s 2 [t.y [U (W (T ))] .

1+ 1 + 2 + m X j=1 1 2 2 1 1 + f( 2 2 2 1 ( 1 + w2 vww 1 1 r) + + rwvw 2 ( m X 2 + 2 2 r)) wvw (2.w. and of sublinear growth in the sense that there exists positive constants k and 2 (0. w. j j t) > 0. for all (t. The proof of Lemma 3. y) = J(t.1 can also be found in [6]. y. T ) D: We observe that we have the terminal condition V (T. w. y) 2 D. y) = sup J(t.18) and the boundary condition Rm 0+ : V (t. w. w. y) = U (w). m: The function U is the investor’s utility function. T ] D. y) 2 [0. y.19) Preliminary estimates This section aims at relating the existence of exponential moments of Y to exponential integrability conditions on the Lévy measures. 1) so that U (w) k(1 + w ) for all w 0: Hence our stochastic control problem is to determine the optimal value function V (t. y)) lj (dz). j = 1. w. non-decreasing. y) 2 [0. such that V (t. It is assumed to be concave. ). bounded from below. (2. We get from the dynamics (2. (2. y) 2 [0. 0.y means expectation conditioned by W (t) = w.17) j yj vyj j=1 j Z 1 (v (t.1 holds with cj = Z 1 Z s j yj + j exp E exp( j Yj (u)du) exp j t j= j jz Proof. w.1) of Yj that Z s Yj (u)du = yj + Zj ( j s) Zj ( j t) j t yj + Zj ( j s) = yj + Zj ( 13 j (s Zj ( t)). the optimal investment strategy. ): The HJB equation associated to our stochastic control problem is 0 = vt + max i 2[0. w.i=1. for all (w. w.where the notation Et. y + z ej ) v (t. T ] 3 (2. Lemma 3. (t. and Yj (t) = yj . 0 for (t.2. y) = U (0). Then 1 lj (dz)(s j 0+ L for Yj (s) t) .1].16) 2At and an investment strategy 2 At . w.1 Assume Condition 2. as well as developing moment estimates for the wealth process and showing that the value function is well-de…ned. :::.

j z 1 lj (dz): j 0+ j=1 1 j j=1 where 2 j+ Proof.w. w. 2 (u)) du t Z s Z s p p + 1 (u) 1 (u)dB1 (u) + 2 (u) 2 (u)dB2 (u) . y) U (0): The sublinear growth condition of U gives that V (t.j . Lemma 3. 2) = 1 (u)( 1 + 1 1 r) + 2 (u)( 2 + 2 1 1 ( 1 (u))2 1 ( 2 (u))2 2 : 2 2 +r De…ne X(s) = exp Z s 2 1 (u) t 1 2 Z s (2 )2 ( p 1 (u)dB1 (u) + Z s 2 2 (u) t 2 1 (u)) 1 (u) du t 14 1 2 Z t s 2 (2 ) ( p 2 2 2 r) (u)dB2 (u) 2 (u)) 2 (u) du : . :::. We know that U (w) U (0) since U is non-decreasing.14) and Itô’s formula that Z s W (s) = w exp (u. > 0: Then sup Et. for 2 At .y (W (s)) 2At 0 m X (j 1 j + ) !1. t t where (u. Lemma 3.j +2 (j j )!2.2 Assume Condition 2.j 1j + ) !1. m. This gives that E [U (W (T ))] U (0). We have by Equation (2. and " =L " denotes equality in law. y) = sup E [U (W (T ))] 2At k 1 + sup E W (T ) : 2At This means that we obtain an upper bound to the optimal value function if we have control of the wealth process.j + (j w exp @2 C( ) = (j rj + j 2 rj + r) Z 1 (j exp 2 j 1 + 1 2 m X 2 j + ) !2.since Yj (s) 0 when yj = Yj (t) 0. j = 1. The proof is analogous to [6. Recall that we have de…ned Zj (0) = 0: The result follows from Equation (2.1 holds with cj = for some 1 j+ 2 (j )!1.j + (j t)A . we only sketch the details.3].15). yj + C( )(s 2j + ) !2. which implies that V (t. Hence. 1. 1 (u) . w.

1]. 2 ) de…nes an optimal investment strategy in feedback form if Equation (2. From now on we assume that Condition 2. m: Proof. Y (s))) i ds < 1. W (s) . Theorem 4. w.1 ([0. :::. 2. and introduce the operator M v := ( 1 ( 1 + 1 1 r) + 15 2 ( 2 + 2 2 r)) wvw . :::. j = 1.1 Assume that v(t.We can prove that X is a martingale. y) = v(t. 1) [0.19). y) V (t. and sup 2At Z T 0 h 2 E ( i (s)) i (s) (W 2 2 (s)) (vw (s. T ) (0.14) admits a unique solution W and h i V (t. T ) D and 2 At . there exist measurable functions i (t. in addition. y on [0.1 ([0. :::. 1) [0. 1)m with continuous extensions of the derivatives to t = 0 and yj = 0. T ) (0.1 in [6]. y) 2 [0. 2: Then v(t.17) with terminal condition (2.2. w. W (s) . y) 2 [0. for all (t. m: This ensures that the value function is well-de…ned. Y (s ))j] lj (dz)ds < 1. T ) [0. y) 2 [0.2. then = ( 1 .w. For j = 1. y) 2 [0. y). being the maximizers for the max-operator in Equation (2. This can be used together with Hölder’s inequality to get the result. j = 1. y) = Et. 4 A veri…cation theorem We state and prove the following veri…cation theorem for our stochastic control problem. Therefore we omit some details. w. w.y U W (T ) .18) and boundary condition (2.j +2 (j j 2 j+ )!2. w. for all (t. i = 1.j .1 holds with cj = 1 j+ 2 (j )!1. Y (s ) + z ej ) v (s.17). m. y) 2 C 1. w. 1) and once continuously di¤ erentiable in t. i = 1. w. assume sup 2At Z T 0 Z 1 0+ E [jv (s. 1)m ) means twice continuously di¤ erentiable in w on (0. T ] D: If. w. 1)m ) \ C([0. W (s) . The proof is similar to Theorem 4. w. Let (t. T ] D: The notation C 1. T ] D) is a solution of the HJB equation (2.

w. y + z ej ) v(t. Y (s))] Z s (vt + L v) (u. y) + E t 2 0 1 3 Z s @vt + v(t. y). w. y) + E 4 max L v A (u. W (s). w.i=1. y) : Explicit solution In this section we construct and verify an explicit solution to the control problem (2. w. as well as an explicit optimal control . by putting s = T and invoking the terminal condition for v: The …rst conclusion in the theorem now follows by observing that the result holds for t = T and w = 0: We prove the second part by verifying that is an admissible control. when the utility function is of the form U (w) = 1 w . Y (u)) du = v(t.2. y) = E U W This proves the theorem. 2 (0. w. max L v = L v. 16 . w. t 1+ 2 1 = v(t. w.1].1 Reduction of the HJB equation In this subsection we reduce the HJB equation (2.+ + 1 2 m X 2 1 1 + 2 2 2 w2 vww + rwvw m X j yj vyj j=1 j Z 1 (v(t.17) to a …rst-order integrodi¤erential equation by making a conjecture that the value function v has a certain form.16). Since is a maximizer. 5 (T ) i V (t. W (u). for all 2 At . w. y) E [U (W (T ))] . Y (u)) du5 i 2[0.i=1.2. We get now that v(t. W (u).1]. i 2[0. 1+ 2 1 which for s = T gives that h v (t. 1): 5. y)) lj (dz) : 0+ j=1 Itô’s formula gives that E [v(s.

4) 1. y) j y j hy j (t. y).2) j=1 m X + j Z 1 (h (t. For our purposes.We conjecture that the value function has the form 1 v(t. (5. y) 2 [0. and i when r i = 0. (5.1) ) + r: If we insert the conjectured value function into the HJB equation (2.1 Note that we can …nd a constant j ( 1. y + z ej ) h (t. 2 )j +j 17 1j 1 i.13 and Remark 4. Theorem 4. y) 2 [0. see for example [7. : [0. (5. w. w. T ) h (T. y) = 1. T ] for some function h(t. Remark 5. 1)m : The terminal condition becomes where (t.17) we get a …rst-order integro-di¤erential equation for h as 0 = ht (t. y) = U (w) = 1 w .2 1+ 2 1 2 1 1 1 2 f ( 1 1 + 2 2 2 + r) + 1 1 (1 D. y) (5.6) Note that this strategy only depends on the parameters for each stock.1]. the optimal fractions of wealth are + r) 1 1 ) ( 1 ( + 2 + 2 2 r) 2) + 2 ( 1 + 2) . 1) ! R as r) 2 2 (5. we will need to be continuously di¤erentiable. y)) lj (dz) . 8y 2 [0. 0+ j=1 [0. Calculations give that our candidates for optimal fractions of wealth are i whenever i 2 (0. 1) 2 ( 2 + [0. 1)m .3) i < 1.5) and 2 =1 1: (5. i. 1) and ( i) = 1 + 2 1 1 i ( 1. 1 0: When 2) = 1 1 (1 + 2 ( 1 + i .14]. (t. This follows from Danskin’s theorem. m X 2 )h (t. and the volatility > 0 such that +j 2j 2: . w. y) + ( 1. y) = w h(t. 2) = max i 2[0. y): We de…ne the function ( 1.i=1. since v(T.

2 )j +j > 0: Therefore. y) by h i RT h (t. 1 2 (s)) ds (5.j ) j j=1 1 j !1.y exp t ( 1 (s) . T ] " exp ( (s) .j ) j=1 18 Rt 0 y 2j (s) ds 3 2 Yj j (s)ds 5 : 2 j !2.y exp 0. y) = 1: The only di¤erence between the two functions is the direction of the time variable t: We show now that g is well-de…ned under an exponential growth hypothesis in 1 and 2 : Lemma 5. 2 (s)) ds 0 . :::. m: Then 0 m X (j 1 j !1. we have that " !# Z T h (t. y) = Ey exp ( 1j 1 +j 2j 2 for some constant 1 (s) .j + j g (t. 2 (s)) ds . y) = Ey exp g (t. Proof. y) exp @kt + j (j 2 j !2. y) = Et.2 Veri…cation of explicit solution In this subsection we de…ne a Feynman-Kac formula that we verify as a classical solution to the related forward problem of Equation (5.j +j 2 j!2. From Remark 5. By the stationarity of Y.7) t Z T !# t ( 1 (s) .1 Assume Condition 2. T ] Rm 0+ : We prefer to re-write the function h on a form that is simpler for us to handle.j for some positive constant k: +j 1 yj A .2).1 we know that j ( 1.5. y) = Et. Rm 0+ :We de…ne now t. y) 2 [0.1 holds with cj = j = 1.y =E for (t. y) 2 [0. We indicate then how we can show that our conjectured solution v coincides with the optimal value function V: De…ne the function h (t. y) := h (T Z t ( 1 (s) .j ) for . 2 (s)) ds 1 (s) + j 0 Ey exp 2 e t Ey 4 Z t + j 0 m Y e (j 1j 1 j!1. (t. Z t g (t. 2 (s)) ds : 0 Note that g (0.

j j=1 +j 2 j !2. m. :::. T ]) Rm 0+ .j ) for j = 1. y+z ej 2 ( y 1 (s) . m. we have that t 0 Z t exp j Z 1 0+ jg (t. y 2 (s)) + cj ( y 1 (s) . we have that E "Z 0 T 2 j !2.2 Assume Condition 2. We therefore give a di¤erent proof. Y (s) + z ej ) 19 # g (t. We will need that g is continuously di¤erentiable in y for h to satisfy Equation (5. that is. y)j 0 Z t Ey exp (j . y) is continuous for all y 2 Rm and g (t. Since y Yj j (s) yj + Zj ( 1 yj A (exp (cj z) 1) .1 holds with cj = (1 2 ) (!1.3 Assume Condition 2. m: Then g 2 C 0.1. y 2 (s)) ds ( y 1 (s) .2). g ( .4 in [6] is incorrect. Lemma 5. The proof is analogous to Lemma 5. Then "Z Z m T 1 X E jg (t. Y (u))j lj (dz) du < 1: y+z ej 1 (s) . :::. Since cj jg (t. T ] : 0+ Proof.1.1 ([0. Y (u) + z ej ) 0 j=1 0+ 0 yj Proof.j ) .2). j = 1.j ) g (t. j = 1. y + z ej ) Z Ey exp j= 1 j !1.j + j 2 j !2.j ) j for k1 > 0 by Lemma 5. y 2 (s)) ds (s) ds j ze js ds 0 Z exp 0 t 0 exp @k1 t + m X (j 1 j !1. To prove that g solves the suitably modi…ed Equation (5. the proof of Lemma 5. ) is once continuously di¤ erentiable for all t 2 [0. we need the following result concerning the expectation of the jumps of g: In our view.j + !2. Lemma 5.3 in [6]. Y (s))j lj (dz) ds + . j s) .j +j # g (t.By independence of the Yj .1 holds with cj = j (j 1 j !1. we get the result by Lemma 3. :::.

