Asset Correlation Estimation in the Vasicek Model

Maximum Likelihood: Analytical vs Numerical Optimization Approach

Andrija Djurovic

www.linkedin.com/in/andrija-djurovic

Andrija Djurovic Asset Correlation and the Vasicek Model 1/9

The Functional Form and Parameters of the
Vasicek-Distributed Variable

The Vasicek distribution is a two-parameter ( 0 < p < 1 and 0 < ρ < 1) continuous
distribution on the range 0 to 1. If a variable x has a Vasicek distribution, then x can be
represented as:  −1 √ 
ϕ (p) − ρz
x =ϕ √
1−ρ

where:
p and ρ are the parameters of the distribution, commonly referred to as the average
default rate and asset correlation, respectively;
z represents the systemic factor drawn from the standard normal distribution; and
ϕ and ϕ−1 denote the distribution and quantile function of the standard normal
distribution, respectively.

Andrija Djurovic Asset Correlation and the Vasicek Model 2/9

The Parameters Estimation Methods

The parameters of the Vasicek distribution can be estimated using one of the following
methods:
1 Direct Moment Matching
2 Indirect Moment Matching
3 Maximizing the Log-Likelihood of the Vasicek Probability Density Function
4 Quantile-Based Estimation
While each method produces nearly unbiased estimators for parameter p, this is not the
case with parameter ρ.

The most frequently used method for estimating ρ in practice is Indirect Moment
Matching. This method is also known as the analytical solution for the Maximum
Likelihood Estimator, but how does it compare with the method of maximizing the
Log-Likelihood of the Vasicek Probability Density Function?

The following slides provide a brief description of these two approaches along with
simulation results for different sample sizes and selected elements of the variable with the
Vasicek distribution.

Andrija Djurovic Asset Correlation and the Vasicek Model 3/9

Indirect Moment Matching

!
µ̂x
p̂ = ϕ p
1 + σ̂x2

σ̂x2
ρ̂ =
1 + σ̂x2

where:
PT
ϕ−1 (xi )
µ̂x is defined as µ̂x = i=1
T
and ϕ−1 denotes the quantile function of the
standard normal distribution; and
P T
(ϕ−1 (xi )−µ̂x )2
σ̂x2 is defined as σ̂x2 = i=1
T −1
with ϕ−1 being the quantile function of the
standard normal distribution.

Andrija Djurovic Asset Correlation and the Vasicek Model 4/9

Maximizing the Log-Likelihood of the Vasicek Probability
Density Function
The Log-Likelihood of the Vasicek Probability Density Function is given by:

T
X
ln(fp,ρ (xi ))
i=1

with  √ 2 
1−ρϕ−1 (x )−ϕ−1 (p)
q 1 ϕ−1 (x )2 − √
1−ρ 2 ρ
fp,ρ (x ) = e
ρ

where:

T represents the number of observations;

fp,ρ (x ) denotes the Vasicek Probability Density Function;
x represents the observed default rates;
p denotes the average default rate; and
ϕ−1 is the quantile function of the standard normal distribution.
PT
xi
With p calculated as i=1 , ρ is derived by maximizing the Log-Likelihood function based on the observed
T
default rates.
Andrija Djurovic Asset Correlation and the Vasicek Model 5/9
Simulation Setup

1 Select the inputs for simulating random data from the Vasicek distribution: sample
size (T = 5), unconditional PD (pd = 0.05), asset correlation (ρ = 0.20),
time-dependent systemic factor (z) from the standard normal distribution with an
autoregression coefficient θ = 0.30.
2 Using the simulated data, estimate the asset correlation by employing Indirect
Moment Matching and maximizing the Log-Likelihood of the Vasicek Probability
Density Function.
3 Store the results and repeat step 2 a total of N times (N = 10, 000).
4 Calculate the average asset correlation for each estimation method, then compare
these averages to the true ρ.
5 Change the sample size T from step 1, and repeat steps 2 through 4.
Practitioners are encouraged to test and simulate the bias of the estimators with other
parameters of the Vasicek-distributed variable.

Andrija Djurovic Asset Correlation and the Vasicek Model 6/9

Simulation Results
Comparison of the Asset Correlation Estimators
IMM MLE True rho

20%

19%
Average Asset Correlation

18%

17%

16%

15%

25 50 75 100
Sample Size (Number of Years)
IMM − Indirect Moment Matching
MLE − Maximizing the Log−Likelihood of the Vasicek pdf

Andrija Djurovic Asset Correlation and the Vasicek Model 7/9

Simulation Results cont.
Comparison of the Asset Correlation Estimators
IMM MLE True rho

20%

19%
Average Asset Correlation

18%

17%

16%

15%

5 6 7 8 9 10
Sample Size (Number of Years)
IMM − Indirect Moment Matching
MLE − Maximizing the Log−Likelihood of the Vasicek pdf

Andrija Djurovic Asset Correlation and the Vasicek Model 8/9

Simulation Results cont.
Distribution of the Asset Correlation per Sample Size

IMM MLE

600

400

5
200

0
600

400

6
200

0
600

400

7
200
Counts

0
600

400

8
200

0
600

400

9
200

0
600

400

10
200

0
0% 10% 20% 30% 40% 50%
rho
IMM − Indirect Moment Matching
MLE − Maximizing the Log−Likelihood of the Vasicek pdf

Andrija Djurovic Asset Correlation and the Vasicek Model 9/9