Assessment 2

Part I – Market Risk Modelling

GARCH and EWMA models

1. Selected Wal-Mart Stores, Inc. stock WMT stock price from 1/22/2019 to

1/19/2024.

2. Analyze and discuss stock return characteristics

As you can see from the graph, the median is 0.000663, the maximum is 0.110723,

and the minimum is -0.120765. The mean is 0.000472, and the mean of the return

distribution is close to 0. The skewness of the return distribution is below 0

indicating that the return distribution has some long-left tail. There is a high

probability that the value is less than 0. The distribution's kurtosis is about 2, which

is very high, indicating that the yield distribution is fat-tailed. The Jarque-Bera is

very high, and the probability is close to 0, so we can indicate that the daily return

distribution is not normal.

Date: 01/20/24 Time: 18:26
Sample (adjusted): 1/23/2019 1/19/2024
Included observations: 1257 after adjustments
Autocorrelation Partial Correlation AC PAC Q-Stat Prob

1 -0.078 -0.078 7.5947 0.006

2 0.033 0.027 8.9362 0.011
3 -0.022 -0.018 9.5634 0.023
4 -0.060 -0.064 14.122 0.007
5 0.001 -0.007 14.125 0.015
6 -0.048 -0.045 17.001 0.009
7 0.031 0.022 18.226 0.011
8 -0.106 -0.104 32.371 0.000
9 0.125 0.108 52.200 0.000
10 -0.047 -0.032 55.005 0.000
11 0.027 0.016 55.938 0.000
12 0.005 -0.001 55.969 0.000
13 -0.011 0.004 56.120 0.000
14 0.033 0.020 57.520 0.000
15 -0.053 -0.034 61.045 0.000
16 0.004 -0.021 61.061 0.000
17 -0.055 -0.029 64.991 0.000
18 0.054 0.031 68.669 0.000
19 -0.017 -0.004 69.026 0.000
20 0.009 0.000 69.141 0.000
21 -0.006 -0.014 69.186 0.000
22 -0.083 -0.077 78.111 0.000
23 0.032 0.006 79.449 0.000
24 0.025 0.046 80.264 0.000
25 0.008 -0.004 80.340 0.000
26 -0.055 -0.050 84.242 0.000
27 -0.008 -0.026 84.319 0.000
28 -0.009 -0.003 84.421 0.000
29 0.029 0.030 85.508 0.000
30 -0.029 -0.046 86.583 0.000
31 -0.090 -0.084 96.995 0.000
32 0.021 0.001 97.588 0.000
33 -0.010 -0.003 97.712 0.000
34 0.052 0.037 101.20 0.000
35 -0.013 -0.007 101.41 0.000
36 -0.049 -0.061 104.51 0.000

Correlogram of R

The probability of serial correlation tends to be zero as seen in the figure. A zero

probability of serial correlation would typically mean there is no evidence of serial

correlation in the residuals of a regression model. In other words, the residuals do

not exhibit any systematic patterns or correlations over time.

Date: 01/20/24 Time: 18:20
Sample (adjusted): 1/23/2019 1/19/2024
Included observations: 1257 after adjustments
Autocorrelation Partial Correlation AC PAC Q-Stat Prob

1 0.339 0.339 144.85 0.000

2 0.233 0.133 213.25 0.000
3 0.213 0.114 270.32 0.000
4 0.065 -0.065 275.65 0.000
5 0.038 -0.014 277.45 0.000
6 0.081 0.063 285.77 0.000
7 0.061 0.031 290.50 0.000
8 0.110 0.080 305.83 0.000
9 0.142 0.072 331.44 0.000
10 0.064 -0.035 336.59 0.000
11 0.115 0.063 353.34 0.000
12 0.073 -0.007 360.10 0.000
13 0.029 -0.011 361.18 0.000
14 0.064 0.032 366.41 0.000
15 0.061 0.022 371.14 0.000
16 0.038 -0.001 372.94 0.000
17 0.053 0.002 376.47 0.000
18 0.028 -0.017 377.46 0.000
19 0.010 -0.012 377.59 0.000
20 0.014 -0.010 377.86 0.000
21 0.018 0.017 378.28 0.000
22 0.015 0.001 378.58 0.000
23 0.041 0.021 380.75 0.000
24 0.010 -0.023 380.88 0.000
25 0.030 0.018 382.04 0.000
26 -0.005 -0.038 382.07 0.000
27 -0.006 0.002 382.12 0.000
28 0.003 0.005 382.13 0.000
29 0.001 0.003 382.13 0.000
30 0.008 0.008 382.22 0.000
31 0.004 -0.011 382.24 0.000
32 -0.000 -0.010 382.24 0.000
33 0.001 0.003 382.24 0.000
34 -0.007 -0.012 382.31 0.000
35 -0.006 0.008 382.36 0.000
36 -0.017 -0.019 382.73 0.000

Correlogram of R2

The probability of fluctuating clusters is also zero. A zero probability indicates that

the data deviates significantly from the constant volatility assumption.

