Excel MPT
Excel MPT
Excel MPT
The Optimal Portfolio Problem Consider the dilemma an investor faces when trying to decide what proportion on his/her wealth should be allocated to the different type of available investments on the market place (i.e. Canadian bonds, Canadian equity, U.S. equity, Real Estate etc.). Historical experience has shown that each of the above stated types of investments exhibit different risk/return characteristics under different economic conditions (see Appendix B). Adding to the complexity of this problem, the above investments also exhibit correlation to one another. An investor may ask the following question How can I minimize portfolio risk for a level of acceptable return?.
Mean-Variance Optimization (MVO) Method MVO, as developed by Markowitz, aims to find the set of optimal portfolios (efficient frontier), which have the highest rate of investment return for a given level of risk (i.e. standard deviation of returns) or the lowest level of risk for a given rate of return. To accomplish this task, the MVO method requires three inputs which can be calculated from the historical returns of the assets being considered for optimization. These inputs are: returns on each investment (ri), standard deviation of returns on each investment (i) and covariance (ij) between the investments. It is important to mention that standard deviation is not the only measure of investment risk studied by investment practitioners, but MVO requires this statistic as input. Other types of risk are tracking error (standard deviation of excess returns versus a benchmark), Value at Risk (VaR) etc. I will discuss some pathologies of MVO related to Parametric versus Numerical VaR.
Mean-Variance Optimization (MVO) Method The main assumption underlying MVO as developed by Markowitz is that returns are Normally distributed. If this is the case, then making use of i as the risk measure in MVO makes sense because the occurrence (probability) of outperformance versus the mean investment performance is exactly offset by the occurrence of underperformance versus the mean performance. Historical experience has demonstrated however that investment returns are not symmetrically distributed (see Appendix C) and as such MVO has been the subject of debate amongst investment practitioners.
Mean-Variance Optimization (MVO) Method The following is the MVO problem: Minimize the portfolio variance i.e. min (t Subject to:
t T
) -------(1.0)
(1.1)
is the nxn
of investment returns, rT is the total portfolio return and n is the number of different investments in the optimization algorithm.
MVO Sample calculation Consider the following two asset MVO problem: An investor wants to build a portfolio of stocks and bonds that exhibits the lowest possible portfolio variance. To simplify calculations further, assume that all of the investors wealth must be invested (without borrowing more!). Let b be the percent of the investors wealth allocated to bonds (1b to stocks!). Let b2 and s2 be the variance of returns of bonds and stocks respectively and let the covariance of returns between these two investments be bs. Making use of (1.0) on the previous page yields the following function to be minimized: f(b)=b2(b2-2bs+s2)+b(2bs -2s2)+s2
MVO Sample calculation Taking the derivative of f(b) with respect to b and equating to zero yields the following minimum variance portfolio (this would be the first portfolio on the efficient frontier): b = -(bs-s2)/(b2-2bs+s2) bs, s2 and b2 are known inputs. We know that the above portfolio is the global minimum variance portfolio because MVO is a strictly convex quadratic programming problem due to the covariance matrix being positive definite. Now as you can see adding more investments to this algorithm and adding other constraints makes for a computationally rigorous problem. Luckily many mainstream software packages have built in quadratic programming solvers. My model makes use of the built in solver within Microsoft Excel, which makes use of the Simplex method to solve the MVO problem (see Appendix E).
MVO Microsoft Excel Optimizer The following quote encapsulates why I built an efficient frontier builder (which I called the Optimizer) in Microsoft Excel Necessity, the mother of invention (George Farquhar (1678-1707)). Current providers of MVO do not calculate key statistics that investment practitioners want to examine and are very expensive. My Optimizer finds the maximum portfolio return then finds the minimum variance portfolio and the return associated with it. These two returns are then subtracted and then divided by 100. This calculation in essence is the different levels of return for which the optimizer will then find the associated minimum variance portfolio. The optimizer then stores these minimum variance portfolios in a spreadsheet in excel to which essential portfolio statistics are calculated and descriptive graphs are built. An easy to use interface was created so that even a novice user of Microsoft Excel can generate efficient frontiers.
