Chapter 3 Methodology
3.2 Defaultable stock valuation
Chapter 3 Methodology
3.1 Defaultable zero-coupon bond
Unlike Treasury bonds, corporate bonds and common stocks are subject to the credit risk. So, we introduce the reduced-form approach of credit risk to value these securities. Also, we assume that a probability is given, which is the risk-neutral measure. Following Duffie and Singleton (1999), framework, the defaultable zero-coupon bond is expressed as
],
3.2 Defaultable stock valuation
Traditionally, in most literature, the normality assumptions are popularly applied in pricing derivatives, where the Brownian motion is adopted. The stock price subjected to the credit risk is assumed to follow a geometric Brownian motion (GBM). Let the stock returns can be described as follows:
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
where t is the expected instantaneous stock returns, is the standard deviation of instantaneous returns and is the standard Brownian motion under the physical measure. Thus, we can describe the dynamics of stock price under the risk-neutral measure as follows:
P
Wt
(3) where is the standard Brownian motion under the risk-neutral measure,
), )
exp((
0
Q t GBM
t t
t S r t W
S
Q
Wt GBM is the
compensator term of GBM model which makes sure that the stock price is under the risk-neutral measure, i.e. 2.
2 1
GBM
3.3 The dynamics of intensity rate
We know that when the stock price of firm is low, it tends to default. Thus, the intensity rate has negative effect on the stock price. Following Takahashi, Kobayashi and Nakagawa (2001), the intensity rate is the decreasing function of stock price, which is easily observed and traded frequently in the financial market. Also, Ayache, Forsyth and Vetzal (2004) considered the intensity rate increases as the stock price decreases. Hence, the intensity function is set as
( , ) ( ) ( )( ) ,
0 0
S
S S S
t
St t t (4)
where )(S0 is the estimated intensity rate at St S0. In Muromachi (1999), the formal function like Equation (4) was found to fit the bonds rated BB+ and below in Japanese market. Moreover, the values of are reasonably suggested range from –1.2 to –2.0.
‧
3.4 The properties of Lévy process
Under the GBM model, the stock return follows a normal distribution. However, as mentioned in previous section, the return distributions of financial securities are always not normally distributed but exhibit fat tail and excess kurtosis. Hence, we will use the Lévy process to characterize the dynamics of stock price. A Lévy process is under the probability space , where is the sample space, F is the information filtration, and P denotes the physical measure, it has the following properties:
Xt
According to Lévy-Khintchine formula, we have
determine the distribution type of . Hence, we will use the Lévy process to characterize the dynamics of stock price. In the late 1980s and the 1990s, such as the normal inverse Gaussian (NIG)
Xt
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
distribution (Barndorff-Nielsen, 1995), the variance Gamma (VG) distribution (Mandan, Carr and Chang, 1998) were widely applied.
In the Lévy process, the stock price under the risk-neutral measure can be described as follows:
St S0exp((rt t)tXt t), (5) where Xt is the Lévy process, is the compensator term of Lévy model which ensures that the
dynamics of stock price are under the risk-neutral measure.
Next, we introduce some properties of the NIG and VG model. Barndoff-Nielsen (1995) proposed the normal inverse Gaussian (NIG) model, which is an inverse Gaussian time-changed Brownian motion. Also, the NIG(,,,) with parameters 0, , 0 and
R
has a characteristic function given by
NIG(x,,,,)exp(ix)exp(( 2 ( ix)2 2 2)), (6) where )( is the characteristic function.
If we have the inverse Gaussian distribution I(, 2 2), the NIG model can be obtained by replacing the time t in the Brownian motion with a inverse Gaussian distribution. Hence, we can simulate XtNIG as follows:
1. generate I(, 2 2).
2. generate ) W ~ N(0,1.
3. set XtNIG I(t) I(t)W.
