SP Jain FIN405 Financial Risk Modelling
Class Discussion Exercises
Topic 4 – Modelling Investment Portfolios
Cholesky Matrix
Up till now, we have worked with models with independent variables. However, this is not
realistic as asset classes, economic variables and other quantitative measures can be correlated
with each other. To address such correlation, one way is to generate correlated random
variables to reflect such relationships. One way is to use a Cholesky Matrix.
Refer to the Cholesky worksheet of the file “Investment Portfolio Model – [Link]”.
If you know how to generate a Cholesky matrix using iterative calculation, excellent! Please
send me your code sequence. If not, you may find the following website useful to compute
the Cholesky matrix from your desired correlation or covariance matrix:
[Link]
1. Explain the relationship between the correlation coefficient and the covariance.
Let ij be the covariance, ij be the correlation coefficient
ij ij i j
2. How does the correlation coefficient matrix and the covariance matrix relate to each other?
The correlation coefficient matrix multiplied by the variance diagonal matrix gives the
covariance matrix.
3. For the two random variables being generated, specify their variance and correlation
coefficients.
The diagonals of the covariance matrix are variances. Thus, the two random variables
have a variance of 2.2% and 1.6%, respectively. The correlation coefficient between the
two is -0.2.
4. Explain how the Cholesky matrix is structured and how it relates to the correlation or
covariance matrix.
The Cholesky matrix, C, is a lower triangle matrix such that product of this matrix and its
transpose, CT, gives either the correlation or covariance matrix.
5. Explain the process in how the Cholesky matrix is used to convert independent random
variables to correlated random variables.
For a vector of random variables, R, when this is multiplied by the Cholesky matrix
representing the lower triangle square root of the correlation matrix, or the covariance
matrix, it gives the vector of correlated random variables. If the Cholesky matrix is a
covariance matrix, then the vector of correlated random variables incorporate the
variances of the individual random variables.
© Copyright 2020. SP Jain Page | 20
Investment Portfolio Model Structure
This week’s exercise we will take the approach of bottom-up learning where instead of being
told the details about the model being studied, you will work this out by answering the
questions below. We will undertake the approach of working backwards in terms of looking
at the model output and the process in order to grasp the objectives of the model. Refer to the
investment portfolio model “Investment Portfolio Model – [Link]” and answer the
following questions.
1. Identify the output variables for this investment portfolio model.
The output variables are the geometric average returns of the portfolio, the total
accumulated portfolio value and whether the portfolio’s return meets a benchmark.
2. Explain briefly how the portfolio value is calculated.
The portfolio value is calculated by the sum of the values of cash, bonds, stocks and
property holdings. Those are determined by the accumulated returns of each asset class,
modelled using various continuous time returns models.
3. What is the structure of the investment portfolio? How long is the investment horizon for
this model?
The investment portfolio comprises cash, bonds, stocks and properties. The returns are
modelled on a monthly basis over ten years. The four asset classes have returns that are
correlated with each other.
4. Briefly explain what basis is used to estimate the distributions and parameters for the
asset returns for this investment model.
The asset returns follow a continuous time model. Each asset class is modelled by its
returns. The cumulative prices are calculated by assuming the returns are continuously
compounded.
For cash, the price of the cash asset is modelled using a Bachelier-Samuelson model (cover
more in detail in Topic 7). For the other asset classes, the returns are assumed to follow a
constant drift with the diffusion following a t-distribution instead. Hence, they are not
modelled using the Bachelier- Samuelson model, but with a process that has a diffusion
with a thicker tail distribution.
The parameters are estimated by taking into account the historical returns from 2013 to
February 2020, with more weighting placed on the most recent two years of history. This
is because the last two years have seen more volatile returns as a result of the central
banks initially seeking to normalise the economy by raising interest rates, but changing
course middle of 2019 as the markets showed signs of weakness.
5. Explain the process in which the correlation between different asset classes in the
portfolio is modelled.
Firstly, uncorrelated normal random variables are generated. A Cholesky matrix on the
correlation coefficients between the four asset classes is used to convert these
© Copyright 2020. SP Jain Page | 21
uncorrelated normal random variables to become correlated normally distributed
random variables. These can either be used directly for modelling cash returns, as we
assumed cash returns to follow a Brownian motion. For the other asset classes, we need
to use correlated U(0,1) random variables. This is achieved by converting the correlated
normal random variables to U(0,1) random variables by using the NORMSDIST function.
Afterwards, we can convert these to t-distribution random variables using the TINV
function so they are now correlated t-distribution random variables.
6. What may be possible reasons for using the t-distribution instead of the normal
distribution to model the returns for bonds, stocks and property?
The t-distribution has a higher kurtosis and can model more extreme observations.
Refer to the “Simulations” worksheet.
7. Explain the results for the modelled portfolio and its implications in a way that a
layperson can understand.
Out of 10 000 simulated portfolios, we expect that the portfolio will outperform against
the 6% p.a. return almost 60% of the time. The mean annual return of the portfolio is just
above 8% with standard deviation of annual returns being 5.6%.
8. A friend looks at the mean and standard deviation of the asset returns from the simulated
results and comments that these do not correspond to the assumptions in the returns
model. Explain briefly why this may be the case.
The reason why the mean asset returns for the different asset classes do not correspond
is because of random variation. However, they are not too far away from the
parameterised means. The standard deviations of the returns of the different asset classes
are smaller than the parameterised values because we have simulated many portfolios
and the annual returns over the ten years of the modelled portfolios reduces the volatility
of the annualised returns. Some years deliver higher returns than others and across
simulations, the returns are going to be closer to the mean annualised returns.
9. Suggest possible scenarios that you may model and make the appropriate amendments
to the model to generate results.
Possible scenarios include:
1. How the portfolio performs relative to a returns benchmark for different asset
allocations.
2. Determining how likely a portfolio allocation will lead to losses greater than a certain
acceptable threshold (for capital guaranteed or low-risk investment where the
manager guarantees the return of the portfolio does not fall below a threshold).
© Copyright 2020. SP Jain Page | 22
10. Identify potential improvements to the model to make it more realistic.
1. Returns for asset classes are autocorrelated – meaning that the returns are dependent
on its history.
2. Varying correlation between asset classes, including potential modelling of market
crash events where all asset classes fall together.
3. Allowing for deposits and withdrawals from the fund, rather than modelling a fixed
asset allocation.
© Copyright 2020. SP Jain Page | 23
