Research on regularized mean–variance portfolio selection strategy with modified Roy safety-first principle

We propose a consolidated risk measure based on variance and the safety-first principle in a mean-risk portfolio optimization framework. The safety-first principle to financial portfolio selection strategy is modified and improved. Our proposed models are subjected to norm regularization to seek near-optimal stable and sparse portfolios. We compare the cumulative wealth of our preferred proposed model to a benchmark, S&P 500 index for the same period. Our proposed portfolio strategies have better out-of-sample performance than the selected alternative portfolio rules in literature and control the downside risk of the portfolio returns.

equities. Rachev (2001) showed how the safety-first approach can be more efficient than the stable Paretian approach in portfolio theory. Levy and Sarnat (1972) and Nawrocki (1999) show how safety-first principle and meanvariance approach are akin. Levy and Sarnat (1972) found out that in the distinctive case where target rate of return is equal to risk-free return, both Roy's safety-first principle and Markowitz's mean-variance approach lead to the same optimal portfolio selection strategy. However, both models have several restrictions which include but are not limited to the following: Firstly, both Markowitz and Roy employ symmetry about the mean variance as a risk measure. Variance takes into account the situation in which return exceeds mean value and this does not impact risk. Secondly, transaction costs are absent in safetyfirst principle and mean-variance approach. According to DeMiguel et al. (2014), their loss in the mean-variance approach with the absence of transaction costs was 49.33 %. Borkovec et al. (2010) pointed out that 40 % of financial market participants attribute the fundamental loss in abnormal return to transaction costs. Lastly, both model frameworks are subjected to parameter uncertainty as returns of assets are considered as deterministic parameters being depicted by a single point estimate which leads to estimation risk.
Variance, as a risk measure, penalizes both under-performance and over-performance equally (Markowitz 1968). However, investors are only worried about underperformance. Roman et al. (2007) obtained a better portfolio by employing three indexes i.e. mean-variance-CVaR in comparison to mean-variance and mean-CVaR. The authors demonstrated that the mean-CVaR portfolio policy results in large variance, which leads to a small Sharpe ratio. Again, the CVaR of the portfolio generating from mean-variance model is large. To eliminate these inconsistencies between the strategies of meanvariance and mean-downside risk models, Roman et al. (2007) proposed to merge CVaR and variance in a multi-objective portfolio selection strategy. Inspired by Roman et al. (2007)'s approach of combining a downside risk measure, CVaR, and variance, we consider a portfolio optimization model with multiple risk measures. Recognizing a need to modify and improve Roy's safety first principle, we consider merging our improved Roysafety first approach with variance as the portfolio risk measure.
The construction of a portfolio of investments is a significant problem faced by investors and institutions. A decision ought to be made to allocate weights to each investment with the intention of striking an appropriate balance between returns and risk. In reality, setting up a new portfolio or rebalancing an existing one requires costs to be incurred and must be inclusive in any realistic analysis. We incorporate proportional transaction costs (Kellerer et al. 2000;Muthuraman and Kumar 2006) which are induced by liquidity costs, tax and brokerage fees (Dumas and Luciano 1991;Kellerer et al. 2000;Lobo et al. 2007) into our portfolio selection model.
In the mean-variance and Roy safety-first models, stock returns are considered deterministic and taken as a single point estimate which results in estimation risk or overfitting (Bawa et al. 1979;Merton 1980). A small variation in the input parameters of the standard Markowitz mean-variance approach and Roy safety-first principle usually lead to changes in the structure of the resulting portfolios (Brandt 2009). To reduce the undesired impact of estimation risk or overfitting, models in the context of robustness (Goldfarb and Iyengar 2003;Tütüncü and Koenig 2004), stochastic programming (Rockafellar and Uryasev 2000), factor models (Green and Hollifield 1992;Nagai 2003;Schultz and Tiedemann 2003) and shrinkage estimators (Jorion 1986;Ledoit and Wolf 2004) have been explored. Another method analogous to our work is the modification of portfolio weights by adding regularizers or additional constraints to the portfolio strategy (Jagannathan and Ma 2003;DeMiguel et al. 2009;Brodie et al. 