Hence, the NIG model can be represented as follows:
XtNIG I(t) I(t)W, (7)
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
where is a location parameter, is the parameter and follows the inverse Gaussian distribution.
) I(t
In the NIG model, there are three parameters, which are the shape parameter , the skewness parameter and the excess kurtosis parameter In this case, the compensator term of the NIG . model is
NIG 2 (1)2, where 2 2.
Madan, Carr and Chang (1998) suggested the variance Gamma (VG) model, which is obtained by Brownian motion at a random time change given by a Gamma distribution. This model has two additional parameters determining the skewness and kurtosis which are and respectively.
Also, the characteristic function of the
v )
, , ,
( v
VG distribution is given by
) .
2 1 1
( ) exp(
) , , , ,
(x v ix ixv 2x2v 1v
(8)
If we have the Gamma distribution G(tv,v), the VG model can be obtained by replacing the time t in the Brownian motion with a Gamma distribution. Hence, we can simulate as follows:
VG
Xt
1. generate G(tv,v).
2. generate ) W ~ N(0,1.
3. set XtVG G(t)W(G(t)).
Hence, the VG model can be represented as follows:
‧
Moreover, Madan, Carr and Chang (1998) stated that the VG model could be expressed as the different parameters of two independent Gamma distributions. In this case, the compensator term of the VG model can be shown as
In addition, if we have the characteristic functions of these distributions, the basic moments can be obtained, i.e., mean, variance, skewness and kurtosis.
3.5 The estimation method
We discuss the method of parameter estimation for the NIG and VG model. We use the maximum likelihood estimator (MLE) to estimate the parameters of NIG and VG distributions. If we have the probability density functions of NIG and VG distributions, we can obtain the maximum log likelihood functions of NIG and VG model. The probability density function of NIG distribution is described as follows (Barndorff-Nielson, 1995):
,
The probability density function of VG distribution is described as follows (Madan, Carr and Chang, 1998):
K1
‧
Then, we can obtain the log likelihood function of NIG and VG model. They can be described respectively as follows:
where and denote the likelihood functions of NIG and VG distributions
respectively. In this application, the initial values of the parameters are set from the moment ()
LNIG LVG()
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
method.
3.6 The valuation of convertible bond
The value of CB can be affected by many factors, such as the stochastic process of underlying stock price (denoted as ), the convertible price (denoted as ) and the credit risk of issuing company. In addition, there are many CBs’ provisions that influence the value of CB such as buy-back, sell-back, adjustment or reset of the conversion price. Brennan and Schwartz (1980) found that the value of CB is not significantly different under the assumption of constant or stochastic interest rates. Hence, for simplicity, we assume that the interest rate is constant. In addition, since the credit events have occurred frequently and the complex provisions such as sell-back and buy-back are popular in the market, we consider these conditions in measuring the value of CB.
S C
rt
We assume that the CB is a zero-coupon bond, in which, the par value is denoted as M. We propose that the valuation is proceeded under the probability space (,F,Q), where is the state; the filtration is the information process; and is the risk neutral probability.
Let be the price of CB at time t with maturity date T,
T t
Ft
F ( )0 Q
) c
, ( Tt
CB , s and b be the converted,
sell-back and buy-back time respectively. Moreover, we assume min(c,s,b) is the terminated time of CB. The terminated time is decided according to the converted time c, sell-back time s and buy-back time b, these three values can be determined based on the comparison of the holding value and terminated value of CB.
‧
Based on the provision of conversion, the payoff of CB is if the CB is converted. Based on the provision of sell-back, the investors can sell the CB back to the issuers on the value of
M , where is a constant extra premium before a predetermined time. Also,
the issuers can buy the CB back from the investors on the value of M(1). Furthermore, if the CB is still remained until maturity, the value of CB is M at maturity date. Therefore, the payoff of CB at maturity date T can be shown as
The rational investors will terminate the CB if they expect the terminated value is larger than the holding value. The holding value and terminated value of CB can be calculated based on the generating stock price. In other words, after generating the stochastic process of underlying stock, we can determine the times of c, s and b, then evaluate the CB at maturity by Equation (11).
