【AP_AOR】Review of Robust Optimization

Robust Optimization

Soyster(1973)在鲁棒优化方面进行了开创性的工作，他提出了一个线性优化模型求出一个对所有凸集数据均可行的解，但是该解过于保守，放弃很多标称问题的最优性.
Ben-Tal和Nemirovski(1998, 1999, 2000)，EI和Oustry(1998)，EI和Lebert(1997)在解决鲁棒模型保守性方面做出了贡献，他们将Robust Optimization应用到椭球不确定集合中，得到了SOCP模型.
Bertsimas和Sim(2004)提出了一种控制方法，通过生成的线性鲁棒对应来控制解的保守程度. 并且该方法可以拓展到离散优化的问题. 他们同时提出了鲁棒线性规划，鲁棒二次规划和鲁棒半定规划模型.
Ben-Tal(2002)的模型对于每个约束的保护程度提供了完全控制，当某个约束方程的系数只有少于预先指定的个数发生变化时，可以保证解是可行的，即使发生变化的数量超过这个指定数值，也能提供一个解可行的概率保证.
Bertsimas(2006)基于鲁棒优化提出了一个一般的方法来处理需求为随机的离散时间供应链最优控制的问题.

金融系统中的鲁棒优化模型

Mulvey(1995)系统分析了鲁棒优化的目标，特点和应用领域，并开启了鲁棒优化在金融领域的应用.
Zeniou(1996)建立了可赎回债券投资决策的鲁棒优化模型，Mulvey和Dawei(1997)分析了优化决策中不确定性问题的诸多解决方法，如随机线性规划，动态规划等，论证并比较了鲁棒优化在资产负债管理中的结果，并且得到了不能依靠随机线性规划解决决策风险中风险厌恶的结论.
Todd和Richard(2003)考虑了证券市场中的牛市和熊市的周期性变化，提出了鲁棒运作策略策略.
鲁棒模型中的目标函数包含了两项内容，第一项表示解的鲁棒性，充分考虑不确定因素的影响，可以使用包含均值和方差的效用函数表示；第二项则是模型的鲁棒性，对不可行解设置惩罚值.
Lobo(1998)在均值-方差框架下考虑方差不确定情况下的投资组合优化问题，并且提出了在Worst Case的情况下分析问题的鲁棒思路，他研究了最坏情况下具有二阶矩边界的方差最小化投资组合问题，并且表明在最坏条件下方差计算是一个SDP问题.
Costa(2002)在假设仅已知期望收益和协方差矩阵的情况下，研究了鲁棒跟踪误差投资组合优化问题.
Ghaoui(2003)等将最坏条件下VaR概念引入到均值-方差模型中，主要解决选择投资组合所使用的方差数据不确定导致的决策失误，并且证明了在Worst Case情况下，VaR可以通过SDP进行精确计算.
Pinara和Tutuncu(2005)提出了选择鲁棒套利机会的优化模型，主要贡献是在期望收益和标准离差数据不完全的情况下，运用多项式内点法得到凸二次规划的鲁棒解.
Popescu(2007)提出随机优化的鲁棒投资组合均值-方差方法，在资产收益向量分布为任意时，将问题转为单变量的目标函数鲁棒问题.
Bertsimas和Pachamanova(2007)研究了考虑交易成本的多阶段投资组合选择的鲁棒优化问题.
Gulpinar和Rustem(2007)提出了多阶段均值-协方差优化问题的鲁棒模型，该模型使用情景树描述不确定性，并给出了QP求解方法，结果表明，基于多阶段情景树的投资组合鲁棒优化模型可以减少由参数不确定性引起的投资决策最优值的波动，但是该研究停留在理论阶段.

目前金融鲁棒优化表现出三个特点：

1. 研究问题主要针对金融机构的实际需求.
2. 研究方法从一般的目标规划扩展到内点法，凸二次规划法，半定规划，LMI等.
3. 研究内容从静态向动态扩展，并考虑交易成本和套利机会等因素.

Recent advances in robust optimization for investment management

Paper Link
Achieving robustness aims at alllowing certain degrees of uncertainty in inputs while generating managable outputs. This goal is accomplished in robust optimization by defining uncertainty sets for representing possible input values and the robust solution is produced from worst-case or minimax approaches. Its main advantages over stochastic programming is less computational complexity because certain uncertainty sets result in robust counterparts of portfolio problems that are formulated as tractable optimization problems.

Robust asset allocation

One of the first robust approaches to asset allocation was introduced by Tutuncu and Koening(2004). They present robust formulations when the expected return vector and the covariance matrix of asset returns are defined by lower and upper bounds and also illustrate how to compute the robust efficient frontier.

Yam(2016) introduce a variation of robust model that allows short positions. They find the effect of uncertainty in expected return is more critical than the uncertainty in covariance matrix for controlling sensitivity.

The practical advantage of applying robust models to strategic asset allocation is demonstrated by Asl and Etula (2012). In their robust asset allocation approach, they performance robust optimization with input estimated from a multi-factor model, which is suitable for estimating expected returns and risk across asset classes.

The out-of-sample performance of robust portfolios is tested by Ben-Tal et al(2010), where they introduce a soft robust approach that relaxes the robustness of the standard robust optimization models. The out-of-sample results demonstrate a higher gain for the soft robust approach while only sacrificing limited downside risk.

Recchia and Scutella (2014) compare performance of several robust asset allocation strategies and comfirm empirically that the relaxed robust models of Ben-Tal (2010) show strong robustness. Moreover, they find that the classical robust model of Tutuncu and Koenig (2004), which minimizes worst-case variance, exhibits how turnover and high diversification.

