The Application of Gibbs Sampler Method to the Distribution of Asset Return Correlation in the
New Basel Accord
鍾麗英
Lyinn Chung 國立台北大學統計學系副教授
余信萱 Hsin-Hsuan Yu 國立台北大學統計學系研究生
李 詩政 Shih-Cheng Lee 元智大學會計學副教授
Abstract
The asset correlation is the key variable for calculating the regula-
tory capital in the IRB approach of New Basel Accord. However, the
range of asset correlation has suffer a lot of controversy since the Sec-
ond Consultative paper (CP2) in 2001. Then, the third Consultative
paper (CP3) announced in April, 2003 formally introduced the asset
return correlation as a decreasing function of default probability. CP3
not only defined that the range of asset correlation from 0.12 to 0.24
for corporations but also addressed a negative relationship between as-
set correlation and probability of default. As CP3 did not explain the
theoretical concept for the formula of asset correlation, there were a
lot of studies discuss the appropriateness of this formula. This study
applied Bayesian method to ASRF model to estimate the posterior
distribution of asset correlation. We use the data for the firms in the
United State from year 2001 to 2005. The empirical results suggest
that the asset correlation is a decreasing function of probability of de-
fault and an increasing function of firm size, and indicate that there
may be some important factors impact the asset correlation were ig- nored.
Keywords: New Basel Accord, asset correlation, one-factor model, Bayesian
1 Introduction
In order to cope with the keen competition of international financial institution and maintain the soundness and stability of the interna- tional bank system, the supervisor institutions of various countries es- tablished a lot of capital controlled regulation to supervise risk, among them the “Basel Capital Accord” is the most representative.
In 1988, the Basel Committee for Banking Supervision announces the
“Basel Capital Accord”, that is Basel I, which defined credit risk and the minimum amount of capital that should be held by the bank. The accord required that banks must hold a minimum of 8% capital. Be- sides, in 1996, the Basel Committee incorporated market risk into the framework. This accord was ratified by the G-10 countries and is still used today to define the minimum amount of capital a bank must hold to cover loss arising from obligor default.
Since 1988, the framework contained in Basel I has been progressively
introduced not only in member countries but also in virtually all other
countries with active international banks. However, the shortcoming
of Basel I have been found gradually, and Committee pay much atten-
tion to the claim of improving the accord. In June 1999, the Commit-
tee issued a proposal , A New Capital Adequacy Framework (CP1),
to replace Basel I. Following extensive communication with banks and
industry groups, the final framework was issued in June 2004 and is
known as Basel II.
Asset correlation is an important component of the Basel II Accord for regulatory capital requirements of credit risk portfolios. It means the correlation of a given firm’s assets with the risk factor that summa- rizes general economic conditions and is a key parameter to determine the shape of the risk weight formulas that is based on the Asymptotic Single Risk Factor (ASRF) model. In the Basel Committee on Bank- ing Supervision (BCBS) document of January 2001 (BCBS, 2001a), asset correlations were assigned a value of 0.2 for all obligors. That is, the asset values of every obligor were assumed to have a factor load- ing of √
0.2 with the common risk factor. However, the fixed value cause much criticism and the range of asset correlation have revised two times, the fist time is 0.1 to 0.2 and now the range for the second time, ignoring the adjustment for the size of company, is between 0.12 and 0.24.
As Basel Committee did not explain the theoretical concept for the
formula of asset correlation, there were many researchers discussing
the appropriateness of this formula from different aspects. This study
applied Bayesian method to ASRF model in order to discuss the prob-
ability distribution of asset return correlation, and the relationship
among probability of default, firm size and asset correlation. Bayesian
inference is the process of fitting a probability model to a set of data
and summarizing the result by a probability distribution on the pa-
rameters of the model. By Bayesian method we can combine the
prior information with data to make plausibility statements or infer-
ences about parameters. However, alternative model can differ in the
specification of the prior distribution, in the specification of the like-
lihood, or both. Sensitivity analysis can be used to assess the effect of alternative prior distribution on the posterior inference. The basic method of sensitivity analysis is to fit several probability models to the same problem. We relaxed the restrictions of the distribution of the factor in one-factor model and investigated three kinds of prior distribution to make sure the sensitivity of the model. The analysis focused on firms in the United States with Standard Industrial Clas- sification (SIC) codes between 2000 and 5999, covering from year2001 to 2005.