y)) lj (dz) 0+ [0.j ) + (5. :::. :::. 1)m : Moreover.5]. the most important results for the case of n stocks. for j = 1.2). and we can easily proceed as in [6] to check that all the assumptions in Theorem 4. which is what we wanted to show. y + z ej ) g (t. y) is a classical solution to the related forward problem of Equation (5. for k2 > 0: We give now a proposition that shows that g (t. Its proof is very much the same as in [6.j ) . and we omit it.Z T ek1 t+k2 0 Z 1 Pm j=1 yj 2 0 E 4exp @ (exp (cj z) m X (j 1 j !1. T ] g 2 C 1.j ) + (12 ) (!1. so that for (t. y) = w h (t. namely 1 v (t.j + (j 2j + 4 ) !2. y) ( 1 . T ] [0.8) j yj gyj (t. Assume now that cj = 8 ((j j 1j + 4 ) !1.j + !2. gt is continuous.6). bi ] . and some " > 0. y) + (5.j +j 2 j !2.j + !2. y) j=1 m X j=1 j Z 1 (g (t.3) (1 ) 2 (!1.j ) + " for j = 1. :::.1 are satis…ed. by the assumptions. m.j ) Zj ( j j=1 1) lj (dz) 13 j s) A5 ds 0+ < 1. m: We note that this implies that the optimal value function V is well-de…ned. w. 6 Generalizations In this section we state. 20 .1 Assume there exists " > 0 such that Condition 2. n.1 is satis…ed with cj = 2 j (j 1 j !1. ) belongs to the domain of the in…nitesimal generator of Y and m X 0 = gt (t. Hence we have proved that our conjectured solution coincides with the optimal value function.1 ((0. Then g (t. i = 1.j + j 2 j !2. y) 2 (0. i (s) 2 [ai . 2 ) g (t. Proposition 5. 1)m ) : From our conjecture of the form of the value function we have now our explicit solution candidate. Proposition 5. without proofs.9) is given by Equations (5. y) : The candidate for the optimal feedback control to (5.

T. 21 . y) = U (w). y) = 1 w h (t. Mikosch and S. T ] The solution to this equation can be shown to be h RT v (t.:::. [2] Barndor¤-Nielsen. E. y + z ej ) v (t. y)) lj (dz). for all (w. T ) Z D: We still have the terminal condition V (T. :::.bi ]. for all (t.bi ]. y) = U (0).and n X c i d: i=1 It can be seen that the additional di¢ culty in this setting is merely notational. O. y) 2 D. w. Barndor¤-Nielsen. :::. 2 i i i )ds . Resnick). n) References [1] Barndor¤-Nielsen. E.n. :::. Shephard. ) (6. 0 j=1 for (t. y) = 1 w Ey e t where ( Rm +: y y 1 (s). n (s) is de…ned as ( = 1 . O. n c i=1 i d ( n X i ( i + i=1 i i r) 1 2 n X i=1 The optimal fractions of wealth are given by the parameters that obtain ( 1 . Finance and Stochastics 2. N. w. Boston: Birkhäuser.:::. (1998): Processes of normal inverse Gaussian type. n ) in Equation (6.:::. (2001a): Modelling by Lévy processes for …nancial econometrics. w.Theory and Applications (eds O.1) + r: =( 1 .E. 0. 41-68.i=1. y) 2 [0. The HJB equation associated to this stochastic control problem is ( ) n n X X 1 2 2 0 = vt + max wvw r) + w vww i( i+ i i i i 2 i 2[aP i .1).i=1. y) 2 [0. in Lévy Processes . w. w.. i=1 i=1 n i=1 c +rwvw m X i d j yj vyj + j=1 m X j 1 (v (t. and the boundary condition V (t. n) max i 2[aP i .n.

(2004): Optimal portfolios when stock prices follow an exponential Lévy process. K. H. 247-257. 49. F. Annals of Applied Probability 13. F. N. Rieder. K. U. 275-303. 245-262. J. (1999): Real Analysis. Finance and Stochastics 5. Review of Economics and Statistics 51. Hernández-Hernández. Karlsen. Karlsen. [16] Merton. Sircar.. 447-467.. 17-44. to appear in Journal of Applied Probability. Shapiro A.. R. [14] Goll. J. E. (2005): Portfolio optimization with unobservable Markov-modulated drift process. (2001b): Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in …nancial economics. [8] Bäuerle. N. Inc. (2001a): Optimal portfolio selection with consumption and non-linear integro-di¤erential equations with gradient constraint: A viscosity solution approach. H.. F. Shephard. K. E. (2001b): Optimal portfolio management rules in a non-Gaussian market with durability and intertemporal substitution. (2003): A complete explicit Solution to the log-optimal portfolio problem.. (2000): Derivatives in Financial Markets with Stochastic Volatility. Rieder. O.. Finance and Stochastics 5. [4] Benth. (1969): Lifetime portfolio selection under uncertainty: The continuous time case. [15] Korn. 167-241 (with discussion). IEEE Transactions on Automatic Control . H. [12] Folland.. G. 774-799. N. No. T. (2003): An optimal consumption model with stochastic volatility.. [13] Fouque. (2000): Perturbation Analysis of Optimization Problems. 442-447. [7] Bonnans.[3] Barndor¤-Nielsen. (2004): Portfolio optimization with Markovmodulated stock prices and interest rates. New York: Springer. Klüppelberg. Finance and Stochastics 8. R. 3. Reikvam K. Ltd. Reikvam K. E. S. E. Mathematical Finance 13(2). Karlsen. [11] Fleming.. Papanicolaou. Cambridge: Cambridge University Press. J. U. R.. (2003): Merton’s portfolio optimization problem in a Black and Scholes market with non-Gaussian stochastic volatility of Ornstein-Uhlenbeck type. Vol. Reikvam K. Singapore: World Scienti…c Publishing Co... New York: John Wiley & Sons. F. K. 22 . D.. 215-244. H. Finance and Stochastics 7. [6] Benth... G. Journal of the Royal Statistical Society: Series B 63. [10] Emmer. Pte. Kallsen.-P. C. (1997): Optimal Portfolios. W. [9] Bäuerle. [5] Benth.

Communications in Statistics: Stochastic Models 13(4). Erratum (1973) 6. [22] Sato. 23 . T. Mathematical Finance 13. [18] Nicolato. (2001): A solution approach to valuation with unhedgeable risks. E. Berlin: Springer. K. Annals of Applied Probability 11. 210-238. Venardos. H. Cambridge: Cambridge University Press. 887-910. E. 5th ed. Quenez. M-C. (2001): Optimal portfolio in partially observed stochastic volatility models. [20] Protter. (1998): Stochastic Di¤ erential Equations. R. [19] Pham. (2003): Option pricing in stochastic volatility models of the Ornstein-Uhlenbeck type. (1971): Optimum consumption and portfolio rules in a continuous time model. T.[17] Merton. H.. Finance and Stochastics 5. [23] Zariphopoulou. (1997): The normal inverse Gaussian Lévy process: Simulation and approximation. [24] Øksendal. (2003): Stochastic Integration and Di¤ erential Equations. P. 213-214. [21] Rydberg. 445-466. B. (1999): Lévy Processes and In…nitely Divisible Distributions. New York: Springer. 373-413. 2nd ed. 61-82. Journal of Economic Theory 3..

A method to …t this model to data is given in the companion paper [19]. This was solved by Merton in [20] and [21] for a Black and Scholes market with constant coe¢ cients. Further. The idea of a factor structure is that the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. Hence. This model assumes that the volatility in asset price dynamics be modelled as a weighted sum of non-Gaussian Ornstein-Uhlenbeck (OU) The author would like to thank his supervisors Holger Rootzén and Fred Espen Benth for valuable discussions. [5] solved a similar problem for one stock and a bond in the more general market model of Barndor¤-Nielsen and Shephard [2]. The model in the present paper can obtain strong correlations between the returns for di¤erent stocks without a¤ecting their marginal distributions. and a few Brownian motions that all stocks share. He is also grateful to Michael Patriksson for guiding him to some theorems from optimization theory. as well as for carefully reading through preliminary versions of this paper. Explicit optimal portfolios for n stocks are obtained. and characterizes the dependence by the use a factor structure. with utility drawn from terminal wealth. the number of model parameters does not grow too fast as the number of stocks n grows. It is a modi…cation of [18]. Recently. 1 . We introduce an extension of the stochastic volatility model proposed in [2]. Chalmers University of Technology and Göteborg University.Portfolio optimization and a factor model in a stochastic volatility market Carl Lindberg Department of Mathematical Sciences. and in the Brownian motions in the di¤usion components. Sweden Abstract The aim of this paper is to …nd explicit optimal portfolio strategies for a n-stock stochastic volatility model. This allows us to obtain stable parameter estimates for relatively large n: We use dynamic programming to solve Merton’s optimization problem for power utility. which is of large practical importance. the dependence between stocks lies both in the stochastic volatility. 1 Introduction We consider a version of the problem of optimal allocation of capital between di¤erent assets. This was not possible in [18].

processes of the form dy (t) = y (t) dt + dz (t) . such as stocks from di¤erent branches of industry. The stochastic optimization problem is solved via dynamic programming and the associated Hamilton-Jakobi-Bellman (HJB) integro-di¤erential equation. we deal with this disadvantage. Instead. where z is a subordinator and > 0. the parameter estimates must be stable for relatively large n: This feature is necessary since we want to be able to apply the optimal strategies to portfolios that contain a considerable number of stocks. The object of this paper is to …nd explicit optimal allocation strategies for the factor model described above. volatility clustering. as well as borrowing and short-selling restrictions on the entire portfolio. Further advantages are that it requires little data and gives explicit optimal portfolio strategies. the dependence between stocks is that they share some of OU processes of the volatility. Portfolio optimization with stochastic volatility has been treated in a num- 2 . the marginal distributions have to be very skew. All results are derived under exponential integrability assumptions on the Lévy measures of the subordinators. A multi-stock extension of [5] was considered by [18]. The latter are called factors. This might not …t data. Therefore. but not too dependent. We use a veri…cation theorem to identify the optimal expected utility from terminal wealth as the solution of a second-order integro-di¤erential equation. A factor model has fewer parameters than a standard model. we de…ne the stochastic volatility matrix implicitly by a factor structure. and skewness. and n Brownian motions in the di¤usion components of all n stocks. It retains all the features of the univariate model of [2]. and in the Brownian motions. This problem is an extension of [18] and [5]. The model was primarily intended for stocks that are dependent. In the present paper. In that paper. we do not use the standard approach: An explicit stochastic volatility matrix. It is a primary objective that the number of parameters in our model do not grow too fast as the number of stocks n grows. We allow for the investor to have restrictions on the fractions of wealth held in each stock. such as semi-heavy tails. In other words. We show that this model can obtain strong correlations between the returns of the stocks without a¤ecting their marginal distributions. This idea allows us to capture several of the observed features in …nancial time series. Thus explicit optimal allocation strategies are obtained. and a few Brownian motions that all stocks share. The reason is that the number of factors can be chosen a lot smaller than the number of stocks. The idea of a factor structure is that the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. The disadvantage of the model used in [18] is that to obtain strong correlations between the returns of di¤erent stocks. This means that the dependence between stocks lies both in the stochastic volatility. We consider an investor who wants to maximize her utility from terminal wealth by investing in n stocks and a bond. We then compute the solution to this equation via a Feynman-Kac representation for power utility. which from an applied perspective is a key feature of the solution. This is given the interpretation that the stocks react to the same news.

ber of articles. In [11], [13], and [27], the stochastic volatility depends on a
stochastic factor, correlated to the di¤usion process of the risky asset. The
paper [7] models the stochastic factor as a continuous-time Markov chain with
…nite state-space that is assumed to be independent of the di¤usion process.
The approach to solve the portfolio optimization problems in [7] and [27] is related to the approach in this paper. In [24], partial observation is used to solve
a portfolio problem with a stochastic volatility process driven by a Brownian
motion correlated to the dynamics of the risky asset. The papers [3], [4], and
[9] treat di¤erent portfolio problems when the risky assets are driven by Lévy
processes, and [8] look at portfolio optimization in a market with unobservable Markov-modulated drift process. Further, [14] derive explicit solutions for
log-optimal portfolios in complete markets in terms of the semimartingale characteristics of the price process, and [17] show that there exists a unique solution
to the optimal investment problem for any arbitrage-free model if and only if
the utility function has asymptotic elasticity strictly less than one. We refer
to [1] and [2] for an introduction to the market model of Barndor¤-Nielsen and
Shephard. See [23] for option pricing in this context.
This paper is divided into six sections. In Section 2 we give a rigorous
formulation of the market model. We discuss the dependence structure of the
market, but also alternative models. In Section 3 we set up the control problem.
Section 4 shows that our problem is well de…ned. We prove our veri…cation
theorem in Section 5, and use it in Section 6 to verify the solution we have
obtained.