The first term in Equation (11) is the payoff when conversion happens, and the second term is the payoff that the investor sells back the CB to the issuer. The third term in Equation (11) is the payoff that the issuer buys the CB back from the investor. The last term is just the payoff that the investor holds the CB until maturity date. From the Equation (11), the value of CB at time under the risk-neutral measure can be described as
t
‧
3.7 Least squares Monte Carlo algorithm
The embedded call option of CB is an American-style, hence we must deal with the early-exercised problem. The least squares Monte Carlo Simulation (LSM) method proposed by the Longstaff and Schwarz (2001) is to evaluate the expected continued holding value compared with the conversion value and then to determine the optimal excise time at each simulated path, i.e., the investor may not convert the CB even if the stock price exceeds the convertible price. That means the investor expects the holding value is more than conversion value. In the LSM method, we assume the investor is allowed to convert the CB into common stock in points, , the continued holding vale of CB is defined as the expected discounted value of cash flow at time conditional on the time information. It can be estimated with a
least squares regression. For example, the conditional holding value is viewed as , where is the stock price, investor makes decision to hold the CB or convert the CB.
tk
The conditional excepted value is obtained by the backward method at each . When the CB is in the money at time , we assume the continued holding value of CB, denoted as , can be calculated by the regression, where is the path of stock price. Since the
continued holding value depends on the expectation of remaining discounted cash flows with respect to the risk-neutral measure , the holding value at time can be expressed as
1
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
k
where is the risk-adjusted rate between time and , is the remaining discounted cash flows of each path between time and . Based on Equation (13), the holding
value is conditional on the known information at time ()
R tj tk C(w,tj,tk,T)
tj tk
tk
Ft .
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y Chapter 4 Empirical results
4.1 Data description
In this subsection, we describe five examples of CBs containing provisions or not and show why we adopt the Lévy model. Since the trading isn’t frequently in Taiwan CB market, we select the samples that contain relatively higher trading volume. There are five samples in this study, including Foxconn, Uni-President, the first issue of Synnex (denoted as Synnex 1), the second issue of Synnex (denoted as Synnex 2) and Yuanta respectively. The stock prices of these samples are obtained from the Taiwan Economic Journal (TEJ). The sample period of these samples orderly begins from November 10, 2006 to May 10, 2007, September 24, 2008 to March 23, 2009, June 25, 2008 to December 12, 2008, January 14, 2011 to July 15, 2011 and July 9, 2010 to December 24, 2010. All of the sample period is 120 trading days. The difference between five samples is the contract that Uni-President and Synnex 2 are three years without additional provisions and the others are five years with complex provisions such as sell-back and call-back. Table 1 briefly describes the properties of these CBs.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 1: Convertible bond
Convertible bond Foxconn Synnex 1 Yuanta Uni-President Synnex 2
Date of issuing 2006/11/10 2008/6/24 2010/7/7 2007/10/25 2011/1/14 Duration 5 years 5 years 5 years 3 years 3 years
Issuing price ＄100 ＄97.7 ＄103 ＄103 ＄103.15
Coupon rate zero zero zero zero zero
Initial conversion price ＄316.55 ＄80 ＄20.8 ＄56 ＄86.8 Provision sell-back, buy-back sell-back, buy-back sell-back, buy-back none none
Premium of buy-back 0% 0% 0% 0% 0%
Premium of sell-back 0% 0% 0% 0% 0%
Interest rate 2.2% 2.6% 1.1% 2.8% 1.1%
Note: This table briefly describes the content of contracts for five samples.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Many empirical studies in finance shows the return distributions of securities often have fat tail and excess kurtosis. These samples also have the properties. Moreover, the time period of return data begins from the whole year before the day that starts to evaluate plus the trading days. If the empirical returns exceed two standard deviations of returns, it may imply the returns of securities containing jump components. Figure 1 shows this phenomenon. Therefore, the CBs are suitable for applying the Lévy model, not the GBM model.