2009). Correcting undesired portfolio weights are as a result of large estimation risk of unknown parameters leading to the achievement of their desired forms and characteristics. Jagannathan and Ma (2003) employed a short sale constraint in the minimum variance framework and found out the non-negativity constraint performs well as those computed with shrinkage estimators and factor models. In DeMiguel et al. (2009), authors added a convex norm ball constraint to the portfolio weight and observed that the norm ball constrained portfolios had better out of sample performance than the portfolio strategies of the naive 1/N diversification (Jagannathan and Ma 2003). The l 2 -norm constrained portfolio in general attained higher Sharpe Ratio than l 1 -norm constrained portfolios. We study the weight constrained portfolios by specifying the general norm as squared l 2 -norm ball.
The optimal portfolio of classical Roy safety-first principle and Markowitz's model in the presence or absence of short sale constraint or stability constraints hold a large number of assets and especially small weights assigned to their proportions. However, holding a large number of assets leads to the investor incurring high transaction costs. Due to this factor and other market and economic frictions, investors often hold only a small number of stocks in their portfolio. This is usually known as a sparse portfolio (a portfolio with few non-zero weights). With sparsity, one pre-sets limit of assets (stocks) with non-zero entries of portfolio allocations.
A method of constructing sparse portfolio termed hard threshold strategy was proposed by Britten-Jones (1999) to statistically test each portfolio weight with a null hypothesis that weight is zero. Using F-test and t-test, Britten-Jones (1999) proposed that if the portfolio weights are statistically not different from zero; assigning zero reduces portfolio risk. Sparse portfolio weights can be derived by cardinality constrained portfolio optimization (CCPO) problem (Chang et al. 2000;Maringer and Kellerer 2003;Ruiz-Torrubiano and Suárez 2015). CCPO takes into account all portfolios of given number of assets and chooses an optimal portfolio. With the inclusion of cardinality constraint, the portfolio selection strategy becomes NP-hard (Moral-Escudero et al. 2006), and standard quadratic program solvers can no longer be adopted to tackle the problem. One resorts to several other methods or relaxations in search of near-optimal solutions at a moderate computational cost. Farrell and Reinhart (1997) suggested classification of assets based on geographical aspect, size, sector, etc. and made a selection of N assets from each of these classes. In Pai and Michel (2009), the cardinality constraints were handled via clustering algorithm to reduce the size of the portfolio. More recently, Ruiz-Torrubiano and Suárez (2015) used a memetic approach that combines a genetic algorithm (GA) with an extended set encoding and quadratic programming (QP) in a mean-variance framework to deal with the cardinality constraint. An alternative relaxation method for which we employ in this paper is by constraining the l 1 -norm (Tibshirani 1996) which is also in connection with the upper bound of the estimation risk as shown in Fan et al. (2012). The convexity nature of this type of regularization makes it more tractable. The use of norm penalty helps investors limit transaction costs and exposure to risky stocks.
As Roy safety-first principle minimizes the chances that the portfolio's return will fall below the minimum acceptable return, introducing it in the mean-variance model helps control the downside risk of the portfolio return. The modification and improvement of Roy safety-first principle and merging it with variance as a consolidated risk measure in a risk-return framework represents the main novelty of this research paper. In achieving our aim, we minimize our modified and improved Roy safety-first principle and impose the lower constraint on the mean return of the portfolio as in Markowitz's mean-variance model. We also investigate the portfolio strategy with variance and Roy's safety first principle as a consolidated risk measure in a mean-risk framework. To address the problem of estimation risk, we constrain the portfolio weights with squared l 2 -norm and proceed to achieve sparsity via l 1 -norm heuristic. We will explore the impact of transaction costs on portfolio selection strategies.
The paper is organized as follows. The next section presents our proposed portfolio selection strategies. It's subsections study Roy safety-first principle and its modification, portfolio revision and stable and sparse portfolio. We perform numerical tests and present computational results of our proposed methods in the subsequent section. Concluding remarks are provided in the last section.