2
2.1

The model
Model de…nitions

For 0 t T < 1, we assume as given a complete probability space ( ; F; P )
with a …ltration fFt g0 t T satisfying the usual conditions. We take m independent subordinators Zj , and denote their Lévy measures by lj (dz); j = 1; :::; m:
We recall that a subordinator is de…ned to be a Lévy process taking values in
[0; 1) : This implies that its sample paths are increasing. The Lévy measure l
of a subordinator satis…es the condition
Z 1
min(1; z)l(dz) < 1:
0+

We assume that we use the cádlág version of Zj : We introduce now a n stock
extension of the model proposed by Barndor¤-Nielsen and Shephard in [2]. Our
model is a generalization of that in [18].
Consider n+q independent Brownian motions Bi : Denote by Yj ; j = 1; :::; m,
the OU stochastic processes whose dynamics are governed by
dYj (t) =

j Yj (t)dt

3

+ dZj (

j t),

(2.1)

where j > 0 denotes the rate of decay. The unusual timing of Zj is chosen so
that the marginal distribution of Yj will be unchanged regardless of the value
of j : We use the …ltration
fFt g0

t T

:= f (B1 (t) ; :::; Bn+q (t) ; Z1 (

1 t) ; :::; Zm

(

m t))g0 t T

;

to make the OU processes and the Wiener processes simultaneously adapted.
We view the processes Yj ; j = 1; :::; m in our model as news processes associated to certain events, and the jump times of Zj ; j = 1; :::; m as news or
the release of information on the market. The stationary process Yj can be
represented as
Rs
(2.2)
Yj (s) = yj e j (s t) + t e j (s u) dZj ( j u); s t;
where yj := Yj (t) ; and yj has the stationary marginal distribution of the process
and is independent of Zj (s) Zj (t) ; s t: Note in particular that if yj 0; then
Yj (s) > 0 8s 2 [t; T ] ; since Zj is non-decreasing. We set Zj (0) = 0; j = 1; :::; m;
and set y := (y1 ; :::; ym ) : We de…ne i2 as
2

i

(t) :=

t;y
i

2

(s) :=

Pm

j=1

!i;j Yj (s) ;

s 2 [t; T ] ;

where !i;j
0 are weights summing to one for each i: The notation
denotes conditioning on Y (t) : Further, we de…ne
Pm
2
2
t;y
i;k (s) := i;k (s) :=
j=1 i;j;k !i;j Yj (s) ; s 2 [t; T ] ;
where

i;j;k

(2.3)
t;y
i

2

()

2 [0; 1] are weights chosen so that
Pq
2
2
i (s) =
k=0 i;k (s) ; 8s 2 [t; T ] :

We will see now that the i;j;k give the volatilities for each Brownian motion.
De…ne the stocks Si ; i = 1; :::; n; to have the dynamics
!
q
X
2
dSi (t) = Si (t)
dt + i;0 (t) dBi (t) +
i;k (t)dBn+k (t) .
i + i i (t)
k=1

Here i are the constant mean rates of return; and i are skewness parameters.
The Brownian motions Bn+k ; k = 1; :::; q; are referred to as the factors, and we
2
will call i + i i (t) the mean rate of return for stock i at time t: Note that
the choice of volatility process i2 preserves the features of the univariate model
in [2]. These stock price dynamics gives us the stock price processes
Z s
Z s
2
1
(u)
du
+
Si (s) = Si (t) exp
i
i;0 (u) dBi (u) (2.4)
i+
i
2
t
t
!
q Z s
X
+
i;k (u)dBn+k (u) :
k=1

t

4

This stock price model does not have statistically independent increments. It
allows for the increments of the returns Ri (t) := log (Si (t) =Si (0)) ; i = 1; :::; n;
to have semi-heavy tails as well as both volatility clustering and skewness. The
increments of the returns Ri are stationary since
Ri (s)

Ri (t) = log

Si (s)
Si (0)

log

Si (t)
Si (0)

= log

Si (s)
Si (t)

=L Ri (s

t) ;

where " =L " denotes equality in law.
We assume the usual risk-free bond dynamics
dR (t) = rR (t) dt;
with interest rate r > 0.
The idea of this model is to model the dependence between stocks in two
ways. First of all the stocks share the news processes Yj ; j = 1; :::; m: This
implies that the volatilities of di¤erent stocks will be similar. Second, we characterize the dependence further by letting stocks depend on common factors.
We will show below that this allows us to obtain high correlations between the
returns for di¤erent stocks without a¤ecting their marginal distributions. Our
n-stock extension preserves the qualities of the univariate model. In addition,
the number of factors can be chosen to be a lot less than the number of stocks.
This means that much less data is required to estimate the model compared
to if an explicit volatility matrix would have been used. This is an important
feature, since …nancial data can typically not be assumed to be stationary for
long periods of time. The concept of characterizing dependence by factors is
not new. For example, it is used in the Factor-ARCH model (see [10]). It is
also indicated in [2].

2.2

The dependence

In this section we describe brie‡y how to estimate the one-stock model from
data. We then calculate explicit formulas for the covariances and correlations
for the increments of the returns between di¤erent stocks.
We assume that we are observing returns Ri ( ) ; Ri (2 ) Ri ( ) ; :::; Ri (k )
Ri ((k 1) ) ; for stock i = 1; :::; n; where
e.g. is one day, and k + 1 is the
number of trading days in our period of observation. We recall the standard
result that if we take 2 to have a Generalized Inverse Gaussian distribution
(GIG) ; and draw an independent N (0; 1)-distributed random variable "; then
x = + 2 + " will have a Generalized Hyperbolic distribution (GH). This
class is quite ‡exible and contains many of the most frequently used marginal
distributions in …nance. We see by this result that if we choose GIG-marginals
of our continuous time volatility processes i2 , we will obtain approximately
GH-marginals of the increments of the returns Ri ; i = 1; :::; n. The existence
and integrability of Lévy measures lj such that the volatility processes i2 will
have GIG-distributed marginals is not obvious. See [1] and [26, Section 17] for

5

m exp ( m jhj) . :::.j Yj s + !2. R2 (t)) 0 =@ 1 2 1 js e + q X k=1 r E 2 !1. It is an important feature of the model in the present paper that we can estimate the covariances of the increments of the returns between di¤erent stocks from data without a¤ecting the marginal distributions. R2 (u)) Cov (R1 (s) .this theory. It is su¢ cient to show this result for the returns Ri . R2 (v)) . 2: It turns out that Corr (R1 (s) . R2 (u)) + Cov (R1 (t) .k (u) 2.t) 1.j Yj t 1 2 2 1 2 2 2 V ar (Y (0)) !2. Our model generates a non-diagonal covariance matrix for the increments of the returns over the same time period. The weights sum to one for each i = 1. u. v 2 [0. h 2 R: Straightforward calculations show that i (h) = !i. t) #! 1 Pm j=1 r m X j=1 jt +e "Z 1 2 Pm j=1 2 1 1 2 2 2 V ar (Y (0)) !1. for s. where the !i. We estimate the rates of decay j by using the autocorrelation function of the volatility processes i2 : The autocorrelation is de…ned by (h) = Cov ( (h)2 . n.k (u)du 0 j min (s.j V ar (Yj (0)) e j js tj 1+2 2 j min(s. R2 (v)) Cov (R1 (t) .j j e js 1+ js 2 j + !1. t.j j 6 : e jt 1+ 2 j jt .j !2. are the weights from the volatility processes. 2.1 exp ( 1 jhj) + ::: + !i. (0)2 ) V ar ( (0)2 ) .j 0. since the returns have stationary increments. T ] : We now calculate the correlation between the returns Ri . First note that Cov (R1 (s) R1 (t) . This was not possible for strong correlations in the model in [18]. i = 1. i = 1. which is the most frequently used measure of dependence in …nance. R2 (u) R2 (v)) = Cov (R1 (s) .

j E i.j E !1. E [R1 (t)] = 1t + 1 m X 1 2 !1. R2 (t)) = 1 + 1 2 q X k=1 1 2 2 E "Z m X !1. for some constant Yj > 0. 0 j=1 min(s.0 Z (u) du + !# j=1 1 2 s 0 2 1 2 2 Z (u) 0 k=1 1 2 st du + 1 (u) ! t 0 = 2 1 2 0 k=1 + and Itô’s isometry (see [28]) 1.j E 1 q X E k=1 2 1 2 m X Z Z min(s.t) t 2.0 (u) dB2 (u) 0 t Yj (u) du 0 s Yj (u) du 0 Z !1. the independence of the increments of Yj gives 7 .k (u)du 0 s Yi (u) du 0 # Z t Yj (u) du 0 : Similarly. we have that E [Yj (t)] = Yj .k (u)dBn+k 2 q Z X + 2 t 2. for all t 2 R: If we assume that u v.k (u)dBn+k + 1s 2 1 2 + 2t m X 1 + 1 2 m X !2.2 i We have by de…nition of E [R1 (s) R2 (t)] Z s =E 1+ 0 q XZ s + Z 1 1.j=1 "Z (u) dB1 (u) (u) j=1 + 1.j E j=1 Z t Yj (u) du : 0 This gives that Cov (R1 (s) .j Cov Z s Yj (u) du.k (u) 2.k (u)du 0 # Z t Yj (u) du 0 : By stationarity.t) 1.j !2.k (u) 2.i !2.

3 Alternative n-stock extensions We introduced in [18] the notion of news-generated dependence.j Yj (t) dZj ( (t) dBi (t) . We meant by this that the dependence between stocks lies in that they react to the same news. j t) . for i = 1. that V ar (Ri (t)) = m X 2 i 1 2 2 2 !i. Yj (v)) =E Yj (u) =e j (v =e j (v Yj Z v 2 Yj (u) + Yj (u) e j (v u h i 2 u) E Yj (0) e j (v u) 2Yj j (v =E e Yj (v) Yj u) u) s) dZ ( 2 Yj j s) V ar (Yj (0)) : The same calculations for v u shows that j jv Cov (Yj (u) .j Yj t ! . similar to above.j Yj 1 (t)A Si (t) dt + 8 i (t)Si i. for example. Yj (v)) = e uj V ar (Yj (0)) .5) 0 t Cov (Yj (u) . and we get Cov = Z Z s Yj (u) du. In other words. 2: 2. by considering the stock dynamics dSi (t) = i+ 2 i i (t) Si (t) dt+ i (t)Si (t) dBi (t)+Si (t) m X j=1 or 0 dSi (t) = @ i + m X j=1 i. the stocks share news processes. 0 s 0 Z Z t Yj (u) du (2. t) : 2 j Finally. Yj (v)) dudv 0 = V ar (Yj (0)) e js +e jt e j js tj 1+2 j min (s. A natural extension of this model is to distinguish between good and bad news. . we get by Itô’s isometry (see [28]). This can be done.j V ar (Yj (0)) e j=1 jt 1+ jt 2 j + !i.that Cov (Yj (u) .