The Optimizer
Covariance Matrix Return STDEV SC91TBIL SCUNOVER S&P/TSX SC91TBIL 4.85% 0.44% 0.00% 0.01% 0.01% SCUNOVER 8.91% 5.08% 0.01% 0.26% 0.24% S&P/TSX 9.07% 16.80% 0.01% 0.24% 2.82% SP500C 11.68% 13.90% 0.01% 0.10% 1.64% MSEAFEC 6.54% 14.35% 0.00% 0.02% 1.49%
10
12.0%
E B
9.0%
8.0%
7.0%
6.0% A 5.0%
A portfolio is considered efficient if it is not possible to obtain a higher return without increasing standard deviation.
8.0%
Return Volatility
4.0% 0.0%
2.0%
4.0%
6.0%
10.0%
12.0%
14.0%
16.0%
SC91TBIL SCUNOVER S&P/TSX SP500C MSEAFEC Total Equity Total Foreign Equity Total Fixed Income Expected Real Return Return Volatility VaR (Absolute) VaR Confidence Min VaR Confidence Max
A 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 100.0% 4.85% 0.44% -$56 -$57 -$55
B 0.0% 69.0% 0.0% 31.0% 0.0% 31.0% 31.0% 69.0% 9.77% 5.94% -$195 -$210 -$184
C 0.0% 59.1% 0.0% 40.9% 0.0% 40.9% 40.9% 59.1% 10.04% 6.81% -$212 -$229 -$200
D 0.0% 49.2% 0.0% 50.8% 0.0% 50.8% 50.8% 49.2% 10.31% 7.82% -$232 -$251 -$217
E 0.0% 39.4% 0.0% 60.6% 0.0% 60.6% 60.6% 39.4% 10.59% 8.94% -$253 -$274 -$236
F 0.0% 29.5% 0.0% 70.5% 0.0% 70.5% 70.5% 29.5% 10.86% 10.12% -$275 -$299 -$256
G 0.0% 19.7% 0.0% 80.3% 0.0% 80.3% 80.3% 19.7% 11.13% 11.35% -$298 -$325 -$277
H 0.0% 9.8% 0.0% 90.2% 0.0% 90.2% 90.2% 9.8% 11.41% 12.62% -$322 -$352 -$298
0 0.0% 0.0% 0.0% 100.0% 0.0% 100.0% 100.0% 0.0% 11.68% 13.90% -$346 -$379 -$320
11
Pathologies: Parametric versus Numerical VaR Calculation Value at Risk (VaR) is a statistic that aims to quantify the maximum amount of wealth an investor is likely to lose over a given amount of time at a specific confidence level. There are two common ways in portfolio theory to calculate VaR. There is the Parametric VaR (typically making use of Standard Normal quantiles) and numerical VaR (making use of historical Profit/Loss histograms). Consider an investor who now wants to build the following portfolio: 5% Cash (SC 91 Day T-Bills) 45% Bonds (SC Universe Bonds) 20% Canadian Equity (S&P/TSX Composite) 15% U.S. Equity (S&P 500) 15% International Equity (MSCI EAFE)
12
Pathologies: Parametric versus Numerical VaR Calculation The 95% Parametric annualized VaR (Absolute) of this continuously rebalanced portfolio (under Normal distribution assumption i.e. assumption in MVO) considering the last 5 years of historical monthly returns with $1000 initial wealth (W) is then: VaR (Absolute) = -W-W = -4.15%(1000)1.645(8.11%)(1000) = -$175 (i.e. 1 in 20 years we can expect to lose more than $175) The 95% confidence interval for this VaR is [-$204,-$155] Now let us examine the difference between Parametric and Numerical VaR considering the above portfolios performance over the last 5 years ending December 31, 2002.
13
Pathologies: Parametric versus Numerical VaR Calculation Below is the Profit/Loss histogram generated using portfolio monthly returns for the last 5 years ending December 31, 2002.
Profit / Loss Histogram
18 16 14 12
10 8 6 4 2 0
Bin -84.17 -63.17 -42.17 -21.17 -0.17 20.83 41.83 62.83 83.83 More
93 03 9 93 03 9 62 .8 33 83 .8 33 M or e
Frequency Cumulative % 1 1.64% 0 1.64% 2 4.92% 9 19.67% 13 40.98% 16 67.21% 15 91.80% 5 100.00% 0 100.00% 0 100.00%
Frequency
69 61
69 61
69 61
69 61
96 05
93 03 9
-6 3. 16 60
-4 2. 16 60
-8 4. 16 60
-0 .1 66 06
-2 1. 16 60
.8 33
20
Bin Levels
14
As you can see the above returns are not Symmetrically distributed. The loss level that delineates the lowest 5% of losses from the other profit/loss levels is approximately -$42.2, therefore the numerical annualized VaR is: Numerical VaR = -$42.2(12)^0.5 = -$146 (i.e. 1 in 20 years we can expect to lose more than $146) This VaR lies outside the Parametric VaRs confidence interval and hence evidence of one of the pathologies of MVO.