Proposed portfolio selection strategies
Investors allocate proportions of their capital among the assets they invest in. For the purpose of this study, these proportions are allocated to stocks. We denote by N the number of risky assets, R is the required return level, and x 0 is the initial risky assets before rebalancing: x 0 k is the proportion of capital initially allocated to asset k, k = 1, 2, 3, ..., N. Let x, x b and x s be N dimensional vectors of controllable variables: x k is the portfolio invested in risky asset k after rebalancing, x b k is the purchases (proportion used) of risky asset k and x s k is the sales (proportion obtained) of risky asset k. The transaction costs incurred when buying risky assets is c b and that of selling risky assets is c s . The financial portfolio is described by a N-dimensional vector of random returns r. The portfolio total random return is R p = f (x, r) = N k=1 x k r k , portfolio expected return vector ) 2 as variance of portfolio return and Q as the variance-covariance matrix of the portfolio return.

Roy safety-first principle and it's modification
Most investors' aim is to maximize returns and minimize risk. The Roy safety-first principle advocates avoiding extreme losses through the minimization of disaster probability. To optimally construct a portfolio strategy, Roy's safety-first principle defines a threshold or a minimum acceptable return R, below which the portfolio wealth is considered to be a disaster. The best portfolio is one that minimizes the chances that the portfolio's return, R p , will fall below a minimum acceptable return, R. In essence, an investor selects his portfolio by solving this optimization problem: where e is a vector with ones as entries and P is a probability measure. Roy employed Bienayme-Tchebycheff 's inequality as the investor is likely not to know the actual probability function and obtained an approximation subject to e T x = 1 Thus, the optimization problem is reformulated as The modification to the Roy's approach we adopt is by using a coherent downside risk measure known as Conditional Value-at-Risk or Expected shortfall as it has a set of desirable properties for a risk measure (Platen and Heath 2006) leading to more accurate estimates of probability. For a detailed study on desirable properties of an ideal risk measure in portfolio theory, we refer the reader to Rachev et al. (2008).
A well-known downside risk measure known as Value-at-Risk focuses on the percentiles of loss distributions and measures the predicted maximum loss at a given probability level. Mathematically it is formulated as α-quantile VaR α (X) = min{z | (F X (z) ≥ α} , where X is a loss random variable and α ∈ (0, 1) is the given probability level. Values for α often used are 90 %, 95 % and 99 %. Considering Value-at-Risk (VaR) has undesirable properties such as non-subadditive and non-smooth etc., Rockafellar and Uryasev (2000) introduced a coherent downside risk measure termed Conditional Value-at-Risk (CVaR) and for α ∈ (0, 1) represented it as where Equivalently, for x ∈ X ⊆ R N and random vector r ∈ R N which represents the actual portfolio return has a continuous density function p(r) First, we will determine the semi-deviation of the random return f(x, r) from the α-quantile VaR α (x). With respect to Bienayme-Tchebycheff 's inequality, the following estimate is valid for VaR α) (x) > R:

Let us consider an α-quantile
and a measure of risk which is termed expected shortfall from the α-quantile VaR α (x) value. The estimate (3) can be written as Therefore (1) can be reformulated by considering the approximation of the right hand side of (4) and obtain the following Telser (1955) considered a portfolio strategy by maximizing portfolio returns under the the constraint of Roy safety-first principle. He solved the optimization problem: Inspired by Telser's approach, we constrain (5) with the minimum mean return vector µ T x ≥ L from below where L is the lower bound of u T x, where L > R. Thus we obtain the optimization problem : In another approach, we consider variance and the modified Roy's safety first-principle as a consolidated risk measure in a mean-risk framework. To this end, we propose the optimization problem: The rest of the modifications is geared towards investigating realistic constraints such as transaction costs, sparsity, and stability to P 0 and P 1 on the financial market.