3.

Capturing Serial Correlation in Return Levels Using ARMA Models. The table

shows that serial correlations are generally not significant. There is no significant

relationship in the ARMA model.

Establish GRACH (1,1) model

In the GARCH (1,1) model, the results are significant. So the model is appropriate.

The Prob. F is much higher. For probabilistic reasons, we cannot reject the existing

hypothesis.
4.

Modeling an appropriate EWMA return volatility.

LogL: EWMA
Method: Maximum Likelihood (BFGS / Marquardt steps)
Date: 01/23/24 Time: 17:16
Sample: 1/24/2019 1/19/2024
Included observations: 1256
Evaluation order: By observation
Convergence achieved after 14 iterations
Coefficient covariance computed using outer product of gradients

Coefficient Std. Error z-Statistic Prob.

LAMBDA(1) 0.944437 0.000861 1097.278 0.0000

Log likelihood 9694.929 Akaike info criterion -15.43619

Avg. log likelihood 7.718893 Schwarz criterion -15.43210
Number of Coefs. 1 Hannan-Quinn criter. -15.43466

5.

Forecasted volatility for 22/1/2024.

volatility for 22/1/2024 is 0.000203.

Dynamic Conditional Correlation – Bivariate GARCH model

2.

GARCH(3,3)model

Dependent Variable: RSPO

Method: ML ARCH - Normal distribution (BFGS / Marquardt steps)
Date: 01/23/24 Time: 18:51
Sample (adjusted): 1/02/1980 9/06/2017
Included observations: 9831 after adjustments
Convergence achieved after 57 iterations
Coefficient covariance computed using outer product of gradients
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*RESID(-2)^2 + C(4)*RESID(
-3)^2 + C(5)*GARCH(-1) + C(6)*GARCH(-2) + C(7)*GARCH(-3)

Variable Coefficient Std. Error z-Statistic Prob.

Variance Equation

C 4.86E-06 1.14E-06 4.262869 0.0000

RESID(-1)^2 0.037316 0.007312 5.103658 0.0000
RESID(-2)^2 0.058285 0.015629 3.729216 0.0002
RESID(-3)^2 0.036056 0.016084 2.241783 0.0250
GARCH(-1) 0.072113 0.204770 0.352166 0.7247
GARCH(-2) 0.890039 0.018705 47.58343 0.0000
GARCH(-3) -0.116740 0.185135 -0.630565 0.5283

R-squared -0.000008 Mean dependent var 3.85E-05

Adjusted R-squared 0.000094 S.D. dependent var 0.014040
S.E. of regression 0.014039 Akaike info criterion -5.876780
Sum squared resid 1.937697 Schwarz criterion -5.871658
Log likelihood 28894.31 Hannan-Quinn criter. -5.875045
Durbin-Watson stat 1.981665

The results start to be significant at the time of the GARCH (3, 3) model.
Establish DCC model

LogL: DCC
Method: Maximum Likelihood (BFGS / Marquardt steps)
Date: 01/24/24 Time: 13:39
Sample: 1/03/1980 9/06/2017
Included observations: 9830
Evaluation order: By observation
Convergence achieved after 19 iterations
Coefficient covariance computed using outer product of gradients

Coefficient Std. Error z-Statistic Prob.

T(1) 0.066354 0.000623 106.5336 0.0000

T(2) 0.910406 0.000928 981.2527 0.0000

Log likelihood 11798.09 Akaike info criterion -2.400018

Avg. log likelihood 1.200212 Schwarz criterion -2.398554
Number of Coefs. 2 Hannan-Quinn criter. -2.399522
3.

All variables used.

std01=@sqrt(garch01)
series std02=@sqrt(garch02)
h_star=rho12*(std01/std02)

Line graph of dynamic hitch ratio

Part II – Credit Risk Modelling

2.

Use Mortgage.csv

The variables chosen are ltv_time and gdp_time.

Choose ltv_time because when that rate goes up, so does the default rate because

people don't want to repay their loans.

The reason for choosing gdp_time is that when GDP growth is higher, the

economy is doing well, and people have more money to repay their loans. Then

the default rate goes down.