3). The second model is a generalization of the model in [18].for constants i. where and are constants. c. the …rst model has the disadvantage that is gives discontinuous stock prices.j = i !i. i = 1. and 2 has the marginal distribution of the volatility process of the stock price. The reason is that we suspect that they will be hard to estimate from data. The di¤erence is mainly that we have normalized the jumps in the stocks price process so that they are always less than the stock price itself. bi ] . ai < bi . This means that we can not obtain marginals from the generalized hyperbolic distribution (for example the N IG-distribution) just by choosing generalized inverse Gaussian marginals of the volatility processes. i = 1. In this section we set up the control problem for the market model of 2. Recall that i2 are weighted sums of the news processes. :::. law enforced restrictions on the fraction of wealth held in a speci…c stock. which is the special case where i. and the trading strategies to obtain it.j : The optimal portfolio problem for this model can be solved using analogous techniques as in [18]. :::. the returns of these models can not in general be written (approximately) on the form x = + 2 + ". c < d. given a certain amount of capital. 3 The control problem We want to solve the problem of how a trader is supposed to invest in a stock market to optimize her expected utility from wealth in a deterministic future point in time. for example. Further. for all t s T: 1 +:::+ n This means that we can consider. and c d. d 2 R.j 2 R. :::. unlike the model we have chosen to work with. m. see Equation (2. We use dynamic programming and stochastic control to …nd the maximum expected utility from terminal wealth. and let the constraints take the form i 2 [ai .1. n. 1)-distributed random variable. :::.j that decides whether a news process Yj is associated with good or bad news. The …rst model is a modi…cation of a model proposed in [2]. and set = ( 1 . and measures the trader’s aversion towards risk in that it concretizes how much the trader is willing to risk to obtain a certain level of wealth. bi . We do this to avoid that the stock prices become negative as a result of large negative jumps. For example. 9 n ). We de…ne a optimal value function V as the maximum amount of expected utility that we can obtain from a trading strategy. n: It is the signs of i. " is a N (0.. .s. as well as short-selling and borrowing of capital. These two extensions both seem reasonable from a modelling perspective. n ) : The fraction of wealth held in the risk-free asset is (1 ::: 1 We choose constants ai . We measure utility by a utility function U: This function is chosen by the trader. We then set up the associated Hamilton-Jakobi-Bellman equation for the optimal value function V: Denote by i (t) the fraction of wealth invested in stock i at time t. j = 1. but we have chosen to not pursue them any further. a.

and c d. for all s 2 [t. we want our wealth to be positive.1 For a constant cj > 0 to be speci…ed below.0 (s) dBi (s) i=1 + W (s) q n X X i (s) i. bi ] . We therefore assume that the following holds: Condition 3. The conditions in the de…nition are natural: We impose trading regulations both on each stock and on the entire portfolio. j = 1. Further. i=1 k=1 with initial wealth W (t): We now de…ne our set of admissible controls. i (s) 2 [ai .1 The set At of admissible controls is given by At := f = ( 1 . R (s) 1 W (s) = where i (s) W (s) =Si (s) is the number of shares of stock i which is held at time s: We assume also that the portfolio is self-…nancing in the sense that no capital is entered or withdrawn. and we should always be able to tell how rich we are. a:s:8t s T.We de…ne the wealth process W as (s) W (s) n (s) W (s) S1 (s) + ::: + Sn (s) S1 (s) Sn (s) (1 ::: 1 (s) n (s)) W (s) + R (s) . :::. m: 0+ 10 . i = 1. Z 1 (ecj z 1) lj (dz) < 1. An investment strategy = f (s) : t s T g is said to be admissible if 2 At .1 for t s T as dW (s) = W (s) n X i (s) + i 2 i i (s) n X i r ds (3. T ] : See [16] for a motivating discussion. De…nition 3. :::.1) i=1 + rW (s) ds + W (s) (s) i.1) existsg. n. :::.k (s) dBn+k (s) . and a unique solution W 0 of 1 (s) + ::: + n (s) Equation (3. n ) : i are adapted to fFs gt s T . This can be formulated mathematically as n Z X W (s) = W (t)+ i=1 s t i (u)W (u) Si (u) Z dSi (u)+ s (1 1 (u) 2 (u)) W (u) R(u) t dR (u) . This condition gives the wealth dynamics for the model from Subsection 2. We will need later some exponential integrability conditions on the Lévy measures.

w. and of sublinear growth in the sense that there exists positive constants k and 2 (0. + We will seek to maximize the value function J(t.y [U (W (T ))] . w.3) 2 At . w. y) 2 [0. w. ): The HJB equation associated to our stochastic control problem is ( n X 1 2 2 2 2 w vww 0 = vt + max r wvw + i i+ i i 2 i i. 8 (w. where the notation Et. ) = Et.bi ]. y) 2 [0. y) = J(t. 0. y. y) = U (0) . w. It is assumed to be concave. (t.k w vww 1 h.0 i 2[ai . y) 2 [0. 8 (t. m: The function U is the investor’s utility function.1 holds. w. bounded from below. 1) by R+ and [0. y + z ej ) 9 = . y) 2 Rm+1 . ym ) 2 Rm . T ] 2At and an investment strategy (3. such that V (t. non-decreasing. y)) lj (dz). w.i n 1 k q m X j=1 j Z 1 (v (t. y. y) = U (w) . y.We know from the theory of subordinators that we have Z 1 i h (eaz 1) lj (dz) E eaZj ( j t) = exp j t (3. y) = sup J(t. Denote (0.6) We recall the motivation to this equation for the convenience of the reader. the optimal investment strategy. T ] Rm +: (3. 1) by R0+ .i=1. 1) so that U (w) k(1 + w ) for all w 0: Hence our stochastic control problem is to determine the optimal value function Rm+1 .w. and Yj (t) = yj .:::.5) and the boundary condition V (t. :::. :::. 11 . 0 for (t.4) + rwvw m X j yj vyj j=1 v (t. w.2) 0+ as long as a cj with cj from Condition 3. + V (t. T ) Rm+1 : We observe that we have the terminal condition + V (T.n. w. w. and assume that y = (y1 .y means expectation conditioned by W (t) = w.k 2 i. w. (3. + (3. i=1 c + + 1 2 1 +:::+ n X d X h i h. ). j = 1.

y [V ( . w = 0.1 holds with Pn ! .w. m: We see that if we choose stopping times so that # t.2 Assume Condition 3. Lemma 4.y (W (s)) 2At i w exp ^ n X i=1 12 j i ! rj + r (s ! t) . y) = sup Et. w. then sup lim #t 2At Et.Assume that the Dynamic Programming Principle holds. m.w. w.w.w. W ( ) .y (W (s)) : Lemma 4. 4 Well-de…nedness of the optimal value function In this section we show that the optimal value function is well-de…ned. t = T.y [V (W ( ) .j j where j = 1. w. y) 2 C 1.w.7).7) where we assume that we may change the order of the supremum operator and the limit operator. Then 1 lj (dz)(s t) j We want to show that the optimal value function V (t. + V (t. T ] then for any stopping time T a:s: and t T. Y ( ) . V (t. (3. :::. we get the HJB equation (3.y E exp( j Yj (u)du) exp yj + j exp j t 0+ j for jz j > 0. If we evaluate Equation (3. :::. T ] Rm + with continuous extensions of the derivatives to t = 0. 1) and once continuously di¤erentiable in (t. That is. j = 1.4). w. we see that V (t.1 holds with cj = j = Z s Z 1 j t.y (W (T )) : 2At This observation points out a direction for us to take in h showing ithis property: We want to obtain an upper bound for sup 2At Et. Y ( ))] t V (t.w.1 Assume Condition 3. y) = sup Et. y) = 0.1 [0. )] : 2At The notation C 1. > 0: Then h sup Et. y) on [0. and yj = 0.1 [0. w. We start by stating a lemma that we will need both in this section and in some of the subsequent sections.y [U (W (T ))] 2At h i k 1 + sup Et. T ] Rm+1 means twice continuously di¤erentiable in + w on (0. If we use the sublinear growth condition on the utility function U. if Rm+1 .2. cj = ^ 2 i=1 2 (j i j + ) i. y) is well-de…ned. It is due to [5].2.

k (u) dBn+k (u) t (u)) 2 i + r: s i (u) i. Therefore.0 (u) dBi (u) i. Y (u)) du + t + s i i t i=1 q Z n X X s i.y (W (s)) " Z = w Et.0 (u) dBi (u) t i=1 s 2 i : We want to be able to conclude that the Wiener integrals are parts of a martingale. Y (u)) du + t !# i (u) i. we de…ne n Z X X(s) := exp + s 2 i=1 t q Z s n XX i=1 k=1 t n Z s X 1 2 i 2 (u) i (u) 2 (2 ) ( i=1 t i.k (u) dBn+k (u) 2 i (u)) 2 i ! (u) du : Due to the exponential integrability conditions on Yj from Lemma 4. + i i (u) t i=1 k=1 (u) where (u.k (u) dBn+k (u) .w.0 (u) dBi (u) ! (u) i.w.y exp + q Z n X X i=1 k=1 n Z X s (u.1 we have that " !# n Z T X 2 E exp < 1: i (t) dt i=1 0 13 .0 m 1 X exp @ ^ 2 cj yj + 2 j=1 Z j 1 0+ fexp (cj z) 1 t) A : 1g lj (dz) (s Proof. so we can control their expectations.1) gives that Z W (s) = w exp n Z X s (u. Itô’s formula and Equation (3. Y (u)) = n X i (u) i 2 1 ( 2 r i=1 We now have that h i Et.

:::.y (W (s)) " !# 12 Z s n Z s X 2 w Et.w.y [X (s)] " Z t.Hence the Wiener integrals are all well-de…ned continuous martingales.w. 140]). so we have that Et.y exp 2 (u. Y (u)) du + t ^ n X i=1 2 j i 0 ! n X m Z X i=1 j=1 t s 2 2 ( 2 i (u)) 2 i (u) du t i=1 ! rj + r (s Et.w. since ! ! n X V (t. y) k 1 + w exp ^ j i rj + r (s t) 0 i=1 m X 1 exp @ ^ 2 cj yj + 2 j=1 14 j Z 1 0+ fexp (cj z) 1g lj (dz) (s 11 t) AA : .w. exp 2 Z s (u.y E i " exp n Z X s 2 2 ( 2 i (u)) i (u) du t i=1 max (1. jbn j) =: ^ . jb1 j .y 4exp @ ^ 2 n Z X !# 12 : !# 12 t) s 13 21 2 (j i j + ) !i. we have by Hölder’s inequality and the fact that X is a martingale that h i Et.j Yj (u) duA5 Pn We can now apply Lemma 4.w.j : We have now proved that both the optimal value function of our control problem and the wealth process are well-de…ned. w. Y (u)) du + 2 2 ( i (u)) i (u) du t t i=1 1 2 Et.w. jan j .y (W (s)) " Z s n Z X t.y =w E exp 2 s (u.y [X(s)] = 1: The de…nition of X gives that h i Et. p.w. Y (u)) du + =w E exp t s 2 ( 2 2 (u)) i (u) du X (s) i t i=1 ! 1 2 # : Further.w. :::. Y (u)) du + t But since t.y (u. Then X(s) is a martingale by Novikov’s condition (see [25. ja1 j .1 m times with j = ^ 2 i=1 2 (j i j + ) !i.w.

w. Y (t))) i dt < 1. :::. From now on we assume that Condition 3. w. T ] is a solution of the HJB equation (3. m: cj = i=1 2 (j i j + ) i. Y (t ))j] lj (dz) dt < 1. y) . y)+:::+ n (t.i n 1 k q 15 Rm+1 : + j yj vyj : .6). Let (t. y) in Equation (3. T ] 2 At . a:s:. 8 (t. W (t) . y) 2 [0.4) with terminal condition (3. T ] Rm+1 : + Further. y) = Et. Theorem 5. j = 1. 5 A veri…cation theorem We prove in this section a veri…cation theorem for our control problem. if there exist measurable functions i (t.j j This ensures that the value function is well-de…ned.Note that we also have that U (0) V (t. then that solution is the optimal value function. n. :::. Assume that sup 2A0 Z 0 T Z 1 E [jv (t. and Equation (3. y) = v (t.5) and boundary condition (3. w. w. y) 2 [0. Y (t ) + z ej ) 0+ v (t. w. w. and sup 2A0 Z T 0 h E ( i (t) 2 i (t) W (t) vw (t. and introduce the operator 1 2 2 2 2 i i.0 w vww h i h. W (t) . y) 2 [0.w. W (t) . c d.1 [0.1 holds with Pn ! . w. T ) M v := n X i i + 2 i i Rm+1 and + r wvw + i=1 1 + 2 X X (T ) . w. This theorem says essentially that if we can …nd a solution to our HJB equation. w. n: Then v (t.2. for i = 1. y) 2 C 1. y) .1) admits a unique solution W 0. :::.k i. being the maximizers for the max-operator 1 (t. w. w.4). 8 (t. bi ] .k w 2 vww + rwvw m X j=1 1 h. y) 2 [ai . i = 1.1 (Veri…cation Theorem) Assume that Rm+1 + v (t. then de…nes an optimal investment strategy in feedback form and V (t.y U W Proof. w. y) V (t. since U is non-decreasing.