41
.8 33
93 03 9
12.0%
-$100
-$200
9.0%
8.0%
-$250
7.0%
-$300
10.0%
12.0%
14.0%
16.0%
Portoflio
VaR (Absolute)
A 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 100.0% 4.85% 0.44% -$56 -$57 -$55 B 0.0% 69.0% 0.0% 31.0% 0.0% 31.0% 31.0% 69.0% 9.77% 5.94% -$195 -$210 -$184 C 0.0% 59.1% 0.0% 40.9% 0.0% 40.9% 40.9% 59.1% 10.04% 6.81% -$212 -$229 -$200 D 0.0% 49.2% 0.0% 50.8% 0.0% 50.8% 50.8% 49.2% 10.31% 7.82% -$232 -$251 -$217 E 0.0% 39.4% 0.0% 60.6% 0.0% 60.6% 60.6% 39.4% 10.59% 8.94% -$253 -$274 -$236 F 0.0% 29.5% 0.0% 70.5% 0.0% 70.5% 70.5% 29.5% 10.86% 10.12% -$275 -$299 -$256 G 0.0% 19.7% 0.0% 80.3% 0.0% 80.3% 80.3% 19.7% 11.13% 11.35% -$298 -$325 -$277
SC91TBIL SCUNOVER S&P/TSX SP500C MSEAFEC Total Equity Total Foreign Equity Total Fixed Income Expected Real Return Return Volatility VaR (Absolute) VaR Confidence Min VaR Confidence Max
15
VaR 95% Confidence Min = -W-1.645Ws((n-1)/2(0.975))1/2 VaR 95% Confidence Max = -W-1.645Ws((n-1)/2(0.025))1/2
16
17
18
20%
15%
10%
Performance
5%
0%
-5%
-10%
-15%
-20% Less than 0% SC 91 T-Bills SC RRB 0% to 1% SC Universe 1% to 3% S&P/TSX S&P 500 3% to 5% MSCI EAFE
Inflationary Environments
19
Frequency
50 40 30 20 10 0
-717 -667 -617 -567 -517 -467 -417 -367 -317 -267 -217 -167 -117 -67 -17 33 83 133 183 233 283 333 383 433
Bin Levels
20
Efficient Frontier
3.0%
2.0%
1.0%
Expected Surplus (% of Assets)
0.0%
-1.0%
-2.0%
-3.0%
-4.0%
10.0%
12.0%
14.0%
16.0%
21
Appendix E - Simplex Method for the MVO problem Before solving the MVO problem with the simplex method (1.1) has to be rewritten in equation form by introducing slack variables (suitable positive numbers) to transform the inequality constraints to equalities. The new minimization problem is then: min(t-tr+mt(-+s2)+nt(-+1))---------(1.2) where:
s is a vector of slack variables m>=0 & n are vectors of Lagrange multipliers for the inequality and equality constraints respectively. 1 is a column vector of ones
Now note that (1.2) is unconstrained and therefore the sufficient condition for its minimum is that all of its partial derivatives must be equal to zero. Performing these partial derivatives leads to a new set of equations to which is applied the simplex algorithm.
22
Appendix F Accuracy and Computing Time Computing the accuracy of the optimal portfolio over different levels of portfolio return manually with the simplex method to ascertain accuracy would be quite rigorous. Instead when building the Optimizer I compared the optimal portfolios to those obtained with Ibbotson Encorr Optimizer, a widely used efficient frontier software. For identical levels of return, both models posted identical optimal portfolios (at least up to the third decimal place). Computing time for the Excel Optimizer is approximately 7 seconds on a Pentium IV 1.8 GHz with 256 MB or RAM. Ibbotsons product will generate an efficient frontier in less than 1 second. However from the input phase to efficient frontier generation my Excel based Optimizer is much simpler and faster to use.
23