Portfolio revision
We consider an extension of problems (7) and (8) in which transaction costs are incurred to rebalance or revise the initial portfolio x 0 , into an efficient portfolio x. A portfolio of investments may require rebalancing on periodical basis because of updated risk, and return information is generated over time. We make the following assumptions on the transaction cost function c.

Assumption 1
The transaction cost function satisfies the following: To achieve portfolio x k from the previous or initial portfolio x 0 k , we make a payment of transaction costs c(x − x 0 ). We incorporate proportional transaction costs (Kellerer et al. 2000;Muthuraman and Kumar 2006;Mitchell and Braun 2013) which are induced by liquidity costs, tax, brokerage fees (Dumas and Luciano 1991;Kellerer et al. 2000;Lobo et al. 2007) into our model. Therefore, proportional transaction cost follows this structure: where where cost of buying is c b k 0 and cost of selling is c s k > 0. The P 1 model with proportional transaction costs (P 1t ) is the optimization problem The model P 1t minimizes the upper bound estimate (5) w.r.t x and superimposes the lower on the average return after the deduction of transaction costs.
Explaining the constraints with respect to transaction costs, the above optimization problem is subjected to a set of linear constraints. Constraint (10) requires the net return of the portfolio after the deduction of transaction costs to be greater or equal to a threshold level L. Constraint (11) is the budget constraint: the capital available to cover transaction costs and shares of stocks. Constraint (12) shows that x k represents the portfolio position to be chosen explicitly through sold shares x s k and purchased shares x b k that are rebalanced adjustments to the initial position x 0 k of stock k. Constraint (13) and constraint (15) are the complementary constraint and non-negative constraint respectively. They prevent any possibility of concurrent purchases and sales (Dybvig 2005). Note that P 0 model with proportional cost (P 0t ) is defined similarly but without the variance term in the objective function.

Stable and sparse portfolio
In a portfolio selection strategy where the dimensionality of the set of candidate assets is high, sparsity is desired. When the number of assets is large, a non-regularized numerical approach will intensify the effects of estimation risk, leading to an unstable and unreliable estimate of the vector x. Typically, portfolio managers want to set up portfolios with suitable balance between risk and return by investing in a small number of assets, thereby limiting their transaction, management, and monitoring costs.
To obtain meaningful and sparse (zero components) results, a regularization procedure is usually adopted. A standard approach is to augment the objective function of interest with a l 0 -norm penalty or adding a cardinality constraint �x� 0 ≤ N ′ to optimization problems P 0t and P 1t , where x 0 is the number of the non-zero entries of x and N ′ is the upper bound limitation of assets to be managed in the portfolio. However, with the inclusion of cardinality constraint, the portfolio selection strategy becomes NP-hard (Moral-Escudero et al. 2006). We therefore impose its equivalent l 1 -norm penalty as employed by Brodie et al. (2009) among others. The l 1 -norm is a convex function of x, and such convex relaxation makes portfolio selection strategy more tractable.
To this end, we suggest to evaluate the portfolio weights by where the l 1 -norm of a vector x ∈ R N is defined by �x� 1 := N k=1 | x k | and τ 1 is an adjustable parameter that controls the sparsity of the portfolios. Similarly, S 0t can be formulated without the variance term in the objective function.
Optimization models S 0t and S 1t have estimation risk or overfitting problem. The l 1 -norm penalty expedites sparsity of x and leads to a subset of assets receiving zero weights. Such sparsity may result in under-diversification and extreme weights of the portfolio. On the contrary, convex norm ball does not produce sparsity but it can efficiently regularize size of portfolio weight vector. Thus, the norm ball constraint used by DeMiguel et al. (2009) can function as a solution to alleviate the problems of underdiversification and extreme weights of the portfolio aside estimation risk. Following DeMiguel et al. 's (2009) work and specifying the general squared l 2 -norm under no short sale constraint, we propose to formulate the portfolio weights by and where τ 1 and τ 2 are tuning parameters controlling sparsity and stability respectively, with �x� 2 2 = x ′ x as the squared l 2 -norm of a vector. We estimate tuning parameters τ 1 and τ 2 by a method of cross-validation. We perform cross-validation for various possible values of the parameters and select the parameter value that produces the minimum crossvalidation average error. A combination of the l 2 -norm penalty and l 1 -norm penalty is referred to as elastic net (Zou and Hastie 2005).