Sadjadi et al (2012) derive robust formulation of portfolio optimization problems for constructing portfolios with a predefined limited number of assets when asset returns are subject to uncertainty and also develop a procedure based on genetic algorithms for finding the optimal robust portfolio.

Gulpinar (2011) analyze robust counterparts of a portfolio problem with discrete asset choice constraints that control the cardinality and buy-in threshold of a portfolio.

General robust allocation models

Deng et al (2013) also optimize the maximization of worst-case Sharpe Ratio, but use uncertainty sets of Sharpe ratio estimators. Their model selects the portfolio with the largest worst-case Sharpe within a given confidence interval, which is shown to be equivalent to maximizing the value-at-risk-adjusted Sharpe ratio based on the observation that Sharpe ratio estimators are normally distributed.

Lutgens and Schotman(2010) model uncertainty through advice obtained from mutiple expert advisors on mean vector and covariance matrices. The robust portfolio is found by evaluating the worst case among multiple recommendations. They analyze two situations where amobiguity in advice exists either only in expected returns or in both the mean and covariance matrix.

The benefits from considering a robust approach based on multiple recommendations are stressed, especially when advisor recommendations are dispersed.

Portfolio selection

Equity markets

A good number of advances in robust portfolio optimization study models for stock portfolios, especially because of the high volatility in stock markets that call for approaches.
In particular, there have been many attempts to accurately model uncertainty in stocks by incorporating attributes of stock returns such as skewness and fat tails for forming robust stock portfolios.
Kawas and Thiele (2011) introduce a log-robust model based on the log-normal behavior of stock prices. While the model based on the traditional log-normal model, the worst-case approach of robust optimization takes into account the fat-tail events that are under-represented in the traditional approach.

The log-robust approach is shown to form portfolios that are more diversified with better VaR performance compared to traditional robust portfolio approaches.

As an extension, Pae and Sabbaghi (2014) develop a log-robust optimization with transaction cost.

The asymmetry of stock returns and increased correlation during stock market downturns are addressed by Kim et al (2015). They demonstrate the value of worst-case information in the stock market for gaining robust performance by introducing a simple rule-based that focuses on worst-case returns for constructing robust portfolios.

Chen and Kwon (2012) develop a robust model for tracking a market index. Instead of explicitly limiting the tracking error of a portfolio, the model maximizes pariwise similarities between assts of the portfolio and the target index, and the robustness is increased by modeling the uncertainty in similarities through a range of possible values.

The proposed robust model allows setting the number of allowed assets and the model is formulated as a 0-1 integer program for identifying the assets included in the tracking portfolio.

A robust approach that avoids the use of uncertainty sets on expected returns is also introduced by Nguyen and Lo (2012). Ranking models have been used in portfolio selection in order to avoid estimating expected returns, and Nguyen and Lo (2012) derive a robust ranking problem that can be applied to portfolio optimization when there is uncertainty in the rankings.

Credit and bond markets

Ben-Tal et al (2010) illustrate their soft robust approach with a bond portfolio portfolio that invests in $49$ bonds in addition to one risk-free asset. The comparison suggests that the soft robust approach reduces the high conservatism of the standard robust optimization approach, which is shown to invest heavily in the risk-free asset.

The impact of credit risk model misspecification is critical for bond portfolios because estimation of default intensity is difficult due to the rarity of default events. Bo and Capponi (2016) develop a dynamic robust bond portfolio model for investing in risky bonds that is robust against misspecifications of the reference credit model, derving Hamilton-Jacobi-Bellman equations for solving their robust problem.

Currency markets

The main advantage of applying robust optimization is the increased flexibility compared to standard heding strategies that only utilize derivatives such as forwards or futures.

Forming robust international portfolios under uncertainty in returns of assets and currencies in a multi-period setting is studied by Fonseca and Rustem (2012). A comparsion between a single stage model and their multi-stage model is included to demonstrate the benefits of their approach.

Derivatives

Injecting portfolios with options for hedging purposes is a common strategy in controlling portfolio risk. Since the payoffs of options are driven by the underlying assets, uncertainty in the returns of underlying assets affects not only allocation in those assets but options as well. Thus, robust optimization has been applied to model the uncertainty in asset returns for portfolios including risky assets as well as options on those assets. In these portfolios, options may provide protection when asset returns are realized outside of uncertainty sets. Also, robust replicating portfolios are used for pricing derivatives as we discussed.

Portfolios containing derivatives

Ling and Xu (2012) assumed a factor model for modeling uncertainty in returns that defines a joint uncertainty set. Nonetheless, the objective of designing a robust portfolio with options is explored to hedge risk from extreme events that are not covered by uncertainty sets and the advantage is illustrated using simulated data and returns from the Chinese stock market.

Options pricing

Robust option pricing is studied by Bandi and Bertsimas (2014), who combine the replicating strategy with robust portfolio optimization. In particular, the worst-case replication error of a replicating is minimized, which consists of stocks and a risk-free asset.

The uncertainty in the underlying in the underlying price dynamics is modeled using polyhedral sets and this results in robust option pricing problems that are linear programs.

Conclusion

More recently, robust models have also been applied to topics of practical interest such risk budgeting (Kapsos, 2017), factor-based investing (Kim, 2017), and incorporating investor views (Hasuike and Mehlawat, 2017). As mentioned above, the worst-case approach can be applied to practical investment setting due to the flexibility of the approach. More importantly, robust optimization becomes a powerful strategy in practice because robustness is achieved without heavily penalizing the computational complexity of the problem.