Part I – Market Risk Modelling (70 marks)

GARCH model and Value-at-Risk (Review Week 4 Practice Note)

1. Get access to Yahoo Finance and download five-year historical daily prices of a stock or stock
index.
2. Divide the sample into two sub-samples: in sample for model estimation, and out-of-sample
for forecasting. The length of the out-of-sample is two weeks (10 trading days)
3. (10 marks) Calculate the logarithm returns for the stock. Analyze the stock return
characteristics, including, return distribution, serial correlation, and volatility clustering for the
in-sample period.
The chosen company which stock is going to be leveraged for performing tasks relating to the
GARCH model and Value-at-Risk in the Market Risk modelling section will be Synopsys, Inc.
(“SNPS” or “the Company”), with the time horizon for the five-year historical daily prices is from
01-Jan-2017 to 15-Dec-2021 as the in-sample period and from 16-Dec-2021 to 30-Dec-2021 as
the out-of-sample period. After computing the returns of the stock prices, below are the key
characteristics of the returns from the in-sample period.
 Return distribution:
The Jarque-Bera statistic of APPL’ stock price returns is 1152.581 which is different from 0
indicating that the data was not normally distributed. In particular, returns of APPL’ stock price
is left-skewed with a negative skewness statistic of -0.256, indicating the stock price was
growing at a fairly moderate level during the last 5 years. Furthermore, the distribution of the
stock price’s returns had heavier tails than a normal distribution (i.e., 7.68), indicating APPL’
stock price was volatile and sensitive to changes in the market, which is represented via the
Company’s slightly high beta of 1.22.

 Serial correlation:
As indicated in the summary of results from correlogram table below, it can be observed that
there is a serial correlation among APPL’s square stock price returns which is represented
through either the high Q-Stat or the p-value being very close to zero for all lags. Each lag
represents a joint test to examine whether at least one of the previous lags or the current one
is significant to auto correlation (“AC”) (i.e., at least one pair of returns from lag 1 up to lag n
experienced AC).

In addition, the AC effect can be captured using the ARMA model. For this particular case, the
ARMA (2,2) model was employed to significantly capture the AC effect as shown in the table
below.
  Volatility clustering:
To test whether the returns of SNPS’s stock price experienced volatility clustering (i.e., ARCH
effect), we can either use the correlogram function to check for the AC effect of the squared
returns or make analysis based on a graph of the returns of APPL’ stock price.
As a result, the returns of the stock price experienced the ARCH effect. Similar to testing for
auto correlation, p-values of all lags indicates rejecting the null hypothesis of there is no ARCH
effect. Furthermore, the graph of APPL’s stock price returns also indicates the volatility
clustering effect as shown below, especially during the first quarter of 2020 when the COVID-19
started to spread.
4. (10 marks) Based on analyses in step 3, build an appropriate GARCH model for the return
volatility using the in-sample period and explain why the model specification is appropriate.
(Note that: you may not need to capture the serial correlation of the return level using ARMA
model, but if you do so, you can earn the reward mark of 5 marks). Coi cách explain GARCH
model

From the analyses of the ARCH effect above, a GARCH model was built to capture such effect,
using the student t’s distribution since the data series was not normally distributed. As a result,
the built model was appropriate for further use in later sections since it successfully captured
both the AC and the volatility clustering issue of the Company’s stock price returns as shown
below.
Second, we can look at the ARCH LM test (say choose lag 5, that is, we jointly test for the serial
correlation from time t-5 up to now)
Prob is also high, indicating the joint test shows evidence of insignificant serial correlation in the
squared
standardized residual.
In summary, the GARCH(1,1) has captured the heteroskedasticity efficiently.

5. (10 marks) Fix the model specification you found in Step 4, fix the size of the in-sample period
as the window size, using 1-step (day) rolling window method to re-estimate the model, make
1-step ahead forecasts of the volatilities, and calculate the 99% Value-at-Risk for the out-of-
sample period.
After using the rolling window method to forecast for the estimation of GARCH, the 99% Value-
at-risk (“VAR”) would then be computed as equals to -2.33*GARCH estimation. Below is the
table of the results of calculating the VAR.
 Out-of-sample GARCH 99% Value- Observed z-
period Date forecast at-risk returns Hits value
1 12/16/2021 0.00002 0.0% -2.4% 0 2.33
2 12/17/2021 0.00002 0.0% -1.4% 0
3 12/20/2021 0.00002 0.0% -0.5% 0
4 12/21/2021 0.00003 0.0% 3.7% 1
5 12/22/2021 0.00004 0.0% 1.6% 1
6 12/23/2021 0.00005 0.0% 0.4% 1
7 12/27/2021 0.00003 0.0% 2.6% 1
8 12/28/2021 0.00004 0.0% -0.9% 0
9 12/29/2021 0.00006 0.0% 0.3% 1
10 12/30/2021 0.00007 0.0% -0.7% 0
Hit rate: 50%