Necessary and sufficient conditions for optimal problems
In this section we would like to identify the necessary and sufficient conditions for optimality of problems (17) and (18). We investigate the Karush-Kuhn-Tucker (KKT) conditions for these problems under the assumption of normality and study whether the constrained problems have optimal solutions.

KKT conditions for optimal problem
The KKT conditions provide necessary conditions for a point to be optimal point for a constrained nonlinear optimal problem. The system has a unique solution i.e.

From the budget constraint
Thus, RSMt can be represented as Similarly, RSMVt can be represented with addition of x T Qx to the objective function.
Let x be a regular point for the problem RSMt. Then the point x is a local minimum of f subject to constraints (20) if there exists Lagrange multipliers 1 , 2 and 3 for the Lagrangian function L = f 1 (x) + 1 g 1 (x) + 2 h 1 (x) + 3 g 2 (x) such that the following are true.
(19) (20) is linear and the functions g 1 and g 2 are convex, then the feasible region = {x : h 1 , g 1 , and g 2 } is a convex set. On the other hand, f 1 is a convex function subject to the variable x. We see that any local minimum for problem (20) is a global minimum too and KKT conditions are also sufficient.
Similarly, the above KKT conditions holds for RSMVt for when f 1 (x) in (20) is VaR α (x)−R + τ 1 �x� 1 + τ 2 �x� 2 2 and when the following are true where Remark 2 Since the feasible region = {x : h 1 , g 1 , and g 2 } in (20) and the objective function f 2 (x) are convex, then the feasible region is a convex set. We can see that the KKT conditions are also sufficient and any local minimum for problem (20) with objective function f 2 (x) is a global minimum as well.

Data and models
In this section, we use optimization models RSMt (17) and RSMVt (18) to construct optimal portfolios and evaluate out-of-sample performance using stocks traded on New York Stock Exchange (NYSE). Historical daily returns of 500 randomly selected stocks over the period January 2003 to December 2015 were extracted from Yahoo Finance. The risk-free rate is proxied by the 6-month US Treasury Bill rate. The selection criterion of the random sampling of stocks is based on stocks being traded throughout the evaluation period. We also use S&P 500 index daily stock price data to test the robustness of our results over the same time period even though 21 % of randomly selected stocks are components of S&P 500. We chose a short term period as distribution of stock prices tend to change shape over time. To solve the portfolio strategies, we consider the lower bound of mean return as threshold return level. We compute initial positions x 0 k , k = 1, . . . , N for constructing portfolios for January 2006 to December 2008 and January 2009 to January 2015 by solving classical Roy safety-first optimization problem: and by setting x 0 =x * , with x * denoting the optimal solution of the above model.
We employ a 6-month rolling estimation window for parameter estimations and construct portfolios on a sub-sample periodical basis. To obtain the portfolio for January January 2009 to December 2015. Portfolio revision is made on a monthly basis since in reality they are related to low transaction, management and monitoring costs.

