1.

‘Data’ spreadsheet: First, we started by computing simple returns of each stock with all the given
data. These returns were used to estimate (on November 13, 2017) the daily volatility and correlation for each
equity, based on the past 250 returns. Given that the number of shares remains fixed throughout the period,
in order to calculate the initial value Xi of each position, being the market closed on the 11th of November,
2017, we decided to make use of the prices of the stocks on the next trading day, (13th of November, 2017)
considering that we can’t trade at closing prices of the day before. Thus, we basically assumed that just for
the 13th of November 2017 both opening and closing prices were coincident. Daily changes in prices (returns)
reflect in everyday changes in relative weights (and in values) of each asset. The weights of stocks with the
highest return increase at the expense of the others (Buy-and-Hold strategy). Assuming normally distributed
returns and constant daily volatility and correlation throughout the period, we computed the individual VaR
(99%,1) each day until March 11, 2022, for each equity position as follows :

V aRi (99%, 1) = ασi Xi (1)

This approach is more efficient compared to computing the VaR with the Non-Parametric method (quantile
approach). Being the latter a location-based estimator depending on distribution’s order, leads to a VaR relying
on just two values! On the other hand, using a Parametric method (sigma-based approach) implies assuming a
shape for the returns distribution (in this case Normal). It increase precision, reduces uncertainty and thus the
standard error of the estimate (narrower VaR confidence interval).

2&3.‘Portfolio VaR’ spreadsheet: In finding the fund’s VaR(99%,1), each day until March 11, 2022 we
applied the Asset-Normal approach :
√
V aRΠ (99%, 1) = α x′ Σx (2)
Subsequently, the portfolio VaR was compared to the Undiversified VaR (UVaR) which is simply the sum of
individual VaR. The subtraction UVaR- VaRΠ reflects diversification benefits, which are equal to CHF 157,348
on average. Therefore, we notice that we gained advantages by reducing the overall risk of a significant amount
thanks to a not unitary (and low) correlation between assets.
Afterwards, we moved on to computing the daily portfolio profit and losses (P&L) starting from November
14, 2017 (since we bought at prices of November 13th we assumed that they were the closing prices and thus we
did not realize any P or L in that day) till the most recent observation. Comparing the actual values of daily
P&L to the VaRΠ computed the day before (it is a forecast of the next day’s worst-case scenario!) we found out
37 observations (losses) exceeding the VaRΠ , i.e. 3.47% of the overall observations. In particular, we figured out
that in 7 cases during 2022 (starting January 3) the actual loss was higher, in absolute value, than the VaRΠ (i.e.
14%). In both cases, results are not compliant with our estimations as we expected just around 1% of exceeding
losses. Likely, this mismatch is due to assumptions (normality, constant volatilities and correlations), sampling
error (we have a limited set of data) and the fact that not always historical data are good predictors of the future.

4&5.‘Analytic VaR’ spreadsheet: Further on, we proceed to compute the Component VaR (CVaR) and the
Relative Component VaR (RCVaR):
CV aRi (99%, 1) = V aRΠ βi wi (3)
CV aRi
RCV aRi = (4)
V aRΠ
By means of this measures, we are able to identify which position should be increased/decreased to reduce
our VaR the most. Plotting the chart of RCVaR in time, we can see that CFR has the higher relative (%)
contribution to overall risk with spikes of 91.79% (and never below 55.86%). Conversely, SREN stock for most
of the time contributes to lowering our Portfolio VaR as it has negative RCVaR up to –2.33% and a maximum
of 4.74%. NOVN ranges between 9.85% and 44.58% during the portfolio lifetime.
On March 11, 2022, we planned the following trades: buy 4’000’000 CHF in stock SREN, sell 1’000’000 CHF
in CFR, and sell 1’000’000 CHF in NOVN. This makes us re-evaluate the VaR using (in principle):
p
V aRΠ+a (99%, 1) = α (x + a)′ Σ(x + a) (5)
Where x is the vector of the amount of each stock at 11 March (xSREN , xCF R , xN OV N ) and a is the vector of
the incremental change in the amount in each stock (4M,-1M,-1M). However, taking a first-order approximation
lets us save time while not paying a high price in terms of error/distance with the true value. This applies only
if the amounts of the vector a are not too big compared with the portfolio size, as the first-order derivative
require. Such a method makes use of the Incremental VaR: IV aR = (∆ VaR)’·a
W here∆V aR (Marginal VaR) is computed as follows:
V aRΠ
∆V aR = βi (6)
W