Hence we have that Z sZ 1 Et. for Borel sets . Y (u ))] lj (dz) du. w. w. y) where L v := M v + m X j=1 j Z 1 (v (t. 69]) tells c us that d [X. W (s) . :::. m. W (u) . This gives us that Et. W (u) . w.y (vt + L v) (u. Y (u ))] Nj ( j du.w. W (u) . Yj ] . w. j = 1. W (u) . the cádlág property of Y. :::. for a semimartingale X: We know from the assumptions that the Itô integrals are martingales and that the integrals with respect to Nj have …nite expectation. w. :::. Y (u )) Nj ( j du. m: We have used that [Yj . Y (u )) du 2At t+ = v (t.w. where we have used the Fubini-Tonelli theorem. W (u) . Y (s)) Z s = v (t. Nj (t. and the fact that. Yj ] is a:e:(path by path) absolutely continuous with respect to c d [Yj . :::. where [ . y + z ej ) 0+ 16 v (t. y) + Et. W (u) .y [v (u.w. dz) . W (u) . W (u) .Itô’s formula gives that v (s.w. the Kunita-Watanabe inequality (see [25.y vt + max L v (u. W (u) . dz) t+ 0+ Z sZ 1 = j Et. Y (u ))g du + + + n Z X t+ s W (u) i=1 t+ q Z s n X X i=1 k=1 m Z s X j=1 t+ i (u) W (u) i. y) + fvt (u. Y (u ) + z ej ) v (u. where the Nj are the Poisson random measure in the Lévy-Khintchine reprec sentation of Zj .y v (u. ] denotes the continuous part of the quadratic covariation. j = 1. p. Y (u )) du t+ Z s v (t.0 (u) i (u ) vw (u. W (u) .w. by c Theorem 26 in [25]. Y (u )) + M v (u. Y (u ) + z ej ) v (u. W (s) . W (u) . m.y [v (s. y)) lj (dz): . w. t+ 0+ for j = 1. y) + Et. Y (u )) dBi (u) i. Y (s))] Z s = v (t. m. Y (u )) dBn+k (u) t+ Z 1 [v (u. ) tlj ( ) is a martingale. W (u) . Yj ] = 0. Further. Y (u ) + z ej ) 0+ v (u.k (u ) vw (u. j = 1.

together i with the assumptions gives that (s. max L v = L v. y) 2 [0. :::. w.j. Further. :::. i. y) = E U W (T ) i . W (s). w. q.y [U (W (T ))] . i = 1. Y (s)) is an admissible control. :::. 1) : We impose also a condition on the weights i. We obtain the solution by constructing a function. y) = V (t.k > 0 . In the next two sections we verify two Feynman-Kac formulas as solutions to our problem. such that all the assumptions in our veri…cation theorem are satis…ed. W (s) . for all 2 At : The …rst conclusion in the theorem now follows by observing that the result holds for t = T and w = 0: We have that i (s. since (t.If we choose s = T and use the terminal condition for v. y) Et.1 [0. This. along with its associated controls.6). and k = 1.k > 0. and we get that h i v (t. Y (s)) are Fs -measurable for t s T.w. since the equality holds for t = T and w = 0 by the terminal and boundary conditions (3. :::. n. The calculations in the …rst part of the theorem hold with equality by letting = .1 We assume that for every h. we need to verify that a well-de…ned function v (t. 2At since is a maximizer. y) 2 C 1. More precisely.w. w. w. T ] 17 Rm+1 + . n. w. we get that v (t.k > 0 for every j = 1. w.k i.5) and (3. i = 1. m: This condition means that each factor has some news processes Yj associated with it. and these Yj are part of the volatility for every stock that is a¤ected by the factor. w. then h.j.y U W (T ) V (t. y) = Et.j. 6 An explicit solution In this section we derive a solution to our optimal control problem when U (w) = w 2 (0. y) : This means that h v (t. for (t. T ] Rm+1 0+ . w.k : Condition 6. y) is assumed to be a measurable function. if h.2.

1).bi ]. n: Our …rst step will be to reduce the HJB equation to a related equation that is simpler to handle.4). :::. W (t) . c 1 (1 2 + m X j=1 ) 1 +:::+ n X X d ( n X i j 1 + 2 i i r i=1 i.:::. w. 6. y: It is obvious that v is continuous in w: If we insert the function v in the HJB equation (3.6).0 (6. 0+ with the terminal condition h (T. we have replaced the problem of …nding a solution to the HJB equation (3.2) In other words. t) j=1 h (t. Our next step is to …nd a well-de…ned function h that satis…es Equation (6. W (t) . (t.n. w. 6. that sup 2A0 Z Z T 1 0+ 0 E [jv (t.4) by the presumably simpler problem of …nding a solution to Equation (6.4) is of the form v (t. and that sup 2A0 Z T 0 h E f i (t) i (t) W (t) vw (t.k + r h i h. for i = 1.i=1. T ] where h is some function of t.2 A solution to the reduced HJB equation In this subsection we de…ne a Feynman-Kac formula and show that it is wellde…ned and continuously di¤erentiable. we get the associated equation 0 = ht (t.k 1 h. y) 2 [0.1) h (t.4) with terminal condition (3. w. W (t) . Y (t))g 2 i dt < 1.i n 1 k q Z i (h (t.5) and boundary condition (3. y) .1). Y (t ))j] lj (dz) dt < 1. 1 (1 2 ) 2 2 i i. y) + max i 2[ai . y + z ej ) 9 = . We show also that it solves the reduced 18 . Y (t ) + z ej ) v (t. w. 8y 2 Rm +: (6. y) m X j y j hy j (y. y) = w Rm+1 + h (t.1 Reduction of the HJB equation We conjecture that the solution to the HJB equation (3. y) = 1. y)) lj (dz).is a solution to the HJB equation (3.

(t.HJB equation (6. . y) exp @kt + ^ . m: Then 0 1 m X n X j j ! y i i.3) as our candidate solution.0 (6. y) := h (T t.j . y) 2 [0. T ] (6.i=1. y) = Ey exp Z t (Y y (s)) ds : 0 The only di¤erence between the two functions is the direction of the time variable t: We will now show that g is well-de…ned. j = 1.bi ].y exp t i (Y y (s)) ds Rm +.j j A g (t. y) = 1: The function : Rm + ! R is de…ned as (y) = max i 2[ai .1). c 1 +:::+ n d T (y) 1 2 T Q (y) where Q is a positive de…nite matrix for every y 2 Rm + . :::.n. .1 Assume Condition 3.i n 1 k q 9 = . y) 2 [0. we prefer to re-write the function h on a form that is simpler for us to deal with.:::. Note that this function satis…es the terminal condition (6. . j j=1 i=1 for (t.y exp t t (Y y (s)) ds 0 Rm + : We de…ne now g (t.:::.i=1. y) 2 [0.bi ]. c 1 +:::+ n 1 (1 2 d ) ( n X i i 2 i i + i=1 X X 1 (1 2 r h i h. with vector notation. and 19 as +r i = i (6. we have that " !# " !# Z Z T h (t.5) + 2 i i r . By the stationarity of Y. and some constant k > 0: Proof. T ] Rm + . where y = Y (0) .n. We set h RT h (t.2) since h (T. y) = Et.1 holds with cj = ^j i=1 j i j !i. Pn Lemma 6. It is straightforward to see that we can write (y) = max i 2[ai .k i.k 1 h. T ] T (Y y (s)) ds = E0. y) = Et.4) + r: For technical reasons.y exp for (t. ) 2 2 i i.

2 Assume Condition 3. ja1 j . :::.j z 1 lj (dz) t : j The result follows from Condition 3. y) 2 t (Y y (s)) ds : 0 We have then that @F (t. m: Then g 2 C 0.1. 20 .j t yj g (t.14]. :::. De…ne the compact intervals An = [0. see for example [6. T ] Rm + : Proof. j = . jb1 j . y) = E Z exp " y t (Y y (s)) ds 0 Z exp E t e 2 Ey 4 t +^ 0 m Y n X i=1 exp ^ j=1 n X i=1 j ij !# 2 i (s) ds Z j i j !i. ::: .1 holds with cj = 1. Remark 4. n = 1. and by Lemma 4. Let (t. :::. ^ Lemma 6. m: We know from Equation (6. T ] An and set F (t. y) e exp + j Z j=1 1 j ^ exp 0+ Pn i=1 j i j !i. We show now that g is continuously di¤erentiable in y. :::. j = 1. is a quadratic program with linear constraints.1 [0.1 that Pn q Y ^ i=1 j i j !i. we have y g (t. We can now apply Danskin’s theorem. It is well known from optimization theory that this problem has a unique solution. y) = @yj @ @yj Z t Rt (Y y (s)) ds e 0 0 (Y y (s))ds .13.5) that for …xed y. :::. jan j . Theorem 2.j .1]. n j Pn i=1 j i j !i. Since y 2 An . 2. (6. see for example [22. jbn j) : Hence. y) = exp Z 1 n. to conclude that is continuously di¤erentiable. Theorem 4. we have by the independence of the Yj .6) i=1 where we recall that ^ = max (1.It follows by the positive de…niteness of Q that we can …nd a constant > 0 such that n X j (y)j +^ j i j i2 . m. for each j = 1.j t 0 !3 Yj (s) ds 5 : Further.

y) @g(t. We conclude the proof by observing that di¤erentiability and continuity are local notions. m: @yj @yj Note that j@F (t. Proposition 6.j.j j i=1 n X j i j !i. From [12. we have that j@F (t.j i. Theorem 2.1 holds with cj = " + 2^ j + n X i=1 (1 2 j i j !i.j 0 ) ^2 @ X X p 1 h.1.1 Assume that Condition 3.y) = E . y) = @yj Z t 0 yj 0 0 k1 tek2 t @ (Y y (s)) e m Y exp j=1 ek3 t+k4 Pm j=1 yj But exp ^ 0 @ js ^ n X i=1 m Y exp j=1 j i j !i.j i=1 Rt ds e Zj ( ! !1 Zj ( j t) A : j t) is integrable by Condition 3.j. t Yj (u)du j yj + Zj ( j t) 0 we have that @F (t.27(b)].i n 1 k q h. ) belongs to the domain of the in…nitesimal generator 21 .27(b)] to show that g (t. Hence. y) 2 [0. and we can apply [12.we can also conclude that ry and the fact that Z is bounded on An . :::. for j = 1.k !i. y)] is di¤erentiable in y on An .j + n X i=1 i.0 !i. :::.j 11 AA . Theorem 2.1). Theorem 2. m: Then g (t.j. y) = E [F (t. j = 1.j ^ (Y y (s))ds 0 Z 0 j t !1 Yj (s) ds A n X j i j !i. and that h i @F (t.27(a)] that @g (t. y) =@yj j is uniformly bounded in y on An .k !h. T ] An . 8y 2 An . y) =@yj j is continuous in t and y: We get now by using [12. y) =@yj is continuous in (t. and that limn!1 An = Rm +: We show now that g is a classical solution to the related forward problem of Equation (6.

j.k !h. Y (s))j lj (dz) ds < 1.k !i.1 holds with ^ cj = j + n X i=1 (1 2 j i j !i. y) 2 [0. y + z ej ) g (t. Rm + : In addition. T ] y 2 Rm +: g (0.8) We have also that gt is continuous.k !i. is bounded and that .3 Assume that Condition 3. We know that x is strictly concave and increasing.j i.k ) @ p @yj h. y + z ej ) g (t. :::. w.j i. with initial value for (t. the integral Z 1 jg (t.j.7) j=1 + m X j Z 1 (g (t. y)j lj (dz) 0+ is continuous in (t. m: Then "Z Z m T 1 X E jg (t. (6.k !h. T ] Rm +: p Proof.j 0 ) ^2 @ X X p h. Lemma 6.1 [0. y) + m X (y) g (t.j. T ] Rm + : We will need some integrability and continuity conditions on the Lévy measure integrals in order to prove this theorem. i. y) 2 [0. k: We have then by inspection that ry 22 . t) (6.j yj for every h. so that g 2 C 1.of Y and 0= gt (t. T ] i. 0+ j=1 Rm + . y) 2 [0. w. y) = 1.k i. This gives that @ ( @yj h.i n 1 k q for j = 1. j. y) j yj gyj (y. y)) lj (dz) .j.j i=1 1 h. # g (t.j j=1 0 0+ for all (t.0 !i. Y (s) + z ej ) + n X 11 AA .j.