Evaluation criteria
With regards to the return that a portfolio can achieve, we calculate and present the annualized out-of-sample Sharpe ratio with or without transaction cost. Sharpe ratio is defined as the ratio of the expected excess return to standard deviation of portfolio return (Sharpe 1966). The expected excess return is the difference between the return of the portfolio and the return obtained through a risk-free security. Mathematically, we can define Sharpe ratio as SRatio = ρ−ρ f e σ , where ρ f is the return from the risk-free security (6-month US Treasury Bill rate) and σ is the standard deviation of portfolio return. Regarding portfolio risk, we consider the risk reduction which is defined as the ratio of portfolio risk measure from the portfolio strategies (17) and (18) to that from (7) and (8) respectively. The other portfolio selection strategies in literature's risk reduction are estimated as the ratio of portfolio variance to that from Markowitz's mean-variance framework.
Sparsity may result in under-diversification and extreme weights of the portfolio. We therefore present the average number of a subset of assets with non-zero weights. We consider another performance metric known as portfolio turnover. A portfolio turnover measures the frequency with which assets in this case stocks are bought and sold. We employ the measure used by DeMiguel et al. (2009) and Kourtis et al. (2012) by defining the turnover rate of the portfolio between t to t + 1 as where x k,t+1 is the portfolio weight for stock k at t + 1, x − k,t+1 is the portfolio weight before revision at t + 1, T − T − 1 represents the length of the non-zero elements in total portfolio return and N, the number of stocks.
The introduction of transaction costs affects the overall profitability of a portfolio strategy. In practice, they lessen the net returns and diminish capital available for future

Symbol Description
RSMt l 1 -Squared l 2 penalized mean-Roy safety-first portfolio RSMVt l 1 -Squared l 2 penalized mean-variance-Roy safety-first portfolio minVu Sample minimum variance without short sale constraint portfolio minVc Sample minimum variance with short sale constraint portfolio MV Sample mean-variance portfolio 1/N Equally-weighted portfolio l 1 -MV l 1 penalized mean-variance portfolio with short sale constraint (Brodie et al. 2009) TMN Linear combination of sample tangency portfolio, sample minimum variance portfolio and 1/N portfolio (Tu and Zhou 2011) VD Minimum variance portfolio resulting from using a diagonal covariance matrix (Kirby and Ostdiek 2012) investments. We assume proportional transaction cost for the purpose of this study (Please refer to Assumption 1 for more details). We investigate the impact of transaction cost on our portfolio strategy by computing Sharpe ratio with transaction cost. In particular, the optimal portfolios can be obtained by solving problems (17) and (18).
To highlight the effect of transaction costs, we consider two situations with transaction costs, c b k = c s k = 0 and c b k = c s k = 0.02.