6. (5 marks) Graph the calculated 99% Value-at-Risk and the observed returns in one figure,
calculate the hits rate (hits =1 if a daily observed return is greater than the calculated 99%
Value-at-Risk and hits =0 otherwise; and ℎ𝑖𝑡𝑠 𝑟𝑎𝑡𝑒=Σℎ𝑖𝑡𝑠𝑁𝑡=1𝑁) and comment about the
accuracy of the model.
99% Value-at-Risk
5.0%

4.0%

3.0%

2.0%

1.0%

0.0%
1 2 3 4 5 6 7 8 9 10
-1.0%

-2.0%

-3.0%

99% Value-at-risk Observed returns

From the graph and the table above, we can state that 50% of the time, the returns of SNPS’
stock price will fall below the threshold of 0.0% with a 99% confident level. However, we cannot
indicate that the forecast model is insufficient since there are only 10 observations. With this
model, financial practitioner can better monitor their portfolios. As such, to enhance the
performance of this model, more data and stress tests are needed to truly reflect its
performance.
Dynamic Conditional Correlation – Bivariate GARCH model (Review Week 5 Eviews Practice
Note)
1. Using one pair of data (including the spot and futures price) from the given data file on
stream (file name: Data for assessment 2). Look at the first tab of the file for the assigned data.
2. (15 marks) Estimate their dynamic conditional correlation using the DCC-MGARCH (1,1)
model.
The target commodity is sugar.
Before performing the GARCH model, we need to similarly test for the ARCH effect. With the
result table as below, we can observe that there is ARCH effect in returns of spot prices (“spot”)
since all the p-values up to the 10th lag are statistically significant

For the future price (“future”), there is no ARCH effect in the first lag of returns for future price
since prob. > 1%. However, there is ARCH effect in returns of future prices from the 2 nd lag since
all the p-values from that point are statistically significant
After performing the GARCH (1,1) model, we can observe that such model did not capture the
ARCH effect for returns of spot. Therefore, the GARCH (2,2) model was performed which
resulted in the ARCH effect being successfully captured.
On the other hand, the GARCH (1,1) model successfully captured the ARCH effects for returns of
future as shown below

After performing the GARCH models for returns on spot and future prices of sugar, the
standardized residual of spot and future (z1 and z2, respectively), the starting values of variance
and covariance for spot and future will then be generated using the “Generate series” function
in eviews. Next, the Maximum Log-Likelihood Estimation was applied to obtain the estimated
theta 1 and theta 2, and the DCC was generated via creating a LogL object.
As a result, we can observe that being a consumer staple good, which is necessary for
consumers’ daily lives, most of the DCC would fluctuate below 1, indicating that the market
projects the prices of such goods will either always grow or be maintained at the spot price
even if the market is experiencing crisis.
3. (10 marks) Based on the estimated dynamic conditional correlation (𝜌𝑡) and estimated
standard deviation of the spot (𝜎𝑆,𝑡) and futures (𝜎𝐹,𝑡), calculate the dynamic optimal hedge
ratio as follows:
ℎ𝑡∗=𝜌𝑡*(𝜎𝑆,𝑡/𝜎𝐹,𝑡 )
 Similar to the DCC, the optimal hedge ratio also indicates that being a consumer staple
goods which is less likely to be impacted by distress in the market, the ratio would hover
around the level above 0 and around 1.
4. (10 marks) Identify at least one variable that significantly drives the dynamic optimal hedge
ratio (Hint: to prove the significance, you can use the OLS regression with dependent variable is
the optimal hedge ratio)
Part II – Credit Risk Modelling (30 marks)
1. Employ the Mortgage.sas7bdat and lgd.sas7bdat datasets downloaded from
http://www.creditriskanalytics.net/datasets.html
2. Build the best Probability of Default model you could and assess its performance. (15 marks)
Above are the results from estimating the Probability of Default (“PD”) model. As we can see,
all the explanatory variables are significant at 1% level of significance. In addition, all the
variables are appropriate in determining the dependence of default time with rationales as
follows:
 FICO_orig_time has an appropriately negative coefficient, indicating that the lower a
company’s FICO score (a credit rating measurement for mortgage), the higher chance
such company will default on the borrowings.
 LTV_orig_time has an appropriately positive coefficient, indicating that the higher a
company’s Loan-to-Value ratio the higher the chance such company will default on the
borrowings since the higher the loan amount compared to the value of the appraised
property the harder a company be willing to recover the property back.
 gdp_time has an appropriately negative coefficient, since the better the economy is
performing, the lower the chance a company will go default on its loans.
 Interest_Rate_orig_time has an appropriately positive coefficient, indicating that the
higher the prime interest rate, which results in higher borrowing rate, the more difficult
it is for firms to meet their debt obligation, hence the default issues.
  hpi_orig_time has an appropriately positive coefficient since the higher the housing
price, the higher the chance of default on loans since it is more difficult for citizen to
meet their obligations.

3. Build the best Loss Given Default model you could and assess its performance. (15 marks)