We have then that E "Z T 0 Z Z 1 0+ jg (t. We are now ready to prove Proposition 6. with continuous extensions to t = 0 and t = T. m. are adapted. This gives that if g solves Equation (6. Y (s))j lj (dz) ds m X n X j i j !i.1. the expectation part is …nite as well. which concludes the …rst part of the proof. Since y ! g (t. Proof Proposition 6. :::. 0+ as long as a cj with cj from Condition 3. 6. since y Yj j (s) yj + Zj ( j s) : The integral with respect to the Lévy measure is …nite by assumption. We observe that the assumptions in Lemmas 6. We recall from Equation (3. :::. We know that Yj j = 1.1. This gives that jg (t.1. cádlág. which is due to Condition 6.0 yj cj j= . T ) Rm + . y) 2 (0. y + z ej ) Z y E exp Z Ey exp 0 g (t. Itô’s formula (see [25. and 6. j # 13 s) j A5 ds 0+ for k2 > 0. Theorem 33]) gives that the 23 .1.1. y)j Z t Y y+z ej (s) ds exp (Y y (s)) + cj exp @k1 t + ^ (Y y (s)) ds 0 0 t 0 t js j ze 1 m X n X j i j !i. and have paths of …nite variation on compacts since Yj (t) Zj ( j t) : We have from [25.7) then gt is continuous in (t.1 holds.3 are satis…ed. Y (s) + z ej ) T ek1 t+k2 0 Z 1 Pm j=1 (exp (cj z) yj 2 0 E 4exp @ ^ g (t.2) that Z 1 h i aZj ( j t) E e = exp j t (eaz 1) lj (dz) . Hence.2. m. The second part can be proved by similar techniques.j yj A ds exp Z t (Y y (s)) ds 0 (exp (cj z) 1) . y) is continuously di¤erentiable. are quadratic pure jump semimartingales. j j=1 i=1 for k1 > 0 by Lemma 6. Theorem 26] that Yj j = 1.j Zj ( j=1 i=1 1) lj (dz) .

y + z ej ) g (t. j = 1. m. since E [g (t. Y (u)) du s 0 j=1 Z s Z 1 m X j + E g (t. Theorem 3. Theorem 2. Leb) . y) lj (dz) 0+ is continuous. Y (u )) lj (dz) du : We see from Lemma 6.35]) that g (t.mapping s ! g (t. 8t 2 [0. Y (s))] < 1 by Lemma 6. Y (s)) g (t. Y (u)) du 0 (g (t. y) = m X j=1 j yj gyj (t. Y (u ) + z ej ) s 0 0+ j=1 g (t. we get E [g (t. ) is in the domain of the in…nitesimal generator of Y: We denote the in…nitesimal generator by G. that y 7! g (t.1 and Yj (t) yj + Zj ( j t) . which gives that Gg (t. Y (s))] h h Rt ii Y y (s) (u))du = E E e 0 (Y 24 . Y (u ) + z ej ) g (t. m: If we take expectation on both sides and apply Fubini’s theorem (see [12. Y (u)) 2 L1 ([0. Y (s)) is a semimartingale with dynamics g (t. are cádlág. dz) . and that Z 1 g (t. T ] : Hence. 0+ where Nj is the Poisson random measure in the Lévy-Khintchine representation of Zj . Y (u ))) Nj ( j du. y)] s Z s m X j = E Yj (u) gyj (t. y) is continuously di¤erentiable. if we note that the Yj j = 1. y) + m X j=1 + m Z X j=1 s 0 Z 1 j Z s Yj (u) gyj (t. s] .3 that E Yj (u) gyj (t. :::. y + z ej ) g (t.37]). Y (s)) = g (t. by letting s # 0 we get by the Fundamental Theorem of Calculus for Lebesque Integrals (see [12. :::. y)) lj (dz) : 0+ The Markov property of Y together with the law of total expectation yields E [g (t. y) + m X j=1 j Z 1 (g (t.

s] + = sup (Y y (v)) e v2[0. We de…ne the function f (s) = e Rs (Y y (u))du 0 : From the mean value theorem and the linear growth assumption on that we get 1 jf (s) f (0)j s 1 sup f 0 (v) s s v2[0.s] + ^ n X i=1 Rv 0 j ij (Y y (u))du y i 2 (v) 25 ! e RT 0 +^ Pn y 2 i=1 j i j i (u) du : . y) : + s For simplicity of calculations. Y (s)) g (t.s] sup v2[0. (Y y (u+s))du (Y y (u))du (Y y (u))du i jFs Rs e 0 ii (Y y (u))du i : E [g (t. o 1 as s # 0: We need to show now that we can interchange limit and integration. y)] s i Rt y 1 h R t+s (Y y (u))du R s (Y y (u))du e 0 (Y (u))du e 0 = E e0 s i R t+s 1 h R t+s (Y y (u))du R s (Y y (u))du (Y y (u))du = E e0 e0 e 0 s h Rt io y 1 n h R t+s (Y y (u))du i + E e0 E e 0 (Y (u))du s n Rs o R t+s y (Y y (u))du 1 =E e 0 e 0 (Y (u))du 1 s g (t + s.h h Rt =E E e 0 h R t+s =E e s h R t+s =E e 0 Thus. T ] : We can verify that R t+s e 0 ! (Y y (u))du 1 Rt (y) e s 0 n e Rs (Y y (u))du 0 (Y y (u))du . y) g (t. we assume that t + s 2 [0.

(6.2.1 [0.6). Y (t ))j] lj (dz) dt < 1. Lemma 4. y) + gt (t. n X sup v2[0.3 Veri…cation of the explicit solution We show in this subsection that Z TZ 1 sup E [jv (t. W (t) .Since each Zj is a non-decreasing process. 6.1 and the Dominated Convergence Theorem that R t+s E e = 0 (Y y (u))du 1 s (y) g (t. w.T ] n X i=1 i=1 j ij 2 i j ij y i + ! 2 (v) n X m X j i j !i.9) . All that is left for us to do is to show that the remaining conditions in the veri…cation theorem 5.5) and boundary condition (3.j Zj ( y 2 i (u) du 1 o 1 A jT ) : Note that there exists a positive constant k" such that z k" e"z . W (t) .j Zj ( i=1 j=1 jT ) . for all z 0: We have by the independence of the Yj . Y (t ) + z ej ) 2A0 0 0+ 26 v (t. y) . which concludes the proof. y) = (y) g (t. y) : n e Rs 0 (Y y (u))du o 1 We can show analogously that gt exists. T ] Rm+1 + that is a solution to the HJB equation (3. which implies R t+s e (Y y (u))du 1 0 0 @ s n X j ij i=1 2 e RT 0 2 i + +^ n e Rs (Y y (u))du 0 n X m X i=1 j=1 Pn i=1 j i j j i j !i. y) 2 C 1. We have then that Gg (t.4) with terminal condition (3. We have found a function v (t.1 hold.

0 !i. Y (t ) + z ej ) 11 AA .j + n X i. 2 A0 : Proof.10) for i = 1. Note that the conditions we verify are slightly modi…ed versions of Equations (6.j i=1 1 h.j 0 ) ^2 @ X X p h.j Zj ( j=1 i=1 27 j 13 j t) A5 dt: . Hence we have solved our problem of …nding the optimal value function. Since ry is bounded and 0yj cj j = . (6. :::.i n 1 k q for j = 1.4 Assume that Condition 3. h (t.1 holds with 4^ 2 cj = j n X i=1 + (1 2 (j i j + 2 ) !i. W (t) .k !h.and that sup 2A0 Z T 0 h E f i (t) i (t) W (t) vw (t. since our conjectured solution v is of the form v (t. Y (t ))j] lj (dz) dt 0 0+ Z 1 (exp (cj z) 1) lj (dz) 0+ Z T ek1 t+k2 0 2 Pm j=1 yj 0 E 4W (t) exp @ ^ m X n X j i j !i. m: Then sup 2A0 for all Z 0 T Z 1 0+ E [W (t) jh (t. y) = w h (t.j i. y) : Lemma 6. we have analogous to Lemma 6. :::.10).1 are satis…ed for our candidate solution v.1 that Z TZ 1 E [W (t) jh (t. w. Y (t))g 2 i dt < 1.9) and (6. all the conditions from the veri…cation theorem 5.k !i.j. n: Once this is done. Y (t ))j] lj (dz) dt < 1.j.j. Y (t ) + z ej ) h (t.

If we apply Hölder’s inequality with p = q = 2 to the expectation part, we get
2
13
0
m X
n
X
j
j
!
Z
(
t)
i
i;j
j
j
A5
E 4W (t) exp @ ^
j

j=1 i=1

h

2

= E W (t)

i1=2

2

0

E 4exp @2 ^

m X
n
X
j i j !i;j Zj (
j

j=1 i=1

131=2
t)
j A5

:

We recall from Equation (3.2) and Lemma 4.2 that both expectations are …nite.
Lemma 6.5 Assume that Condition 3.1 holds with
8^ 2

cj =
j

n
X
i=1

+

(j i j + 4 ) !i;j
2

(1

0

X

)^ @

2

(6.11)

X p

h;j;k !h;j i;j;k !i;j

+

n
X

i;j;0 !i;j

i=1

1 h;i n 1 k q

for j = 1; :::; m: Then
Z

T

0

h
E f

i

(t)

2

i

(t) W (t) h (t; Y (t))g

i

11

AA ;

dt < 1:

Proof. It follows from the de…nitions of h and g that they have the same
growth. We have then by Lemma 6.1 that
Z

0

T

h
E f

i

^ 2 k" ekT

(t)
Z

i
T

(t) W (t) h (t; Y (t))g
2
0
2

E 4W (t)

0

2

i

dt

13
m X
n
X
j i j !i;j Yj A5
"
dt;
exp @ 2 ^ +
2 j=1 i=1
j

Pm Pn j j! Yj
;
since we can …nd positive constants k" such that i2 k" exp 2" j=1 i=1 i i;j
j
for every i = 1; :::; n: We can now apply Hölder’s inequality with p = q = 2.
This gives that
2
0
13
m X
n
X
"
j
j
!
Y
i
i;j
j
2
A5
E 4W (t) exp @ 2 ^ +
2 j=1 i=1
j
h

4

E W (t)

i1=2

2

0

E 4exp @(4 ^ + ")
28

131=2
m X
n
X
j i j !i;j Yj A5
j=1 i=1

j

;

which are both …nite by the assumptions, for " su¢ ciently small.
To conclude, we assume that Condition 3.1 holds with cj as in Equation
(6.11). This ensures that all our results are valid for " su¢ ciently small. Further,
we note that there exist measurable functions i (y) that are the maximizers for
the max-operator in Equation (3.4). These optimal allocation strategies are the
solution of the quadratic program of Equation (6.4). This gives that Equation
(3.1) admits a unique positive solution W by [25, Ch. 2, Thm. 37]. Hence, all
the assumptions in Theorem 5.1 are satis…ed, and we have solved the problem.

References
[1] Barndor¤-Nielsen, O. E., Shephard, N. (2001): Modelling by Lévy processes
for …nancial econometric, in: O. E. Barndor¤-Nielsen, T. Mikosch and
S. Resnick (Eds.), Lévy Processes - Theory and Applications, Boston:
Birkhäuser, 283-318.
[2] Barndor¤-Nielsen, O. E., Shephard, N. (2001): Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in …nancial economics,
Journal of the Royal Statistical Society: Series B 63 (with discussion), 167241.
[3] Benth, F. E., Karlsen, K. H., Reikvam, K. (2001a): Optimal portfolio selection with consumption and non-linear integro-di¤erential equations with
gradient constraint: A viscosity solution approach, Finance and Stochastics
5, 275-303.
[4] Benth, F. E., Karlsen, K. H., Reikvam, K. (2001b): Optimal portfolio management rules in a non-Gaussian market with durability and intertemporal
substitution, Finance and Stochastics 5, 447-467.
[5] Benth, F. E., Karlsen, K. H., Reikvam, K. (2003): Merton’s portfolio optimization problem in a Black and Scholes market with non-Gaussian stochastic volatility of Ornstein-Uhlenbeck type, Mathematical Finance 13(2),
215-244.
[6] Bonnans, J. F., Shapiro, A. (2000): Perturbation Analysis of Optimization
Problems, New York: Springer.
[7] Bäuerle, N., Rieder, U. (2004): Portfolio optimization with Markovmodulated stock prices and interest rates, IEEE Transactions on Automatic
Control, Vol. 49, No. 3, 442-447.
[8] Bäuerle, N., Rieder, U. (2005): Portfolio optimization with unobservable
Markov-modulated drift process, to appear in Journal of Applied Probability.
[9] Emmer, S., Klüppelberg, C. (2004): Optimal portfolios when stock prices
follow an exponential Lévy process, Finance and Stochastics 8, 17-44.
29