Computational results
Tables 2 and 3 shows the annualized out-of-sample metrics for different periods of each portfolio considered in this study: RSMt is the regularized mean-safety first portfolio with transaction costs, RSMVt is the regularized mean-variance-safety-first portfolio with transaction costs. The others are selected alternative portfolio strategies in literature. Analyzing portfolio risk, the regularized portfolio rules have lower risk than other portfolio selection strategies. This can be attributed to the downside risk measure, Roy safety-first principle considered in the regularized strategies. The variance and the safety-first principle as a consolidated risk measure in RSMVt have lower risk reduction compared to RSMt.
The portfolio turnover indicates how frequently assets in a portfolio are bought and sold. This performance measure is preferred to be small. In terms of turnover, the naive 1/N portfolio has the lowest turnover. The two portfolio selection strategies considered in this paper have relatively low turnover rates with RSMVt the preferred choice. The portfolio rules in selected from literature have a high turnover rate as compared to the regularized models in this paper. The highest turnover comes from TMN, MV and minVu portfolios leading to smaller Sharpe ratios. We observe that with large turnovers, feasible transaction costs lowers the monetary gains of many selected portfolios strategies in literature, as seen by Sharpe ratio deductions after transaction costs have been considered.
The Sharpe ratio allows investors to analyse risk-adjusted returns in exchange for the level of risk they are assuming. The higher the Sharpe ratio, the more returns the investor gets per unit of risk. The lower the Sharpe ratio, the more risk the investor bears to get more returns. Comparing all strategies in this study, the regularized portfolios have the highest Sharpe ratios with RSMVt leading the way. With regards to sparsity, more than 30 % of stocks are selected by RSMt and RSMVt in all the evaluation periods. The l 1 -penalized mean-variance model selects at least 50 % of the stocks across the evaluation periods. The smaller set of sparse portfolio optimizes the budget allocation by focusing on stocks believed to foster diversification.
From Table 2, the sample-based portfolios i.e. minVu, minVc, MV and TMN perform worse due to a large number of stocks that increases the degree of estimation risk or over-fitting. Apart from VD, all other non-regularized portfolios selected from literature (minVu, minVc, MV, TMN) perform worse than 1/N regarding both Sharpe ratio and turnover. We observe an increment in Sharpe ratio, lower risk reduction, lower turnover for post financial crisis sub-sample periods. Among the selected alternative portfolio strategies considered in this paper, l 1 MV has a better out-of-sample performance than minVu, minVc, MV, 1/N, TMN and VD. In comparing periods 2006-2008to 2009-2015 we observe that during the financial crisis period, the Sharpe ratio is lower and the risk reduction is much higher with turnover ratio also higher. This amounts to lower returns during 2006-2008 as compared to the other periods. The Sharpe ratio is much higher and the risk reduction is lower in periods 2009-2011 and 2012-2015 as compared to periods 2006-2008 and 2009-2015. In all the sub-sample periods 2006-2008, 2009-2011 and 2012-2015 including the period 2009-2015, our proposed models RSMt and RSMVt have better out-of-sample performance than the selected alternative models.
To further gain financial insights, we use S&P 500 index data from January 2009 to December 2015 as a benchmark portfolio. With a starting wealth value of $1, we compare the cumulative wealth of portfolio strategies RSMt, RSMVt and l 1 MV to that of S&P 500 index. To provide evidential proof, Fig. 1 plots the cumulative wealth of the portfolios strategies relative to S&P 500 benchmark.
In Fig. 1, $1 is used as initial wealth and it grows at a monthly return of the portfolio selection strategies considered. The figure shows distinctly the higher performance of the regularized portfolio over S&P 500 benchmark.

Conclusion
In this paper, we seek near-optimal sparse and stable portfolios to reduce the difficulty of portfolio management. Theoretical results are established to guarantee the stability and sparsity of our novel portfolio strategies. Computational evidence indicates that l 1 -squared l 2 penalized mean-Roy safety-first portfolio and l 1 -squared l 2 penalized meanvariance-Roy safety-first portfolio are able to choose optimal sparse and stable portfolios while maintaining satisfactory out-of-sample performance.
We compare the performance of our proposed models (l 1 -squared l 2 penalized mean-Roy safety-first portfolio and l 1 -squared l 2 penalized mean-variance-Roy safety-first portfolio) of optimal asset allocation relative to selected alternative portfolio strategies in literature (minVu, minVc, MV, 1/N, l 1 MV, TMN and VD). Our results show that our regularized proposed models have a better out-of-sample performance with high Sharpe  Fig. 1 Cumulative wealth of portfolio strategy RSMt, RSMVt and l 1 MV relative to S&P 500 portfolio ratios and relatively low turnover rates. Except for VD, the Sharpe ratio of 1/N portfolio when compared to other selected non-regularized portfolio rules considered in this paper is higher, which shows that estimation errors in returns shrink gains from other selected classical portfolio strategies in literature. The norm penalty improves Sharpe ratio and turnover. As a result, l 1 MV has a better out-of-sample performance than minVu, minVc, MV, 1/N, TMN and VD.
To gain more financial acumen, we compare our proposed models and the best performing portfolio strategy among the selected models from literature considered in this paper, to a benchmark, S&P 500 index. The results indicate that given an initial wealth of $1, the excess returns from l 1 -squared l 2 penalized mean-variance-Roy safety-first portfolio is the highest. Our proposed models for optimal asset allocation are favourable since they overcome unsteady and extreme portfolio weights induced by estimation error due to parameter uncertainty.