6 Portfolio Optimization-IT 6.1 Introduction |Wo have presented Markowitz’s mean variance model for portfotio optimization re he last chapter. Bnt contrary to its theoretical reputation, this mode! i its original form has not found much favor with the practitioners $0 construct large scale portfolios. ‘There are several theoretical and practical reasons for not using sree odel extensively in practice ~ particularly when the aumber of assets in the portfolio is large. This chapter aims to understand these reasons ‘and then present aera other models which have been developed to improve Markowitz's model both. theoretically and computationally. 6.2 Markowitz’s Model: Some Theoretical and Computational Issues ‘Theoretically, Markowitz’s model is known to be valid if returns 1/8 are multivari- ate normally distributed and the investor is ris averse in the sense that he/she prefers less standard deviation of the portfolio to more, But one not fully con- are if the validity of the standard deviation as a measure of risk. An investor is ertainly unhappy to have small or negative profit, but feels happy te have larger profit, In other words, this means that the investor's peroeption about risk is not symmetric about the mean. There are several empirical studies, which reveal that soewt 7 are not normally or even symmetrically distributed. In this sconrs, 02° possible approach seems to be to consider the skewness and kurtosis of the distri- ree a in addition to the mean and variance and extend the Markowita’s model to generate the efficient frontier in (mean, variance, skeyness, kurtosis)-space. This fenehoon the approach of some of the models, eg. Konno and Sumki (79], and468 Financial Mathematios: An Introduction Joro avd Na [71], but has not found much favor becouse the resulting optimiza: Jere problem is not easy to handle. An alternsit™ ‘and popular approach is to smoduce certain new measures of risk which CAPT information about the possi thle portfolio losses implied by the tail of ‘che return distribution, even in the case ae eothe distribution is not symmetric. This takes co of those situations where The return distribution is heavily tailed. These ‘isk measures are called downside sm safety first risk measures which alan to Tae the probability that the port: iio lose is below a certain acceptable level, Commonly referred as the benchmark vovtne disaster level. Thus these risk mensures 5° quantile based risk measures nd are different from standard deviation oF ‘other moment based risk measures: Some of the most popular quantile based isk measures are value at tisk (VaR) od conditional value at risk (CVAaR). Since Gjownside risk measures of individual seourities cannot be casily aggregated into portfolio dovinside risk measures (we vsoed the entire joint distribution of security returns), their application in prac: tice requires computationally, intensive non parametric estimation, simulation and optimization techniques. ‘rere is another major problem associated with the classical Markowitz’s model, This model gives us an optimal portfolio assuming that we have perfect Information about pi’s and d,s for the assets ‘that we are considering. Therefore an important practical issue 18 the estimation vg and G's. A reasonable approach for estimating these data is 0 057 ti past returns 7: which represents the return of # asset from time (¢-1) to f, where £= 1,2,.../T. How: eer, it has been observed that small changes 1% the time series Ty lead to changes fn the pi’s and i's that often lead to significant changes in the optimal portfolio. wae ida fundamental wealmess of the Markow/ita model, no matter how cleverly u's and oi/'s are computed. This is Desanse the optimal portfolio construction 16 vi ny sensitive to small changes in the data, Only one small change in one pi BA produce a totally different portfolio, Tp Tact recent research (Chopra. and Ziembs [B0]) has revealed that exrors nthe “sctimation of means pi can be more damaging than errors in other parameters. This has motivated researchers to employ robust ‘optimization techniques in Markowita’s w odel, eg, Ben-Tal (9), and Tatiznet! and Koenig [139]. A much simpler approach is 10 consider portfolio optimization under paninimmax rule (Cai et al. (23]) and provide some flexibility by allowing pi to Tie va some interval a; < pis by (Deng et al. (36) "The mean-variance model of Markowitz, in general, results in a dense quadrotie programming problem. If the number ee saaets in the portfolio is large then Tt veeemes very difficult to obtain an optimal olution of such large-scale dense quadratic programming problem on @ veal time