Our empirical results confirm that the asset correlation is a decreasing function of probability of default, as suggested by BCBS(2001b,2002), and an increasing function of firm asset size, agreeing with what de- scribed in BCBS(2004). At the same time, our results indicate that there may be some important factors impact the asset correlation were not considered in the formula of asset correlation.
The paper is organized as follows. Chapter 2 summarizes the IRB
approach and also provides several empirical evidences on the asset
correlation. The framework of ASRF model, bayesian method and
Gibbs sampler on ASRF model will be described in chapter 3. Chap-
ter 4 presents the results of the posterior distribution of asset correla-
tion with by Z-score and firm size categories, respectively. Chapter 5
summarizes the results.
2 Literature review
2.1 Overview of Basel II
The objective of Basel II is to modernize the existing capital require- ments framework to make it more comprehensive and risk-sensitive, taking account of many modern financial institutions’ thorough risk management practices. A key aspect of the new framework is its flexibility. It provides institutions with the opportunity to adopt the approaches most appropriate to their situation and to the sophisti- cation of their risk management. In fact, the new capital adequacy scheme is based on three pillars:
1. minimum capital requirements;
2. supervisory review process;
3. market discipline.
With regard to the first pillar, the Committee proposes two approaches.
The first, called standardized approach, adopts external ratings, such
as those provided by rating agencies, export credit agencies, and other
qualified institutions. The second approach opens to internal rating
systems developed by banks, this is so called IRB approach, subject
to the meeting of specific criteria yet to be defined, and to validation
by the relevant national supervisory authority. The internal rating
approach also confers varying degrees of autonomy to banks in the
estimate of the parameters determining risk weightings, and conse-
quently, capital requirements: The foundation approach entails less
autonomy, and the advanced approach a greater one.
2.2 IRB approach
The approach based on internal ratings represents an important change in the Basel Committee’s principles for the determination of regula- tory capital requirements. The amount of capital required to support the economic risks of banking activities is calculated using the internal rating systems of banks. The IRB system requires four basic inputs, both in the foundation and advanced approach. In the IRB approach, institutions are allowed to determine the borrowers’ probabilities of default(PD), while those using the advanced IRB approach will also be permitted to use their own estimates of loss given default(LGD), exposure at default(EAD) and maturity(M). The four factors men- tioned above correspond to the risk parameters on which the Basel II IRB approach is built:
Probabilities of default (PD)
All banks whether using the foundation or the advanced methodology have to provide an internal estimate of the PD associated with the borrowers in each borrower grade. Probability of default per rating grade gives the average percentage of obligors that default in this rat- ing grade under normal business conditions.
Exposure at default (EAD)
Which gives an estimate of the amount outstanding in cases the bor- rower default. In most cases, EAD is equal to the nominal amount of the exposure but for certain exposures - e.g. those with undrawn commitments - it includes an estimate of future lending prior to de- fault.
Loss given default (LGD)
Loss given default rate, which is equal to one minus the recovery rate,
gives the percentage of exposure the bank might lose in case the bor-
rower default. Note that, average loss over long period of time under- state LGD rates during an economic downturn. Therefore, it needs to be adjusted upward to appropriately reflect adverse entomic condi- tions.
Maturity (M)
Where maturity is treated as an explicit risk component, like in the advanced approach, banks are expected to provide supervisors with the effective contractual maturity of their exposures.
2.3 Derivation of the Capital requirement
The capital requirements of banking books are derived from risk weight formulas, which were developed from Asymptotic Risk Factor (ASRF) model. It is believed that the precursor to the formula is the working paper of (Gordy, 2003). In the ASRF models, Z it is assumed the normalized asset return for firm i at time t. This variable can be decomposed in the following way:
Z it = √
ρX t + p
1 − ρ it (1)
where
X t ∼ N (0, 1) it ∼ N (0, 1)
The component it represents the risk specific to institution i, and X t represents a common factor to all firms in the portfolio. √
ρ is the factor loading of the systematic risk and is often interpreted as the sensitivity to systematic. Under the assumption of distribution N ormal(0, 1), the correlation between the normalized asset return of any two borrower is ρ.
If a borrower’s return falls short of some threshold c, i.e.