Dependent Variable: DEFAULT_TIME

Method: Least Squares
Date: 01/24/24 Time: 14:41
Sample: 1 622489
Included observations: 622219

Variable Coefficient Std. Error t-Statistic Prob.

C -0.008881 0.000778 -11.42207 0.0000

LTV_TIME 0.000454 8.30E-06 54.64413 0.0000
GDP_TIME -0.003230 0.000106 -30.61263 0.0000

R-squared 0.009192 Mean dependent var 0.024353

Adjusted R-squared 0.009189 S.D. dependent var 0.154143
S.E. of regression 0.153433 Akaike info criterion -0.911096
Sum squared resid 14648.08 Schwarz criterion -0.911042
Log likelihood 283453.7 Hannan-Quinn criter. -0.911081
F-statistic 2886.308 Durbin-Watson stat 2.005558
Prob(F-statistic) 0.000000

The results show that our predictions are correct. Ltv_time is significantly positively

correlated with default time, while gdp_time is significantly negatively correlated

with default time.

Dependent Variable: DEFAULT_TIME
Method: ML - Binary Probit (Newton-Raphson / Marquardt steps)
Date: 01/24/24 Time: 14:49
Sample: 1 622489
Included observations: 622219
Convergence achieved after 8 iterations
Coefficient covariance computed using observed Hessian

Variable Coefficient Std. Error z-Statistic Prob.

C -2.581085 0.013789 -187.1847 0.0000

LTV_TIME 0.007432 0.000139 53.42333 0.0000
GDP_TIME -0.046017 0.001639 -28.07689 0.0000

McFadden R-squared 0.036602 Mean dependent var 0.024353

S.D. dependent var 0.154143 S.E. of regression 0.153586
Akaike info criterion 0.220683 Sum squared resid 14677.26
Schwarz criterion 0.220737 Log likelihood -68653.44
Hannan-Quinn criter. 0.220698 Deviance 137306.9
Restr. deviance 142523.6 Restr. log likelihood -71261.79
LR statistic 5216.697 Avg. log likelihood -0.110336
Prob(LR statistic) 0.000000

Obs with Dep=0 607066 Total obs 622219

Obs with Dep=1 15153

Use lgd.csv

Dependent Variable: Y_PROBIT

Method: Least Squares
Date: 01/24/24 Time: 14:56
Sample: 1 2545
Included observations: 2545

Variable Coefficient Std. Error t-Statistic Prob.

C -3.100729 0.023313 -133.0070 0.0000

LGD_TIME 6.355656 0.058224 109.1581 0.0000

R-squared 0.824117 Mean dependent var -1.650813

Adjusted R-squared 0.824048 S.D. dependent var 2.304122
S.E. of regression 0.966501 Akaike info criterion 2.770517
Sum squared resid 2375.480 Schwarz criterion 2.775108
Log likelihood -3523.483 Hannan-Quinn criter. 2.772183
F-statistic 11915.49 Durbin-Watson stat 1.784717
Prob(F-statistic) 0.000000
3. Find relevant literature on the probability of the default model.

Zhang, Qingfen, "MODELING THE PROBABILITY OF MORTGAGE DEFAULT VIA

LOGISTIC REGRESSION AND SURVIVAL ANALYSIS" (2015). Open Access Master's

Theses. Paper 541.

Above is the title of the articles I found.

The default probability model in the article utilizes logistic regression and survival

analysis to predict the likelihood of mortgage default. Logistic regression is

commonly used in credit risk modeling and provides a direct prediction of the

probability of default within a specific timeframe. On the other hand, survival

analysis models the time to an event and is particularly useful when dealing with

censored data, such as in the case of mortgage defaults. The model considers

variables such as the Case Shiller house price index, FICO score, months on book,

and loan-to-value ratio to estimate the probability of default. The performance

of the model was evaluated using metrics such as percent concordant, Somers' D

(Gini), c-statistic (AUC), and KS statistic, with the selected model demonstrating

high accuracy in predicting default probability. The study also highlights the

potential benefits of leveraging survival analysis in broader areas such as

profitability management within the banking industry.

While simple one-dimensional models consider only one or two variables in

predicting the likelihood of mortgage default, the model presented in the article

utilizes multiple variables to estimate the probability of default. The model

considers variables such as the Case Shiller Home Price Index, FICO score, months
on the books, and loan-to-value ratio to predict the probability of default. By

incorporating multiple variables, the model can capture the complex relationship

between the different factors that lead to mortgage defaults, thus predicting the

probability of default more accurately. In contrast, simple one-dimensional

models are unable to capture these complex relationships and are likely to result

in less accurate predictions of default probabilities.