basis. This has motivated Tsimization: A Ly-Risk Model 169 Mean Absolute Deviation Based Portfolio Op searchers to consider mean-absolute deviation of the portfolio as a measure of risk (Konno and Yamazaki [80)) so that the resulting optimization problem reduces to a inear programming problem. These models are called L1-risk models because hese are based on Ly metric on RY, In this terminology, “Markowitz’s model can be termed as a Lz-risk model since it uses variance aS & isk measure which is based ta the notion of Lz metric. In the same spirit, the models based on the minimax tale portfolio selection strategies can be termed as Lyq-risk models. It is simple to Mote that L, and Lu Fisk models wil result ih a Linear programming formulation. ‘This ig in contrast to an Lo- risk model which results i! @ quadratic programming emulation, Therefore computationally, Ly and Le tisk models are easier to han- ide i comparison to La risk model. The Ly» La and Le are three standard moment tpaved risk models which have been studied in the lterait® The above discussion suggests that the choice of « proper risk measure is very crucial to study portfolio selection problems, But then what would be the guiding principles for this choice? The concept of coherent risk measures (Artaner et al. (51) Jean important contribution in this regard. This stipulates an axiomatic approach to the study of risk measure by presenting certain desirable properties. Some other issues associated with Markowite's ‘model are very natural and these have been incorporated in the originall model. “These ot with regard to its extension for the multi-period scenario and also to incorporate transaction costs. We shall discuss some of the models in the subsequent sections which addresses the above issues. 6.3 Mean Absolute Deviation Based Portfolio Optimization: A LyRisk Model risk measure and formulate the corresponding portfolio optimization problem with this risk measur ‘Let an investor has initial wealth My which is to be invested inn assets 4 (i=1,2,.-:n). Let 1; be the return of the vjom variable. Also let x be the amount of money to be asset a; which is a ranc invested in the asset a; out of the total fund Mo ‘We now define 4; = E(ri) and qi = Ei = pil), =1,2,--- expected return rate of the asset a, and gi denotes the expecte cro from its mean, Then the expected return of & portfolio (x1, X2y---/%n) i8 given We first introduce the Li wn. Then pi denotes the .d absolute deviation by170 Financial Mathematics: An Introduction We now define Ly-tisk measure or the mean absolute deviation of the portfolio rr %p,--+/En)- Definition 6.3.1 (Li-Risk Measure of a Portfolio) Let (1,X2,+-+/n) be the given portfolio. Then its Ly-risk measure or mean absolute deviation is defined as | i In terms of the Ly-risk measure 1, (x1, %2,---»%x), the Ly-tisk model of the portfolio optimization problem is formulated as > ra (5, ns] ee i= Wry eX (1) Oexsu G=12,.00) where (a > 1) is a parameter representing the minimum rate of retum required by the investor. Also u; is the maximum amount of the money which can be invested in the asset a). Tn (6.1) it may be noted that by defining w; = xi/Mo, (= 1,2,...,1), the portfolio (t1,%2,--+/%x) can also be represented in terms of its weight vector (1 1,..-,i0,). Traditionally in the portfolio optimization literature, a portfo- Tio has been represented by its weight vector (20, W2,-..,W,) but it can also be equivalently represented in terms of its amount allocation vector (81, ¥2,-++/%n)- Tn practice, the historical data is used to estimate the parameters in the opti- mization problem (6.1). Let ry be the realization of the random variable 7; during the period t (f= 1,2,...,T). We assume that 7 is available through the historical data and the expected value of 7; can be approximated by the average derived from the data. This gives az mes E() = 5 yt (62) Also wy (t-%y-++/%x) can be approximated byMean Absolute Deviation Based Portfolio Optimlastion: A Ly-Risk Model 171 | cerns ) Sta eee ane tai [= In (6.3) it may be noted that, due to the absolute value function, the expres sion on the right hand side becomes a nonlinear and non smooth function of Cer Rayo Xe): Using (6.2) and (6.3), problem (6.2) can be reformulated as Ses ~ Hz ‘= : (6.3) win AY. i subject to (6.4) If we now denote by y; and employ the definition of the absolute Sew = pa a function then we get a= ye ~ Ha] a = Max (Xe. wx - Yu vos) (65) ia i Therefore problem (6. 