Final Project Report IE590: Financial Engineering Seongjin Shin | Jish Bhattacharya Tae Yong Yoon | Luis Zertuche August 05, 2014 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Abstract: Conditional Value at Risk, an extension of Value at Risk, is fnancial mathematical tool aimed at quantifying and reducing possible investment losses. The purpose of this project is to study Value at Risk and Conditional Value at Risk as a portfolio risk metrics in management and asset selection. A practical approach was taken to understand the topic. First a basket of stocks was selected with diversifcation in mind: stocks from diferent types of industries, verifed to be negative-to-posively correlated. For the selected stocks, a statistical validation of the expected return was preformed based on a two-year data set. On diferent one-year data set for the same assets, validated by the previous step, a Monte Carlo one-day price simulation was performed. In order to estimate price fuctuations, the Geometric Brownian Motion model was used as the Monte Carlo sampled function. Value at Risk and Conditional Value at Risk were estimated and compared from the simulation data; these results were used convert the portfolio selection into a linear programming optimization problem, with the efcient portfolio frontier as the fnal result. 1. Introduction Events like the great recession of 2008, the dot-com bubble of 2000, a political upheaval or weather-related disasters are forcing investors to give more attention to risk management. Many investors do not know exactly what is in their portfolios, and more important, how those assets work or do not work together by virtue of their covariance structure. For example, consider all the unintended risk an investor is incurring when assembling a portfolio composed from a few diferent but correlated mutual. What the investor may not realize is that all funds contained a considerable amount of the same stocks. Instead of diversifying risk, the investor has concentrated it. A key and intuitive way to mitigate risk incurred in investment is by diversifying a portfolio by including a large number of assets in it, while ensuring that we dont incur punitive costs of maintaining such a large portfolio. Finding out the proportions of our budget to invest in each asset can be a daunting problem if not precisely characterized with the aid of quantitative tools. Another equally important and challenging problem is how to characterize the risk of such kind of portfolio. The ideas stated above are what !"#$$% $F I&'(!TRI)% E&*I&EERI&*
1 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering we considered an interesting for an applied project: We were keen on applying any learned concepts to real data using both quantitative and computational tools. 2. Data Collection We started by assembling a tentative portfolio with fve technology stocks and four stocks from energy and soft-drinks sectors, aiming at diversifcation. We verifed such diversifcation by means of ensuring the correlation coefcients across our 9 stocks ranged from negative to positive. Daily closing prices were collected for these 9 stocks in two separate data sets: A two-year data set for statistical parameter validation and, one- year data set for the simulation and risk optimization; year 2011-2012 and year 2013 respectively. The collected stocks are comprised of: AAPL, INTC, MSFT, SAP, COKE, WMT, XOM, BRK-A, and DDD. 3. Statistical Validation The (1 ) * 100% confdence interval estimate for the mean with standard deviation unknown is: - tn-1 S/(n) + tn+1 S/ (n) where tn-1 is the critical value of the t distribution with n 1 degrees of freedom for an area of /2 in the upper tail. Two-sided 95% confdence intervals on the mean daily return for the portfolio of 9 stocks were computed in Minitab. As per the Figure 1, the baskets of 9 stocks have a mix of positive and negatively correlated stocks for a well-balanced portfolio. As noted from the matrix, the tech stocks are somewhat positively correlated which raises the overall Value at Risk for the portfolio. Having energy and Retail stocks in the portfolio, hedges the overall risk to some extent. From the Figure 2 with 95% confdence, we can conclude that the mean daily return falls with the 95% Confdence interval and 0 is included in the range. Moreover, the p-value for all the stocks in t-tests is greater than a value of 0.05 or level of signifcance. With a sample size of 252, the normality assumption is not overly restrictive and the use of t distribution is likely to be appropriate. Since the data is normally distributed, the expected value of sample Standard Deviation follows: !"#$$% $F I&'(!TRI)% E&*I&EERI&*
+ Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering E(s 2 ) ~ 2 We use the expected return of 9 stocks from this validation and supply as a parameter to Monte Carlo Simulation program. Figure 1: Correlation Matrix of the Portfolio (Daily Closing Price 2013) Figure 2: Sample Data Analysis with 95% Confdence Interval !"#$$% $F I&'(!TRI)% E&*I&EERI&*
, Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering 4. Theoretical Background and Methodology We established a data processing stream that takes as n number of stocks input for any time interval and produces an efcient frontier of optimal asset weights for a target rate of return. Before detailing the implementation steps we will describe the quantitative tools and concepts that went into the scripting of our data pipeline. Geometric Brownian Motion for Price Estimates First conceived in early twentieth-century France by fnancial theory pioneer and mathematician Louis Bachelier, Geometric Brownian Motion is a formalism that attempts to model fuctuations in asset prices as a stochastic process, akin to the random motion of gas particles. In essence the Brownian Motion equation tries to account for change in price over a given period of time, by mixing what is currently known about the stock price with a random process. The known or certain component is called drift (d) and is calculated using the expected value of the stocks price historical data ( data ), eroded by half of the historical variance ( 2 data ): d = data
2 data /2 The random or uncertain component of the price change is modeled as a sample from a target distribution of prices. What distribution to choose, its implications and how to properly characterize it, is a whole other topic beyond the scope of this project; we will briefy touch back on it when we elaborate on some of the limitations of this model. For this report and data pipeline, we choose to use normally-distributed asset prices as a working and still-useful assumption. When put together, both the certain and uncertain components result in the GBM equation for the daily price: !"#$$% $F I&'(!TRI)% E&*I&EERI&*
4 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Where is the number of time-steps to be simulated, and the function generates a pseudo random number sampled from the Standard Normal Distribution. Value at Risk and Conditional Value at Risk Value at Risk (VaR) is a metric of the incurred risk of an investment expressed as the most likely maximum loss of a portfolio or an asset give a confdence interval (CI) and time horizon. Put in other words, VaR multiplied by the capital in the investment allows investors to calculate the most probable amount of money they would lose within the defned time horizon. An intuitive to way to think about VaR is in terms of its position as in the distribution function of returns; a distribution of rate of returns serves this purpose too, if we are interested in the negative rate of return instead of the actual amount of lost dollars. In such scenario of a distribution of prices or rate of returns, VaR will be the frontier line that delimits the CI expressed as the area under the curve of the probability density function. Fig. 1) Value at risk for a Confdence Interval of 95% It is implied by fgure 1 that if we were to numerically recreate the distribution as histogram, produced either by a Monte Carlo simulation (as in this report), or from measured price data, the VaR can be interpreted as the quantile in which the !"#$$% $F I&'(!TRI)% E&*I&EERI&*
5 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering accumulated histogram area-under-the-curve has reached the CI percentage. This is true if we approach the VaR line frontier from the right, or if approached from the left, it corresponds to the (1 CI) value. This kind formulation of the VaR allows for an easy computation when following a numeric approach; one simply needs to sum the areas of the extreme histogram bins and stop when this value is equal to either CI or (1 CI). This is expressed analytically as: With as probability density function or histogram of the rate of return. From this VaR defnition of risk one apparent problem becomes obvious: Although the VaR frontier quintile is the most likely monetary loss or negative rate of return, there are other less likely scenarios with much higher losses; basically any point to the left of the VaR frontier falls in such category of less likelihood coupled with higher loss. This specifc limitation can partially surpassed if somehow we take into account information about the whole loss tail after the VaR frontier to obtain a measure of risk. This is what is achieved by the Conditional Value at Risk (CVaR) metric. CVaR can be conceptualized as the expected value of the portfolio or asset losses given that we have surpassed the VaR frontier in the distribution. Formally, is the conditional probability of the expected value of the loosing tail given that we have exceeded the VaR quantile in the distribution: Where, r is the portfolio or asset rate of return. From the point of view of the histogram of the probability function, and relevant to the computational implementation, CVaR can be seen as the inner product between the vector of number of elements in each bin of the histogram losing tail, and the matching vector of rate o returns, converted into probability by dividing over the total number of elements in the !"#$$% $F I&'(!TRI)% E&*I&EERI&*
- Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering histogram; put another way, CVaR is the weighted sum of rate of returns from the negative infnity to -VaR, each one weighted by its respective probability. Consequently, since CVaR takes all the possible losing-scenarios, it is a more conservative metric than VaR, as illustrated in fgure 2. Fig. 2) VaR and CVaR for a Confdence Interval of 95% One can realize intuitively that CVaR is more sensitive to extreme events. For example if we were computing a distribution with high kurtosis (with fat tails) our CVaR value would move more and more to the left of the VaR quantile line. Summary of Data Pipeline Using the validated statistical parameters for the stock average daily return and standard deviation, a Monte Carlo simulation was used to estimate daily stock prices. Using such simulated data, Value at risk and Conditional Value at Risk were estimated for each stock; subsequently mean-C-VAR was used as objective function in a portfolio optimization problem to select optimal weights and plot the efcient frontier. 5. Simulation Results Below we see a comparison between individual-stock Value at Risk and Conditional Value at Risk as negative rate of returns. Under the VaR and CVaR frontiers we plot the !"#$$% $F I&'(!TRI)% E&*I&EERI&*
. Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering actual Monte Carlo simulation data condensed in a distribution histogram with 1000 bins as a design parameter. Figure 3: VaR and C-Var over their respective RR distributions from Monte Carlo samples of daily prices !"#$$% $F I&'(!TRI)% E&*I&EERI&*
8 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Figure 4: Individual Value at Risk vs. Conditional Value at Risk RRs We then formulated our portfolio weight selection problem as an optimization scheme that minims -CVaR. We used linear programming over various target portfolio rate of returns to obtain the efcient frontier of minimum portfolio CVaR. The target portfolio rate of returns ranged from the minimum to the maximum individual rates of return for our basket of stocks. Such process is summarized in the following linear programming in standard form: minimize [CVaR] T w i subject to [r i ] T w i = r p-target w i = 1 w i 0 Where r p-target is the portfolio target rate of return, w i is the vector of portfolio weights and ri is the vector of individual rate of returns. Then, using this formula, we repeated the optimization step varying r p-target from min (r i ) to max (r i ) to construct an efcient frontier (Figure 5). !"#$$% $F I&'(!TRI)% E&*I&EERI&*
9 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Figure 5: C-VaR Optimization and Efcient Frontier We learned from reviewing the literaturethat CVaR has superior mathematical properties versus VAR. C-VAR is a so-called coherent risk measure; for instance, the C-VAR of a portfolio is a continuous and convex function with respect to positions in instruments, whereas the VAR may be even a discontinuous function. This is what made the above optimization step easy and convenient. 6. Discussion Ways to Improve Statistical Validation When evaluating parameters such as mean and sigma for estimators, there is still a risk of over ftting on the test set because the parameters can be tweaked until the estimators performs optimally. This way the knowledge about the test set can interfere with model and evaluation metrics. To address this problem, yet another part of dataset can be held out as validation set; training proceeds on the training set, after which evaluation is done on the validation set, and when the experiment seems to be successful, fnal evaluation can be done on the test set. !"#$$% $F I&'(!TRI)% E&*I&EERI&*
10 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Cross Validation using k-fold design can solve this problem. The training set is split into k smaller sets and the following procedure is carried out for each of the k- folds: A model is trained using k-1 of the folds as training data The frst partition of the data is treated as test dataset. The model is ft to all the remaining partitions. The resulting model is validated on the remaining partitions of the data. The performance measure reported by k-fold cross-validation is then the average of the values computed in the iteration. This approach can be computationally intensive, but does not waste too much data as it is the case when fxing an arbitrary test set, which is a major advantage in problems where the number of samples is very small. Discussion on Portfolio Risk and Measures to Mitigate The basic principle of portfolio selection is to increase the expected return on their portfolios and to reduce the standard deviation of the return. A portfolio that gives the highest expected return for a given standard deviation, or the lowest standard deviation for a given expected return, is known as an efcient portfolio. To work out which portfolios are efcient, an investor must be able to state the expected return and standard deviation of each stock and the degree of correlation between each pair of stocks. For any individual stock there are two sources of risk. First is the risk that stems from the pervasive macroeconomic factors. This cannot be eliminated by diversifcation which is called as systemic risk. Diversifcation eliminates specifc risk, and diversifed investors can therefore ignore it when deciding whether to buy or sell a stock. The expected risk premium on a stock is afected by factor or macroeconomic risk; it is not afected by specifc risk. The expected risk premium on a stock depends on the expected risk premium associated with each factor and the stocks sensitivity to each factors (b1, b2, b3, etc.). Expected risk premium can be expressed as: r rf = b1(r factor1 r f ) + b2(r factor2 r f ) + + bn(r factorn r f ) !"#$$% $F I&'(!TRI)% E&*I&EERI&*