Z it < c ⇔ Y it = 1 i = 1, ..., N t , t = 1, ...T (2)
Where Y it is an indicator variable and we assume that Y = 1 indicates that the firm defaults.
The unconditional probability of default as shown in (Vasicek, 1997) is as the form:
P D i = P (Y i = 1)
= P (Z it < c)
= P ( √
ρX t + p
1 − ρ it < c)
= Φ(c)
Conditional on a realization x of the common random factor at time t then becomes the conditional probability of default:
P (Y it = 1|X t = x) = P ( it < c − √
√ ρx
1 − ρ |X t = x)
(3)
= Φ( c − √
√ ρx
1 − ρ |X t = x)
P (Y it = 1|X t = x 99.9 ) = Φ( c − √
ρ Φ −1 (0.001)
√ 1 − ρ )
where X denoted the systematic risk factor, x 99.9 is the 99.9 th per- centile of the systematic risk factor, meaning that a worse outcome of the systematic risk factor only has a 0.01 percent change. and the conditional expected loss is:
E[L i |X t = x 99.9 ] = LGD · Φ( c − √
ρ Φ −1 (0.001)
√ 1 − ρ )
(4)
= LGD · Φ( Φ −1 (P D i ) − √
ρ Φ −1 (0.001)
√ 1 − ρ )
As mentioned in BCBS(2004a), the sum of expected loss(EL) and
unexpected loss(UL) are derived from the exposure’s conditional ex-
pected loss(CEL) by the formula
U L + EL = E[L i |X t = x 99.9 ]
(5)
= LGD · Φ( Φ −1 (P D i ) − √
ρ Φ −1 (0.001)
√ 1 − ρ )
Finally, in Basel II the calibration of the risk weight only to unexpected losses. Thus the formula of capital requirement given by BCBS(2004b) is of the form:
K = LGD[Φ( Φ −1 (P D i ) − √
ρ Φ −1 (0.001)
√ 1 − ρ ) − P D] × (6)
1 − b(P D)(M − 2.5) 1 − 1.5b(P D) Where
U L = LGD[Φ( Φ −1 (P D i ) − √
ρ Φ −1 (0.999)
√ 1 − ρ )] − LGD · P D
= (U L + EL) − EL and
1 − b(P D)(M − 2.5) 1 − 1.5b(P D)
is full maturity adjustment as function of PD and M, with b(P D) = (0.11852 − 0.05478 × log(P D)) 2
means the smoothed (regression) maturity adjustment (smoothed over PDs).
The risk-weighted assets and capital requirements are related in a straightforward manner and the resulting formula is the following:
RW A i = K(P D, ρ) · 12.5 · EAD (7)
The factor 12.5 compensate for the solvability coefficient of 0.08, and we can see that the asset correlation finally determine the shape of the risk weight formulas.
2.4 Asset Correlation in the IRB Approach
The asset correlation of Basel risk weight for corporate, sovereign and interbank exposures is given by
ρ i = 0.12 · 1 − e −50P D
i1 − e −50 + 0.24 · (1 − 1 − e −50P D
i1 − e −50 ) (8)
For corporate borrowers, the correlations ρ are first computed by Equation (8) and then modified as follows
ρ − 0.04, S i ≤ EU R
ρ − 0.04 × (1 − (S i − 5)/45), 5EU R < S i ≤ 50EU R ρ, S i > 50EU R :
Where the S i is annual sales for firm i and 0.04 · (1 − S
i45 −5 ) is size adjustment between 5 m EUR and 50 m EUR. Firms in the corpo- rate portfolio with a yearly turnover below 50 m EUR are treated as SMEs. They receive a capital relief dependent on their firm size which is measured by yearly turnover. As a consequence of the combined de- pendency on PD and firm size, the asset correlation for SMEs varies between 0.08 and 0.24.