[10] Engle, R. F., Ng, V. K., Rothschild, M. (1990): Asset pricing with a factorARCH covariance structure, Journal of Econometrics 45, 213-237.
[11] Fleming, W. H., Hernández-Hernández, D. (2003): An optimal consumption model with stochastic volatility, Finance and Stochastics 7, 245-262.
[12] Folland, G. B. (1999): Real Analysis, New York: John Wiley & Sons, Inc.
[13] Fouque, J.-P., Papanicolaou, G., Sircar, K. R. (2000): Derivatives in Financial Markets with Stochastic Volatility, Cambridge: Cambridge University
Press.
[14] Goll, T., Kallsen, J. (2003): A complete explicit solution to the log-optimal
portfolio problem, Annals of Applied Probability 13, 774-799.
[15] Kabanov, Y., Klüppelberg, C. (2004): A geometric approach to portfolio
optimization in models with transaction costs, Finance and Stochastics 8,
207-227.
[16] Korn, R. (1997): Optimal Portfolios, Singapore: World Scienti…c Publishing Co. Pte. Ltd.
[17] Kramkov, D., Schachermayer, W. (1999): The asymptotic elasticity of utility functions and optimal investment in incomplete markets, Annals of Applied Probability 9, 904-950.
[18] Lindberg, C. (2005): News-generated dependence and optimal portfolios
for n stocks in a market of Barndor¤-Nielsen and Shephard type, to appear
in Mathematical Finance.
[19] Lindberg, C (2005): The estimation of a stochastic volatility model based
in the number of trades, submitted.
[20] Merton, R. (1969): Lifetime portfolio selection under uncertainty: The
continuous time case, Review of Economics and Statistics 51, 247-257.
[21] Merton, R. (1971): Optimum consumption and portfolio rules in a continuous time model, Journal of Economic Theory 3, 373-413; Erratum (1973)
6, 213-214.
[22] Nash, S. G., Sofer, A. (1996): Linear and Nonlinear Programming, Singapore: The MacGraw-Hill Companies, Inc.
[23] Nicolato, E., Venardos, E. (2003): Option pricing in stochastic volatility
models of the Ornstein-Uhlenbeck type, Mathematical Finance 13, 445-466.
[24] Pham, H., Quenez, M.-C. (2001): Optimal portfolio in partially observed
stochastic volatility models, Annals of Applied Probability 11, 210-238.
[25] Protter, P. (2003): Stochastic Integration and Di¤ erential Equations, 2nd
ed. New York: Springer.
30

(1998): Stochastic Di¤ erential Equations. B. (1999): Lévy Processes and In…nitely Divisible Distributions. 61-82. Berlin: Springer. Finance and Stochastics 5. [27] Zariphopoulou.[26] Sato. 31 . T. [28] Øksendal. (2001): A solution approach to valuation with unhedgeable risks. Cambridge: Cambridge University Press. K. 5th ed.

we get an economical interpretation of the non-Gaussian Ornstein-Uhlenbeck processes that de…ne the stochastic volatility in the model. 1 . Further. and to Henrik Röhs at SIX . 1) : It is a key theoretical feature of the framework in [5] that the centered returns divided by the volatility are also i:i:d: and N (0. Many alternatives that seek to overcome these ‡aws have been proposed. 1) : This suggests that we identify the daily number of trades with the volatility. our approach is easier to implement than the quadratic variation method. We also consider the factor model in [13]. The results indicate a good model …t. which is an n-stock extension of the model in [5]. 1 Introduction It is a well-known empirical fact that many characteristics of stock price data are not captured by the classical Black and Scholes model. we verify that we can divide the centered returns with a constant times the number of trades in a trading day to obtain normalized returns that are i:i:d: and N (0. but also stable parameter estimates. and model the number of trades with the model in [5]. Hence. He is also grateful to Holger Rootzén for carefully reading through preliminary versions of this paper.Stockholm Information Exchange for supplying the time series. Chalmers University of Technology and Göteborg University.The estimation of a stochastic volatility model based on the number of trades Carl Lindberg Department of Mathematical Sciences. An illustrative statistical analysis is performed on data from the OMX Stockholmsbörsen. and requires much less data. Sweden Abstract This paper presents statistical methods for …tting the stochastic volatility model of Barndor¤-Nielsen and Shephard [5] to data. We argue that the straightforward approach to estimating the Generalized Hyperbolic (GH) distribution from …nancial return data is inappropriate because the GH-distribution is "almost" overparameterized. A common approach is to assume that the volatility is stochastic. To overcome this problem. Barndor¤-Nielsen and Shephard The author would like to thank Holger Rootzén and Fred Espen Benth for valuable discussions.

A drawback with this volatility model has been the di¢ culty to estimate the parameters of the model from data. The paper [13] makes an attempt to remedy this. In [13]. but not too dependent. In addition. with n Brownian motions in the di¤usion components of all n stocks. the model requires little data since no volatility matrix has to be estimated. in reality the model does not hold on the microscale. [6]. where z is a subordinator and > 0: It turns out that this framework allows us to capture several of the observed features in …nancial time series. which in turn was an extension of [5]. This model retains the features of the univariate model of [5]. the stocks are assumed to share some of the OU processes of the volatility. It is shown that this dependence generates covariance between the returns of di¤erent stocks. Further. None of these features are present in the mathematical model. we need the marginal distributions to be very skew. We therefore propose a di¤erent strategy that only uses daily data. and skewness. see for example [3]. it is analytically tractable. This paper builds on a model proposed in [13]. For example. This makes it in theory possible to recover the volatility process from observed stock prices. and that we can obtain strong correlations without a¤ecting the marginal distributions of the returns of the stocks. In this paper we develop statistical methods for estimating the models in [5] 2 . To characterize dependence by factors is common in discrete time …nance. [12]. This might not …t data. The disadvantage is that to obtain strong correlations between the returns Ri (t) = log (Si (t) =Si (t 1)) of di¤erent stocks Si . We have chosen to not use a full explicit stochastic volatility matrix. In addition. see [4]. it is hard to implement in a statistically sound way due to peculiarities in intraday data. The model is primarily intended for stocks that are dependent. Perhaps the most intuitive approach to do this is to analyze the quadratic variation of the stock price process. and even if is only regarded as an approximation this approach still requires very much data. the stochastic volatility matrix is de…ned implicitly by a factor structure. [13]. The model in [13] is a further development of the n-stock stochastic volatility model in [12]. However. This is given the interpretation that the stocks react to the same news. such as stocks from di¤erent branches of industry. the stock market is closed at night. and [15]. The idea of a factor structure is that the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. see for example [7] and [10]. such as semi-heavy tails. volatility clustering. [14]. since the statistical estimation of such models quickly becomes infeasible as the number of assets grows. and a few Brownian motions that all stocks share. In [12]. and there is more intense trading on certain hours of the day. and it gives explicit optimal portfolio strategies.[5] model the stochastic volatility in asset price dynamics as a weighted sum of non-Gaussian Ornstein-Uhlenbeck processes of the form dy = y (t) dt + dz (t) .

we also discuss our data set. Further. and requires much less data. 1) . which implies that its sample paths are increasing.1) where j > 0 denotes the rate of decay. 1) : We argue that one can not estimate the Generalized Hyperbolic (GH) distribution directly from …nancial return data. j = 1. F. we assume as given a complete probability space ( . due to that the GHdistribution is "almost" overparameterized. 1) : This suggests that we identify the daily number of trades with the volatility. and the GH-distribution. The section also contains a discussion. It is a generalization of that in [12]. 1) : It is an important feature of the stochastic volatility framework in [5] that the centered returns divided by the volatility are also i:i:d: and N (0. It also implies an economical interpretation of the daily average stochastic volatility. We then introduce the discrete time analogue in Section 3. and denote the Lévy measures of Zj by lj (dz). t for " 2 N (0. We assume that we use the cádlág version of Zj . it is easier to implement than the quadratic variation method. Bn+q (t) . We are inspired by [1] to verify that we can divide the centered returns by a constant times the number of trades in a trading day to get a sample that is i:i:d: and N (0. The results indicate a good model …t. Here. the stationarity assumptions. m. Z1 ( 3 1 t) . :::. and it hints that we can view the continuous time volatility as the intensity with which the trades arrive. the OU stochastic processes whose dynamics are governed by dYj (t) = j Yj (t)dt + dZj ( j t). j = 1.and [13] from data. :::. In Section 2 we present the continuous time model. Our approach gives more stable parameter estimates than if we analyzed only the marginal distribution of the returns directly with the standard maximum likelihood approach. The unusual timing of Zj is chosen so that the marginal distribution of Yj will be unchanged regardless of the value of j : The …ltration Ft = (B1 (t) . m: We present now a n-stock extension of the model proposed by Barndor¤Nielsen and Shephard in [5]. The data analysis is presented in Section 4. A statistical analysis is performed on data from the OMX Stockholmsbörsen. (2. :::. Take n + q independent Brownian motions Bi : Denote by Yj . P ) with a …ltration fFt g0 t<1 satisfying the usual conditions. Introduce m independent subordinators Zj : Recall that a subordinator is de…ned to be a Lévy process that takes values in [0. The models are …rst discretized under the assumption that Z t (s) dB (s) (t) ". and model the number of trades within the model in [5]. Zm ( m t)) . 2 The continuous time model For 0 t < 1. :::.

(2.0 (t) dBi (t) + i + i i (t) i.j Yj (t) . :::. We set Zj (0) = 0.j where i.5) Si (t) = Si (0) exp i (s) i+ i 2 0 ! Z t q Z t X + i. t 0. . j = 1.k 0 are weights summing to one for each i: Further. i. 4 i = 1. The stationary process Yj can be represented as R0 Yj (t) = 1 exp (s) dZj ( j t + s) . The Brownian motions Bn+k . :::. t 0: It can also be written as Yj (t) = yj e jt + Rt 0 j (t e s) dZj ( j s).is used to make the OU processes and the Wiener processes simultaneously adapted. 1] are chosen so that Pq 2 i (t) = k=0 2 i. we are back in the univariate model of [5].3) 0. :::. :::. m as news or the release of information on the market. as news processes associated to certain events. i = 1. :::. and i are skewness parameters. and yj has the stationary marginal distribution of the process and is independent of Zj (t) Zj (0) . n. are referred to as the factors. t where !i. t 2 [0. t 0: Note in particular that if yj 0. since Zj is non-decreasing.k (t)dBn+k (t) .k (s)dBn+k (s) : 0 k=1 0 This stock price model allows for the increments of the returns Ric (t) := log (Si (t) =Si (0)) . and the jump times of Zj .j.4) 2 The processes i.j. q.k (t) := j=1 i. m. n. k=1 Here i are the constant mean rates of return. Pm 2 0. then Yj (t) > 0 8t 0.k !i. :::. i (t) := j=1 !i. and we 2 will call i + i i (t) the mean rate of return for stock i at time t: Note that if we choose n = 1. The stock price dynamics gives us the stock price processes Z t 2 1 ds (2. k = 1. and write y := (y1 . We follow the interpretations in [12] and [13]. m. :::. and view the processes Yj .k are the volatilities for factor k for each stock i: De…ne the stocks Si . j = 1. t 0: (2. j = 1. to have the dynamics ! q X 2 dSi (t) = Si (t) dt + i.k (t) . ym ) : The volatility processes i2 are de…ned as Pm 2 (2.j Yj (t) .0 (s) dBi (s) + i.2) where yj := Yj (0) .

for stock i = 1. with Rc de…ned by Equations (2. We assume from now on that the time units are chosen so that = 1: The approximation to assume that Z t (s) dB (s) (t) ".5) and (2. It was shown in [13] that this allows us to obtain strong correlations between the returns for di¤erent stocks without a¤ecting their marginal distributions. and a few Brownian motions that all stocks share. the increments of the returns Ric are stationary since Ric (s) Ric (t) = log Si (s) Si (0) log Si (t) Si (0) = log Si (s) Si (t) =L Ric (s t) . The underlying idea of this model is to capture the dependence between stocks in two ways: First. 3 The discrete time model and the data In this section we introduce a discrete time version of our model. The reason is that the number of parameters to be estimated is smaller. In addition. :::.1 The discrete time model Assume that we observe returns Ric ( ) . :::.6) where " =L " denotes equality in law. the obvious approach to the problem of estimating the model parameters is discussed. m. (2. Here is one time unit.1) t with " 2 N (0. the number of factors can be chosen to be a lot less than the number of stocks. The model has also the feature of preserving the qualities of the univariate model. by letting the stocks share the news processes Yj . In addition. This makes it possible to consider more assets than if we modelled the covariance structure by an explicit volatility matrix. This motivates the following approximate discrete 5 . j = 1. Second. :::. the di¤usion components of the stocks contain one Brownian motion that is unique for each stock. with n Brownian motions in the di¤usion components of all n stocks.6). (3. Further. Finally we give an alternative approach to analyzing the data. see for example [7] and [10]. n.to have semi-heavy tails and for both volatility clustering and skewness. Ric (2 ) Ric ( ) . and present the data that we use in our analysis. 1) may be reasonable unless some j are large so that the volatility processes will be volatile. 3. we allow for the volatilities of di¤erent stocks to be similar. and d + 1 is the number of consecutive observations. Ric (d ) Ric ((d 1) ) . The idea to characterize dependence by factors is of course not new.