11 Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering A diversifed portfolio that is constructed to have zero sensitivity to each macroeconomic factor is essentially risk-free rate of interest. If the portfolio had a higher return, investors could make a risk-free proft by borrowing to buy the portfolio. If it ofered a lower return, one can make an arbitrage proft by selling the diversifed zero- sensitivity portfolio and invest the proceeds towards T-bills. In our study, we noticed that some stocks will be more sensitive to a particular factor than other stocks. Exxon Mobil would be more sensitive to an oil factor than say, Coco-Cola which is more dependent on underlying corn prices and supplies. If oil factor picks up unexpected changes in oil prices, will be higher for Exxon Mobil. To mitigate risk, we diversifed our portfolio initially from technology to energy and soft drink stocks and noticed that the VaR and average return got adjusted with positive and negatively correlated stocks. Ideally, it would have been better to add more stocks to the portfolio from additional sectors such as alternative energy, retail, construction and automotive sectors as such. However, to keep our study in scope, the portfolio was restricted to nine stocks. Distribution Assumptions for VaR and CVaR From inspecting the histogram plots with VaR and CVaR, its easy to realize that these measures are very sensitive to the distribution we assume our price data to have. One can imagine how the kurtosis and skewness of the distribution have considerable impact, as both would change the amount of area under the tails of our histogram plot. As the tail-ends of the distribution become larger, we will consistently see our CVaR increase. This is also a good way to illustrate why investors conceived CVaR over VaR. Consider the distributions below in which both have the same VaR but diferent CVaR: !"#$$% $F I&'(!TRI)% E&*I&EERI&*
1+ Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering Figure 6: Same VaR, diferent C-VaR due to nature of distributions. The most striking implication that follows from such idea is that we might consistently under estimate or overestimate the risk in our portfolio just by assuming an inappropriate price distribution. 7. Conclusions We created a data fow that takes daily stock prices as raw data input and validates their statistical parameters based on a larger data set for the same stocks. Then we used the validated parameters to simulate one-day fuctuations in price using a Monte Carlo simulation, based on the Geometric Brownian Motion model of pricing. From the Monte Carlo samples, we estimated the future one-day price of our stocks. We also estimated VaR and CVaR risk metrics and optimized CVaR to obtain portfolios that lie in an efcient frontier. We learned about the importance of risk measures as key quantitative information for any investor. Specifcally about VaR and CVaR, we learned that a striking limitation of such models is distribution. A valuable insight was the realization that risk metrics are contingent upon the parameters and data that go into them. Two diferent indicators might be in disagreement, or be too liberal or too conservative, depending on what information we feed to the models. Knowing what is inside the black-box of a risk !"#$$% $F I&'(!TRI)% E&*I&EERI&*
1, Measures of Financial Risk and their usage in Risk Management Team 14 08/05/14 IE590 Financial Engineering metric model is the best way a rational investor can make the most informed decisions about a portfolio or asset. . As parallel outcome of this project, we increased our knowledge about robabilistic reasoning, linear optimization and fnancial computational tools, which are extremely valuable tools for investors given the enormous growth of information available nowadays. 8. References Valdemar Antonio, D. F. Portfolio management Using Value at Risk: A Comparision Between Genetic Algorithms and Particle Swarm Optimization. Erasmus Universiteit Rotterdam, 67. Luenberger, D. G. (1998). Investment science. New York: Oxford University Press. Shephard, N. Markov chain Monte Carlo methods for stochastic volatility models. Journal of Econometrics, 281-316. Uryasev, S. Portfolio optimization by minimizing conditional Value at Risk via nondiferentiable optimization. Computational Optimization and Applications, 391-415.