The Basel Committee has also provided specific mappings between
probability of default p and asset correlation ρ for retail portfolios,
the formulas are as follows:
Residential mortgages
ρ i = 0.15
Qualifying revolving retail exposures
ρ i = 0.04
Other retail exposures
ρ i = 0.03 · 1 − e 35p
i1 − e −35 + 0.16 · (1 − 1 − e −35p
i1 − e −35 )
The asset correlation function implies two systematic dependencies:
1. Asset correlation decrease with increasing PDs 2. Asset correlation increase with firm size.
The two empirical observations are the same as (Lopez, 2004)
2.5 Literature about Asset Correlation
Lopez (2004) In this paper, the ASRF approach was imposed on the
KMV methodology for determining credit risk capital charges in
order to examine the relationship between average asset corre-
lation, firm probability of default and firm asset size measured
by the book value of assets. Using data from year-end 2000,
credit portfolios consisting of US, Japanese, and European firms
are analyzed. Subportfolios based on either firm default proba-
bility or asset size were constructed for the univariate analysis,
and subportfolios based on both variables were used for the bi-
variate analysis. The empirical results suggest that average asset
correlation is a decreasing function of probability of default and
increasing function of firm size. The result suggest that these
factors may need to be accounted for in the regulatory capital requirements for credit risk.
Dietsch and J.Petey (2004) use a one-factor credit risk model to provide new estimates of stationary default probabilities and asset correlations for SMEs in the German and French market and conclude that the relationship between PDs and asset correlations is not negative but positive, especially at the industry level, in the two countries. It is also possible to distinguish different segments inside the SMEs’ population: at least between very small and small SMEs and large SMEs. However, their evidence support for a decrease of asset correlation with firm size. Their results show that the asset correlations in SME population are very week (1-3%) on average and decrease with size.
D¨ ullmann and Scheule (2003) The asset correlation is estimated from time series of ten years with default histories of 53280 Ger- man companies. The sample is divided into categories that are homogenous with respect to default probability (PD) and firm size. In order to estimate the model parameters in the ASRF model we apply three different estimation methods: maximum likelihood estimator (MLE), full-information maximum likelihood estimator (FIML) and method of moments estimator (MME) three kinds of methods to estimate the relationship between asset correlation, probability of default and firm size. Finally, they find that the asset correlation overall increases with firm size but do not observe clear relation between asset correlation and probabil- ity of default.
R¨ osch (2002) examines the effect of the business cycle on correla-
tion in a one factor model. He use database of German corporate
bankruptcies and find that correlations are throughout lower than assumed in the New Basel Accord. Though he does not specifi- cally address the issue of the relationship between probability of default and asset correlation, his findings are still relevant at this point. This is because in some sectors he includes lagged values of the default rates for the respective sector. As described before, the estimates for the parameters of these variables can be inter- preted as explaining parts of the correlation with the common risk factor. When compared to the other included regression variables for a certain sector, the lagged default rates showed smaller stan- dard errors and were all positive. This strongly contradicts the Basel-II assumption of a decreasing asset correlation with PD.
Hamerle and R¨ osch (2003) have made strong use of additional
explanatory variables within a model. It turns out that the in-
clusion of these regression variables reduce the estimates for asset
correlation significantly. To give an example: They use default
data from the U.S market and the rating system of Standard and
Poor’s. For the BB grade the model without regression variables
produces an estimate of 6.04% compared to a 0.67% when the
Federal Reserve Funds Rate is included as a one year lagged vari-
able. This enormous increase in precision can be explained by the
fact, that the business cycle is reflected in macroeconomic quan-
tities. Therefore, the inclusion of these figures will increase the fit
of the model drastically. The given estimates for asset correlation
are non-monotone in the rating categories and do therefore not
support the Basel-II assumptions.
3 Research Method
3.1 Bayesian Statistical Analysis
Bayesian inference is the process of fitting a probability model to a set of data and summarizing the result by a probability distribution on the parameters of the model. It is a coherent framework for com- bining “prior information” with data to make plausibility statements or inferences about parameters. The relationship between prior infor- mation and data set can be seen from the Bayes Rule:
p(θ|y) = p(θ, y)
p(y) = p(y|θ)p(θ)
p(y) = p(y|θ)p(θ) R
Θ p(θ)p(y|θ)dθ (9) Bayes’s Rule is often written as:
p(θ|y) ∝ p(θ)p(y|θ) i.e.
P osterior ∝ P rior × Likelihood
The posterior distribution can be seen as a compromise between the prior and the data. In general, this can be seen on the two well known relationships:
E[θ] = E[E[θ|y]] (10)
V ar(θ) = E[V ar(θ|y)] + V ar(E[θ|y]) (11) The Equation (10) means that the prior mean is the average of all possible posterior means, and the Equation (11) tells us that the pos- terior variance is, on average, smaller than the prior variance.