4) gets transformed tosubject to Lie = HRS Ye oa Yn HORS Ye a (66) Me where the first two constraints in (6.6) follow from (6.5) and the definition of maximum. Denoting (ri — wi) by ew @= 1,2, j £=1,2,...,7), we can rewrite (6.6) as Min subject to (=1,2,...,T) (@=12,...7) j (67) @=1,2,...,0)- Remark 6.3.1 The formulation (6.4) is essentially nonlinear and non smooth. But since this nonlinearity and non smoothness occurs due to the presence of ‘absolute value function only, it can be handled in a reasonably simple manner 4Mean Absolute Deviation Based Portfolio Optimization: A Ly-Risk Model 173 as explained abone. The resulting linear programming problem (6.7) can be solved efficiently even when n is large. One obvious question at this-stage is to enquire if there is any relationship between L; and L2-risk models. The below given theorem answers this question. Theorem 6.3.1 Let (r1,72-++/tn) be multivariate normally distributed. Then for a given portfolio x = (x1, X2)+--/%n) wry (x) = zoe, where the standard deviation a(x) is given by eanennd= Sere} = a Proof, Let (0j;) € R™" be the variance-covariance matrix of (r1,72, under the given hypothesis, ) ' rit; is normally distributed with mean Yin and ‘= standard deviation a= |), yy ou, i Therefore 1 uw mt sg Sloe ata) 2 we = eal - o(-ay)a which on substitution (12/202) = s gives oa w(x) = e “© vase a o In view of Theorem 6.3.1, for the multivariate normal case, both Ly and L2- risk models are equivalent. In other words, Markowitz's mean-variance portfolioi : & fe f & 174 Financial Mathematics: Am Introduction selection strategy will be the same as the one given by mean-absolute deviation selection strategy. Even in the case when normality assumption does not hold, through certain case studies, it has been shown that minimizing the Ly-risk pro- duces portfolios which are comparable to Markowitz’s mean-variance model which minimizes Lo-risk, i.e, standard deviation, However as one expects the variance of mean-absolute deviation portfolio is always at least as large as the corresponding mean-variance portfolio. But in actual applications, this difference is small. In fact Konno and Yamazaki [80] applied both L; and L,-risk models in Tokyo Stock Market by using historical data of 224 stocks in NIKKEL 225 index. They generated efficient frontiers and observed that the difference of the standard devi- ation of the optimal portfolio generated by Lz and Ly-risk models is at most 10% for what ever value of a. Of course two frontiers will coincide if r’s are multivari- ate normally distributed. Thus this difference can be largely attributed to the non normality of the data. Therefore irrespective of the distribution scenario, L;-risk model provides a good alternative to Markowitz’s L>-tisk model. Advantages of the Formulation (6.7) We now list somo of the advantages of L;-risk model (6.7) over the classical Markowitz’s model which minimizes the Lo-risk. (i) The formulation (6.7) is a linear programming problem and hence can be solved much mote efficiently in comparison to quadratic programming problem which is obtained in Markowitz’s model. ‘This is particulary significant for portfolios having large number of assets. (ii) We do not need to calculate the variance-covariance matrix to set up the Ly-risk model. (ai) In the formulation (6.7) there are always (2T+2) constraints regardless of the number of assets included in the model. ‘This allows to handle very large portfolios on a real time basis. (iv) An optimal solution x* = (24,x3,...,27,) of problem (6.7) contains at most (21+2) positive components if uj = +00 for i= 1,2,...,m. This means that an optimal portfolio consists of at most (2T+2) assets regardless of the size of n. Therefore we can use T as a control variable when we wish to restrict the number of assets in the portfolio.Mean Absolute Deviation Based Portfolio Optimization; An L-Risk Model 175 as explained above. The resulting linear programming problem (6.73on: An efficiently even when n is large. One obvious question at this-stage is to enquire if there is any relatio. between Ly and Ly-risk models. The below given theorem answers this questicil. ‘Theorem 6.3.1 Let (r1,12,-..+Tn) be multivariate normally distributed. Then for « given portfolio x = (x1,X2,---»%s) wy, (x) = 2009, where the standard deviation o(x) is given by EEA] = I O(%1,X Proof. Let (9) € R" be the variance-covariance matrix of (71,12,-.-,Tx)- Then under the given hypothesis, )” rixj is normally distributed with mean Yinei ana standard deviation ‘Therefore m,0)= Ilex = ) a * V2r0(x) J 207(x) ; 2 cs Vino [vee (am) which on substitution (u?/20?) = s gives wr (2) = fe oe | oe 2000, In view of Theorem 6.3.1, for the multivariate normal case, both Ly and Ly risk models are equivalent. In other words, Markowita’s mean-variance portfolio o