1
These measures attempt to quantify the impact on the overall VaR of changing the current positions by a small
(∆V aR) or large amount (IV aR). Therefore, in the IV aR, Marginal VaR is multiplied by the amount of the
change in the portfolio given by the trade. The VaR approximation following the trade is obtained by adding
the Incremental VaR to the previous (11 March) V aRΠ . After the trade, we would have a reduction of CHF
31,420 (=-IVaR) in the fund’s VaR, thus we would have a less risky portfolio.
To evaluate the accuracy of the IVaR, we compared the results provided by the two methods. This gave
us a difference of 8,983 CHF (due to convexity not explained by the 1°order derivative), only the 3.21% of the
accurate procedure (exact relation). Therefore it strengthens what we expected: using the IVaR is faster, not
computationally long and the final result does not differ so much from the actual value, in this case it only
slightly underestimates the after trade V aRΠ !

6.‘Stocks’ Betas’ spreadsheet: We computed each stocks’ betas with respect the SMI Market factor (our
benchmark) based on the full sample of returns. Therefore, we measured the covariances between all the stocks
and the SMI and the variance of SMI. Being β the standard CAPM measure of systematic risk (or volatility)
of a security/portfolio compared to the market, it gauges the tendency of the return of a security to co-move
with the return of the stock market.
Given that, we can state that CFR has the highest systematic risk (1.303), therefore in principle, according
to CAPM theory, adding this stock to a portfolio will increase the portfolio’s risk (but may also increase its
expected return). NOVN, with the lowest systematic risk (0.962), is theoretically less volatile than the market
and including this stock in a portfolio should makes it less risky than the same portfolio without the stock.
Empirically, real data often does not support CAPM theory, as a consequence of its assumptions.

7.‘EWMA&GARCH(1,1)’ spreadsheet: Going forward, we will be looking at time-varying volatility using

two different approaches, which have the goal of modelling the tomorrow volatility considering the dependence
(in the 2nd moment) with past observations, given the mounting evidence that volatility tends to cluster. The
intuition behind these two models is to give more recent observations a higher weight than older data.
Firstly, a RiskMetrics model (EWMA) developed by JPM was utilized, where the persistence of the variance
process is governate by the parameter lambda, settled equal to 0,94.

σt2 = λσt−1
2 2
+ (1 − λ)rt−1 (7)

The resulting weighting scheme achieved by iterating backward has an exponentially decline feature, thereby
data more than 100 days old are essentially weightless.
In addition, we used a GARCH (1,1) model, which is an extension of the previous model but where we now
add a significant feature of the conditional volatility i.e. a mean-reverting behaviour. This aspect is achieved
thanks to the the presence of a new parameter γ, which pulls the variance back to its long-run mean VL = ωγ .
Particularly, the bigger γ is, the faster the volatility will be pulled back to VL .

σt2 = γVL + αrt−1

2 2
+ βσt−1 (8)

We calibrated the parameters of the model (α, β, γ) by maximizing the sum of the individual Log-Likelihood.
In this case, by looking at the two plots of the conditional Variance over time, we can notice that they are some-
how similar, but not perfectly overlapped as we can see from the difference of 0.017% between the March 11,
2022 conditional Variance forecast of the two models. We can notice an high peak for conditional Volatility
of around 4,752% (GARCH(1,1)) on 13/03/2020 probably due to the uncertainty after the beginning of the
pandemic. While the higher recorded value we have for EWMA is 3.577% on 25/03/2020 (around same period).

8.’Sharpe Diagonal VaR’ spreadsheet: In conclusion, we computed the Sharpe Diagonal VaR (99%,1)
of the portfolio on March 11 (just tomorrow’s VaR). In this case, for the value of σ 2 Π , we plugged in the
following:
2
σΠ = βΠ σm + w′ Dϵ w
2 2
(9)
We notice that in our case we couldn’t get rid of the term that captures idiosyncratic component (w’·Dϵ · w)
because we are taking into account a portfolio which is neither well-diversified nor large enough, as well as
being it characterized by an asset that due to its high weight (0,92) is dominating the others in the portfolio.
In this situation idiosyncratic shocks don’t offset each other. Precisely, we plugged in the formula above both
2
predicted σm (SMI variances from ’point 7’) as well as Dϵ and we ended up with a σΠ equal to 2.457% using the
RiskMatrics model and 2.804% using the GARCH(1,1) model, much higher values if compared with A-N σΠ
of 1.119%. These values obviously lead to a far greater V aRΠ for the last two models (numerical comparison
reported in the excel file). In conclusion, considering that as we said in the previous point volatility changes
over time and tends to cluster over periods, we could state that both EWMA and GARCH(1,1), being time-
varying volatility models, are more reliable than a model (such as the Asset Normal) where volatility is assumed
constant in time.