Take for a concrete example, a case where the prior is θ ∼ Beta(4, 3)
and the observation y=2 is from Bin(5, θ). The relationship between
prior, likelihood and posterior can be seen as Figure 1.It seems like that the posterior distribution may have a mean somewhere between the locations of the peaks of f p(y|θ) and p(θ), and this graphic indi- cates how prior and likelihood combine to produce the posterior.
Figure 1: Prior vs. Likelihood vs. Posterior
3.2 Gibbs Sampler
From Equation (9), we can see that to derive the posterior distribution
needs through the technique of integration. However, if the parame-
ter which we are interested in is a high dimension one, it must be a
stiff matter, that is why we use Gibbs sampler. The Gibbs sampler
is a technique for generating random variables from a distribution in-
directly, without having to calculate the density. Through the use
of techniques like the Gibbs sampler, we are able to avoid difficult
calculations. For Θ = (θ 1 , θ 2 , . . . , θ k ), the fundamental idea of Gibbs
sampler is, from some starting vector Θ (0) one creates Θ (i+1) from Θ (i)
by “updating” in succession the k coordinates of Θ (i) drawing at ran-
dom from the conditional distribution of each variable given current
values of all the others.
In order to illustrate the basic idea, consider the simple example of Gibbs sampling from the joint discrete distribution of Θ = (θ 1 , θ 2 ) given in Table 1 below:
Table 1: pmf of Θ = (θ 1 , θ 2 )
θ 2 = 1 θ 2 = 2 θ 1 = 1 0.2 0.1 θ 1 = 2 0.2 0.2 θ 1 = 3 0.1 0.2
This joint distribution has conditional distributions:
θ 1 g(θ 1 |θ 2 = 1)
1 2/5
2 2/5
3 1/5
θ 1 g(θ 1 |θ 2 = 2)
1 1/5
2 2/5
3 2/5
and
θ 2 g(θ 2 |θ 1 = 1)
1 2/3
2 1/3
θ 2 g(θ 2 |θ 1 = 2)
1 1/2
2 1/2
θ 2 g(θ 2 |θ 1 = 3)
1 1/3
2 2/3
Suppose that one beings Gibbs sampling with initial Θ 0 = (2, 1)
Then generates a replacement for θ 1 = 2 by generating a value from g(θ 1 |θ 2 = 1), and suppose that θ 1 = 3 is generates. Then one must generate a replacement for θ 2 = 1 using the distribution specified by g(θ 2 |θ 1 = 3). If θ 2 is generated, after one complete Gibbs cycle we obtained:
Θ 1 = (3, 2)
After generating a long string of Θ in this way, the theory of Markov Chains implies that the relative frequency distribution of the gener- ations will approximate the distribution in Table 1. (Besag, 1974) establish that this set is sufficient to determine the joint or any mar- gins distribution.
3.3 The Asymptotic Single Risk Factor Model
In the ASRF models, Z it is assumed the normalized asset return for firm i at time t. This variable can be decomposed in the following way:
Z it = √
ρX t + p
1 − ρ it (12)
where
X t ∼ N (µ x = 0, σ x 2 = 1) it iid ∼ N (µ = 0, σ 2 = 1)
and X t is independent of it , which means that Z it is considered to have a standardized Gaussian distribution.(i.e. µ z = 0, σ z 2 = 1) The component it represents the risk specific to institution i, and X t rep- resents a common factor to all firms in the portfolio. √
ρ is the factor loading of the systematic risk and often interpreted as the sensitivity to systematic, and ρ is the correlation between the normalized asset return of any two borrower.