Figure 1: The returns of all samples
4.2 The goodness of fit test
In order to know how the fitted density adapts to the empirical data, in this subsection, we adopt two methods such as QQ plot and two-sample Kolmogorov–Smirnov (two-sample KS) to test
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the goodness of fitted distribution for the empirical data. In the QQ plot test, according to the Glivekno-Cantelli theorem, let a sample be , and its cumulative probability density function be
yn
y y1, 2,, )
, (yi
F , furthermore, the ordered sample Y(1),Y(2),,Y(n) satisfy the condition:
limmax|( 0.5) ( (); )| 0.
1
i
N i
N i N F Y
Also, the figure for the sequential pairs are based on the sequence of {((i0.5) N),F(Y(i);))}Ni1.
It is so called as the PP plot which is an approximately 450 straight line. Moreover, the figure depends on the following sequence called QQ plot that is also an approximately 450 straight line as {(Y(i),F1((i0.5)/N);))}iN1.
QQ plot is a probability plot, which is a graphical method for comparing two probability distributions by plotting their quintiles against each other. If the plotted figure is deviated from the 450 straight line too far, which means that the empirical data do not fit the assumed distribution well.
The QQ plots of GBM, NIG and VG model are shown in figure 2. All of the data display the GBM model can’t capture fat tail well. In addition, the VG model of Synnex 1 also can’t capture this phenomenon. Even so, the empirical results show that the Lévy model is better than the GBM model in the fitness.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Figure 2: The QQ plot of all samples
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
In the description of two-sample KS test that is used to determine whether the empirical and fitted distribution differ significantly. Let be the cumulative probability density of empirical distribution, be the cumulative probability density of fitted distribution, the two-sample KS test can be calculated as
Fe
Ff
KS max|Fe(x) Ff(x)|.
R
x
This statistical test is based on the difference between the cumulative probability density of empirical distribution and the cumulative probability density of fitted distribution.
Table 2: KS two-sample test of five samples
KS two-sample test GBM Model NIG Model VG Model
Foxconn 0.0791 0.0627 0.0655 Uni-President 0.0741 0.0391 0.0511
Synnex 1 0.0765 0.0540 0.1092
Synnex 2 0.0647 0.0360 0.0575
Yuanta 0.0820 0.0632 0.0729
Note: The test is all at 10% significance level.
As shown in Table 2, the values of two-sample KS test for Foxconn are 0.0791, 0.0627 and 0.0655 in the GBM, NIG and VG model at 10% significance level respectively. Similarly, the values of two-sample KS test for Uni-President are 0.0741, 0.0391 and 0.0511 in the GBM, NIG and VG model at 10% significance level respectively. Also, the others are shown in Table 2. In GBM model, the cumulative probability density of empirical distribution doesn’t fit the cumulative probability density of fitted distribution among these samples. In the Lévy model, except the VG model of Synnex 1, the cumulative probability density of empirical distribution fits the cumulative
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
probability density of fitted distribution well. In sum, whatever QQ plot or two-sample KS test, we find that the goodness of Lévy model is superior to the GBM model. Then, we estimate the parameters of these three models in the next subsection.
4.3 Parameter estimation of stock price and calibration of intensity rate
Before pricing CBs, we have to estimate the parameters of GBM, NIG and VG distribution by using the maximum likelihood method described in Equation (10). Moreover, if we have the parameters of Lévy and GBM model, then the parameter of intensity rate can be estimated by calibration. Thus, before pricing CBs, the parameter is calibrated by minimizing the difference between the market prices and the model prices given by
, ) (
min
arg stMarket stModel 2
where is the market CB price of each trading day, also, is the model CB price of each trading day.