3) k=1 where we have used that q X '2i.j. However.k (t) "n+k (t) .k '2. we have i. 2 (t)) Cov ( R1 (t) R2 (t)j 1 (t) .k : k=1 It is important to be able to estimate the dependence structure of the discrete time model from data. but is necessary from an applied perspective due to our inability to estimate the Yj : Inserting this into Equation (3.time version of the model in Equations (2. for two stocks Ri i = 1.k 1 (t) "1 (t) + '1.5) and (2.4) holds for i = 1.k 2 1 1 (t) "n+k (t)A (t) "n+k (t) 1 (t) .k =E = 0v u u @t1 k=1 q X '22. :::. 2. That is. such that Equation (2. since we have little hope of estimating the news processes Yj accurately. 1. 2 (t)) 1 (t) 2 (t) 20v u q q X X u 2 t 4 @ 1 '1.0 = 1 '2i. In fact. (3. 2 3 (t)5 . q. and not with the underlying news processes Yj : This is a reasonable restriction. Since the volatility is stochastic we can not hope for constant correlation. :::. :::.k i . we require that for some 'i.k 0. 1) variables.2) gives the discrete time model Ri (t) = i + i v u u 2 t1 i (t) + q X '2i. k = 0. (3. n: In this paper we will only work with the discrete time volatilities i2 (1) . n.k i (t) "i (t) + k=1 q X 'i.k k=1 1 1 (t) 2 (t) q X = '1. and "i ( ) are sequences of independent N (0. This might not be totally realistic. :::. i = 1.k = 'i. :::. Ri (t) = i 2 i i + (t) + i.k : 2 k=1 (t) "2 (t) + q X k=1 k=1 6 '2. we assume for simplicity that i. Therefore. the conditional correlation between the returns of di¤erent stocks turns out to be constant. i2 (d) .k are such that the volatilities for each factor are fractions of the total volatility.0 (t) "i (t) + q X i.k i (t) "n+k (t) . we have that Corr ( R1 (t) R2 (t)j 1 (t) . 2.2) k=1 where t = 1.6).

and [16]. This choice was made primarily for two reasons. We treat only the univariate version of our model. The Normalized Inverse Gaussian distribution (N IG) is used to illustrate the discussion. [5].k=1 . A fundamental paper is [7]. However. we can apply standard factor analysis to the i:i:d: normalized returns i ( ) : This gives us the parameters 'i. 3. too. Ericsson survived and its stock started to increase in value. the reason that the approach fails is valid for the more general GH-distribution.3 Limitations of the GH-distribution In this section we discuss a natural way to estimate the model parameters. First the time series is long enough to give a fairly large amount of data. and the cumulative number of trades on each day. e.k=q is a diagonal matrix where the non-zero elewhere = ('i. The stocks are Ericsson B. The time series were kindly given to us by SIX . The reason for this is that by then the long period of decreasing stock prices was somewhat distant in time. and AstraZeneca. SKF B. Volvo B.g. [2]. 3. These …ve are all large companies and are also among the most traded stocks at the exchange. :::. see 2 + 2.k . We have chosen to analyze data from the time period August 1.4) i=n.Stockholm Information Exchange. We stop right before the summer of 2004. Next. that is n = 1. This choice of time period of observation gives d = 208: The data we have used is the daily closing prices for each stock. We have chosen to start our period of observation after the summer of 2003. and Pq 2 ments are 1 k=1 'i. and still short enough to make it reasonable to assume stationarity. Atlas Copco B. which are given indices i = 1. and in the spring of 2003 so did the exchange as a whole. During this time period the company Ericsson. This second reason is also supported by economical considerations. The N IG-distribution has been shown to …t …nancial return data p well. However.k : This result implies that if we …nd a good model for the volatility 2 ( ) . which was the most in‡uential stock on the exchange. The N IG-distribution has parameters = 7 . 2004.In general the conditional correlation matrix is given by 0 + . and why this approach is not successful. had been close to bankruptcy. Constant conditional correlation is a common feature in discrete time …nance. (3. The OMX Stockholmsbörsen as a whole decreased in value three years in a row from spring 2000 to spring 2003.k )i. we discuss the data to be analyzed.2 The stock market data We consider …ve di¤erent stocks from the OMX Stockholmsbörsen. and hence also the implicit constant conditional correlation matrix that describes the correlation between the returns. The summers are avoided since we suspect that they will give us problems with non-stationarity due to less activity on the exchange. 2003 to June 1. . 5.

5 2 x 10 -3 Figure 3.3000 14 12 2500 10 2000 8 1500 6 1000 4 500 2 -0.1 -0.5) a set of IG parameters. It turns out that this does not hold for our data set. since both parameter sets are estimated from the same data. . This gives through Equation (3. and 0 j j : The N IG-parameters can be estimated from data by the maximum likelihood method in a straightforward way. Parameter set 2: Dotted line. and . > 0. Parameter set 1: Solid line. 1) variables. and its density function is fN IG (x. Parameter set 1: Solid line. .05 0 0. where and are the same as in the N IG-distribution. Recall that the volatility 2 will be observable. It is well known that if 2 has an IG distribution and is standard normal.1 0.5) has a N IG distribution. The IG-distribution has density fIG (x. 8 . The Inverse Gaussian distribution (IG) is related to the N IG-distribution. . .5) gives normalized returns ( ) that appear much like i:i:d: N (0. p where q (x) = 1 + x2 and K1 denotes the modi…ed Bessel function of the third kind with index 1: The domain of the parameters is 2 R. then r= 2 + + (3. Right: IG densities for Ericsson. Then we would expect to 2 see that our estimated ( ) had approximately the same IG-distribution as the N IG-distributed returns imply. Assume further that we manage to estimate the volatility 2 2 process ( ) so that ( ) and r ( ) in Equation (3. . We illustrate this with the following example.05 0 0. ) = p 2 exp ( )x 3 2 1 2 exp 2 x 1 + 2 x . ) p 2 = exp 2 1 x q K1 x q e x. x > 0.1: Left: N IG densities for Ericsson. .5 1 1. Suppose now that we estimate the N IG-distribution from a set of returns r ( ) with the maximum likelihood method. Parameter set 2: Dotted line.

0:0650. there is a substantial di¤erence between the IG-densities. 1) : If we can do this. Still. 2. = 108. the N IG-distribution: If we have information only about the returns of an asset. These were 1. 1 = 73:8. we can model the i2 within the framework in [5]. Accordingly. That is. the GH-distribution is also "almost" overparameterized. In other words.When we estimated the N IG parameters for the observed Ericsson returns our MATLAB routine came up with two very di¤erent maximum likelihood estimates. We need to …nd a way to single out which set of N IG parameter values that are. 1 1 = 0:0591 = 0:114 These parameter sets give the N IG and IG densities that are shown in Figure 3. there are too many of them. This made the parameter estimates very unstable. one can obtain a good …t of the N IG-distribution to return data for many IG-distributions. 9 .1. 1 = 211. From those it could be seen that there were directions in the parameter space along which the likelihood function for the N IG-distribution was very ‡at. the N IG-distribution is "almost" overparameterized.1). one should try to …nd ways to measure i2 with parameters i and i such that the normalized returns i ( ) are i:i:d: and N (0. 4 Analysis In this section we outline an approach to …tting the discrete time model to data. then r= 2 + + has a GH-distribution. This veri…es the validity of the univariate discrete time model. the correct ones.1 Method We propose that instead of using the returns directly one should try to estimate the model through Equation (4. the likelihood functions for the two sets were within 0:2% of each other. 1 1 = = 0:0092. depending on the starting values. Equation (3.2 and estimating the N IG and IG distributions to the returns and volatility. Loosely speaking. The methods are illustrated by applying them to the data set from Subsection 3. This example indicates that the problem with estimating the N IG-distribution from returns is not that it is hard to …nd good parameter estimates. in some sense. 1 1 = 14:7. Rather.5) actually holds in more general. respectively. We realize from this result that the GH-distribution must have the same problem as its subset. If 2 has a Generalized Inverse Gaussian (GIG) distribution and is standard normal. We see that the N IG-distributions are virtually indistinguishable. 4. We also made a number of pro…le likelihood function plots.

with i and i …xed. Since the N IG-distribution is very ‡exible.1) become independent N (0.1). Since the concepts of stochastic time change and stochastic volatility are related. The model in Equation (4. This implies that our discrete time volatility model in Equation (3. Ri (d) .2) is inspired by [1].which allows us to understand better the structure of the process that generated the returns Ri ( ) : The e¤ort to simultaneously …t the returns Ri ( ) to the 2 N IG-distribution.2) is well-de…ned. Equation (3. i (0)) = log ( ( i (1)) ::: 0 d 1 X B Ri (t) = @ 2 t=1 ( i (d))) 2 i+ i 2 i (t) 2 i (t) + log 2 i (t) 1 C A: We recall that the continuous time volatility is de…ned as a linear combination of news processes Yj from Equation (2. and …t the N IG and IG distributions to the returns and the estimated volatility process later. See Equation (3. an average of the continuous time volatility on that trading day. That is. To …nd the parameters in the 2 distribution of i ( ) is the next priority. The analysis is done in four steps. we want to use a similar idea with daily data for our discretized model. in a sense. :::. In other words.5) then gives an implied distribution of the returns Ri ( ) that we have a good understanding of. as L ( i . 1. 1) : Here we assume that the discrete time volatility processes i2 is a constant times the number of trades zi ( ) on each trading day.2). Find volatility processes i2 and parameters that the normalized returns i ( ) = Ri (t) i i 2 + i i (t) and = i i for each stock so (t) (4. a 1=2 . a > 0: Hence our volatility model in Equation (4. The understanding of the model lies …rst of all in 2 getting i ( ) and the model for i ( ) correct. and the normalized returns i ( ) to the normal distribution would probably not make the …t better. Note that for X 2 IG ( . we verify 10 . which uses the cumulative number of trades as a stochastic clock. the volatility process i ( ) to the IG-distribution. we have that aX 2 IG a1=2 .3) is. we are likely to get almost as good estimates by …rst trying to obtain normality of i ( ) .2) This means that we can write the loglikelihood function L for the observations Ri (1) . we can view the continuous time volatility as the intensity with which new trades "arrive". 2 i () = i zi ( ): (4. The paper [1] shows that the intraday cumulative number of trades contains enough information to allow us to obtain almost perfect predictions of the volatility in the near future. Further. ) .

the i ( ) might behave like an i:i:d: sample. In other words. The next step is to …nd parameters i and i so that the empirical distribup 2 2 2 tions of i ( ) from Equation (4. i (0)2 ) (h) = . option pricing theory. by minimizing the least squared distance between the theoretical and empirical autocorrelation function. distribution. i (d) . We require the rates of decay j to be equal for all stocks to allow for the news processes Yj to be shared by di¤erent stocks. especially for large . we know that this will be hard.4) that we need the matrix = ('i. 1) : The reason is that such a model would make it easier to apply results from. we have also speci…ed the N IG-distribution for Ri ( ) : We could do this estimation simultaneously for IG and N IG. where the !j 0. We know from Equation i=n. we have asserted that our continuous time stochastic volatility model is reasonable. Hence. does not take into account the information released at time t. Further. and model the number of trades within the model in [5].k=1 to estimate the correlation between the returns of di¤erent stocks. we get an economical interpretation of the volatility. Since Equation (4. portfolio optimization.k )i. We next use the estimates of the volatility processes i2 to estimate the rates of decay j : This can be done by using the autocorrelation function of the continuous time volatility process i2 : The autocorrelation i is de…ned by Cov ( i (h)2 .k=q (3.1) to obtain i ( ) that are i:i:d: and N (0. 1). :::. this implies that we identify the daily number of trades with the volatility. If we can do this. We recall that in factor analysis it is assumed that a vector x of observed variables with mean 0 can be written as x = f + e: Here is a constant n q matrix of factor 11 . are the weights from the volatility processes that sum to one. since the N IGdistribution is very special we know that even if we would get a slightly better …t this way. This is due to that a previsible model with i (t) 2 Ft 1 . 3. Note 2 that it would be desirable to …nd a previsible model for i ( ) such that the normalized returns i ( ) are i:i:d: and N (0. for example.2 that we can use a constant times the number of trades as i ( ) in Equation (4.1) holds in the model in [5]. 4. The …nal step is to apply factor analysis to the normalized returns to …nd the correlation between the di¤erent stocks. h 0: i V ar ( i (0)2 ) Straightforward calculations show that i (h) = !1 exp ( 1 jhj) + ::: + !m exp ( m jhj) . but their distribution will have thicker tails than the standard normal.2) …t the IG i . it would be at the cost of less understanding of the process. The impact of this information could be considerable. However. We estimate the rates of decay j from the discrete time volatilities i (1) . However. 2. i i . such as the GARCH. or risk management.