Under ASRF model, the correlation between two borrowers i and j comes:
Corr(Z i , Z j ) = Cov(Z i , Z j ) pV ar(Z i )pV ar(Z j )
(13)
= E(Z i Z j ) − E(Z i )E(Z j )
σ zi σ zj
where
E(Z
iZ
j) = E[( √
ρX + p
1 − ρ
i)( √
ρX + p
1 − ρ
j)]
= E[ρX
2+ √ ρ p
1 − ρX
i+
√ rho p
1 − ρX
j+ (1 − ρ)
ij]
= ρE(x
2) + √ ρ p
1 − ρE(X)E(
i) + √ ρ p
1 − ρE(X)E(
j) + (1 − ρ)E(
ij)
= ρσ
2x+ ρµ
2x+ 2 √ ρ p
1 − ρµ
xµ
+ (1 − ρ)µ
2x(14)
and
E(Z i )E(Z j ) = [ √
ρµ x + p
1 − ρµ ] 2
= ρµ 2 x + 2 √ ρ p
1 − ρµ x µ + (1 − ρ)µ 2 (15)
From Equation (14) and (15) we can obtain the covariance of Z i and Z j
Cov(Z i , Z j ) = E(Z i Z j ) − E(Z i )E(Z j )
= ρσ x 2 + ρµ 2 x + 2 √ ρ p
1 − ρµ x µ + (1 − ρ)µ 2 x − (ρµ 2 x + 2 √
ρ p
1 − ρµ x µ + (1 − ρ)µ 2 )
= ρσ x 2
So the correlation between Z i and Z j becomes:
Corr(Z i , Z j ) = ρσ x 2
σ zi σ zj (16)
Under the assumption of ASRF model σ x 2
、σ zi
、σ zj all equal to one, so the correlation between any two borrowers is ρ, but if the distribution of systematic risk factor, X, and idiosyncratic risk factor, , are not standard normal distribution the asset correlation will not equal to ρ.
3.4 ASRF Model with Bayesian Method
In this paper, we relax the restrictions on the distribution of system-
atic risk factor and idiosyncratic risk factor and let them follow normal
distribution with parameters (µ x , σ x 2 ) and (µ , σ 2 ). Not alike the prop-
erties in frequency method that the parameter is a fixed variable, in
bayesian method all parameters could have their own distribution, and we assume µ x and µ follow normal distribution with parameters (0,1), at the same time, σ x 2 and σ 2 follow inverse Gamma(1, 1).
Because of the alteration in the distribution, the asset correlation be- tween two borrowers in Equation (16) is not ρ any more; however, we should adjust the value to solve the true asset correlation, and our ASRF model using Bayesian method becomes:
Z it | µ Z σ Z 2 ∼ N ormal(µ Z , σ Z 2 ) X t | µ x σ x 2 ∼ N ormal(µ x , σ x 2 )
it | µ σ 2 ∼ N ormal(µ , σ 2 ) where
µ Z = √
ρµ X + p
1 − ρµ σ Z 2 = ρσ x 2 + (1 − ρ)σ 2
µ x ∼ N ormal(0, 1) µ ∼ N ormal(0, 1)
σ 2 x ∼ inv − Gamma(1, 1) σ 2 ∼ inv − Gamma(1, 1)
ρ ∼ U nif orm(0, 1) The actual correlation becomes ρ T RU E , not ρ
ρ T RU E = Cov(Z j , Z k ) pV ar(Z j )pV ar(Z k )
= ρ σ 2 x σ Z 2
For the input data, we use the firm’s equity daily return as a proxy
of asset return, as mentioned in (Henneke, 2005), it has been industry
practice to use equity correlations as a proxy for asset correlations.
For systematic risk factor, we applied stock index returns to estimate, and for idiosyncratic risk factor, , the theory of CAPM is adopted.
We use no-intercept regression analysis and remain residuals as the idiosyncratic risk factor.
The cause of choosing prior distribution for µ x and µ as N ormal is due to the properties of data, and inverse Gamma distribution for variance of systematic risk factor and idiosyncratic risk factor is be- cause of the positive nature of variance. Additionally, variance of all distribution follows Chi square distribution, which is a special case of Gamma distribution, and we take inverse Gamma but not Gamma is just a usual practice of Bayesian method. For the square of loading of systematic risk factor, ρ, we take three kinds of prior: U nif orm(0, 1), N ormal(0, 1)I(0, 1) 1 and Beta(0.5, 0.5) to see if they have similar ten- dency. If they do have similar tendency, we can say that our likelihood belief is strong enough, so no matter under what kinds of prior, we can obtain same conclusion.
3.5 Posterior distribution of asset correlation
— Application of Gibbs Sampler
In this paper, we use Gibbs sampler to simulate a sequence to ap- proximate the posterior distribution. Through the use of techniques like Gibbs sampler, we are able to avoid difficult calculations. In the following content, we illustrate how to use Gibbs sampler to obtain the posterior distribution of asset correlation in our Bayesian ASRF model:
1