Market
st stModel
The estimated parameters of Foxconn, Uni-President, Synnex 1, Synnex 2 and Yuanta are listed in Table 3. In Foxconn, the estimated values of GBM parameters and are 0.0024 and 0.0212 respectively. The estimated values of NIG parameters , , and are 0.0027, 47.7661, -2.1822 and 0.0235 respectively. The estimated values of VG parameters , , and are 0.0039, 0.0219, 0.5411 and -0.0023 respectively. Also, the parameters of intensity rate in the GBM, NIG and VG model are -1.0671, -1.9530 and –2.3513 respectively. Other samples are also listed in Table 3. In Table 3, we find that the parameters of intensity rate are negative. Most of these values are between -1 and -2. Moreover, in the VG model, control the skewness. We find that
is almost zero, it means the distribution is symmetric.
‧
Table 3: Estimated parameters in different models
Foxconn Uni-President Synnex 1 Synnex 2 Yuanta
Notes: The parameters of GBM, NIG and VG model are estimated by MLE. The parameter of intensity rate is estimated by the calibration. (‧) means the standard error of the parameter.
‧
4.4 The performance of convertible bond
In this subsection, we display the performances of five samples with different contracts. After estimating the parameters, the LSM algorithm is applied to evaluate the theoretical value of CBs.
Moreover, if the model prices of CBs are obtained, the performance index, which is root mean square error (RMSE), can measure the performance of these prices. In this index, it gives the larger weight for the larger deviations, and can be calculated as follows:
1 ,
where is the number of trading days. Also, the pricing errors are defined as the deviation between market and model prices. It can be measured as follows:
n
The empirical results of five samples are displayed in Table 4. The RMSE values of Foxconn are 0.0536, 0.0496 and 0.0487 in the GBM, NIG and VG model respectively. In addition, the RMSE values of Uni-President are 0.1197, 0.1151 and 0.1072 in the GBM, NIG and VG model respectively. Also, the other samples are displayed in Table 4. In this table, we can find that both the performances of VG and NIG model are better than the GBM model, in which the VG model has the best performance except Synnex 1. It is shown that the NIG model is better than the other models.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Table 4: Compariative performance of five samples
RMSE GBM Model NIG Model VG Model
Foxconn 0.0536 0.0496 0.0487 Uni-President 0.1197 0.1151 0.1072
Synnex 1 0.1478 0.1401 0.1467
Synnex 2 0.0482 0.0459 0.0367
Yuanta 0.1079 0.1015 0.0877
Note: In the performances of these five examples, the Lévy model is better than the GBM model.
Moreover, the pricing errors and SK ratios are shown in figure 3, where is the stock price and
S
K is the exercised price. In addition, the solid and dash lines represent the pricing errors and SK ratios respectively. We find that most of our samples show CBs are out-of-money, the pricing errors are positive like in Carayannopoulos (1996).
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Figure 3: The pricing errors of all samples
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y Chapter 5 Conclusions
In recent years, many firms gradually have increased their funds by issuing convertible bonds.
We find that the issuing volume of Taiwan’s CBs has grown up from 0.9 billion in 1999 to 290 billion in 2011. Therefore, convertible bonds as financing tools are more important for firms. The reason is that the coupon rate of convertible bonds is lower than the coupon rate of straight bonds, in this way, firms can reduce the financing cost by issuing convertible bonds instead of straight bonds. In addition, firms give investors the potential opportunity to obtain capital gains from
We find that the issuing volume of Taiwan’s CBs has grown up from 0.9 billion in 1999 to 290 billion in 2011. Therefore, convertible bonds as financing tools are more important for firms. The reason is that the coupon rate of convertible bonds is lower than the coupon rate of straight bonds, in this way, firms can reduce the financing cost by issuing convertible bonds instead of straight bonds. In addition, firms give investors the potential opportunity to obtain capital gains from