Journal of Applied Operational Research (2013) 5(2), 42–47
© Tadbir Operational Research Group Ltd. All rights reserved. www.tadbir.ca
ISSN 1735-8523 (Print), ISSN 1927-0089 (Online)
Transition and reversion of Japanese
corporate rating structure under the recent
credit crises
Motohiro Hagiwara 1.*, Yasuhiro Matsushita 2 and Katsuaki Tanaka 3
1
School of Commerce, Meiji University, Tokyo, Japan
SET Software, Nagoya, Japan
3
Faculty of Business Administration, Setsunan University, Osaka, Japan
2
Abstract. This study attempts to know the details of transition of rating structure of U.S. leading rating agencies (S&P,
Moody’s) and Japanese agencies (JCR and R&I) under recent credit crises starting from BNP Paribas shock in August 2007
using Artificial Neural Network (ANN). This study checks the transition of recent bond rating structure under the credit crises
based on accounting information giving Altman Z-score. Japanese corporate bond ratings are transformed to normally
distributed variables using published 5 -Year actual default probability, and are modeled as functions by key ratios giving
Z-score. As remarkable findings, rating structures of all agencies before crises (2005-2007) lost explanatory power under
the credit crises (2008-). Only Moody’s seem to have recovered explanatory power by 2010 and those of S&P, JCR and
R&I have not. But exposures of five factors of all agencies have not experienced significant change but slight change in exposures
of future and current profitability or liquidity under crises. We would like to point out that the risk accompanying selecting
the rating agencies might be enlarged and should be much concerned.
Keywords: credit risk; bond pricing; risk management; artificial neural network
* Received September 2012. Accepted May 2013
Introduction
The majority of previous research attempting to explain bond ratings systems are based on financial and other
quantitative data. Among prior research using qualitative analyses, the most commonly assumed was a model
where financial data of issuing corporations are set as explanatory variables. In these particular research horizons,
pioneering theses were presented by Kaplan and Urwitz (1979).
In addition, in previous researches including Altman(1968) and West (1973), a number of models were presented,
which were designed to estimate rating transition probability by applying an ordered probit model or a logit
model, to more complex ones designed to simultaneously estimate default probability and rating transition probability.
Carty and Jerome (1994) and Cyert and Thompson(1968) were among those who attempted to analyze credit
risks of corporations and banks. Analyses using actual data held by banks include those by Barkman (1981) and
Betancourt (1999). Furthermore, analyses using panel data include those by Blume et al (1998).
* Correspondence: Motohiro Hagiwara, School of Commerce, Meiji
University, 1-1 Kanda-Surugadai, Chiyoda-ku, Tokyo 101-8301.
E-mail: [email protected]
M Hagiwara et al
43
In comparison with studies on rating structure or predictions, the group of studies analyzing the differences in
rating structures of each agency or time (or cross-sectional) consistency of informational content of rating information
was a relatively small stream. Time or cross-sectional informational consistency is an important characteristic that
an ideal rating information should fulfill. Consistency means that, if factors that determines the rating information are
the same, each rating agency publish the same rating in different time and sectors. After the BNP Paribas-shock
in 2007, the neutrality of rating agency is doubted and serious criticism on rating information and rating agencies
occurred concerning time or cross-sectional consistency. Most recently, downgrades of sovereign ratings have
crucial effects on global economy. Therefore, continuous check of time or cross-sectional consistency, or if under
the inconsistency, continuous tracing of transition of rating structure of each agency by quantitative methods is
appealing for investors, issuers and financial authorities.
As related studies on topics above, Altman and Rijken (2005) surveys on the use of agency credit ratings reveal that
some investors believe that credit-rating agencies are relatively slow in adjusting their ratings. A well-accepted
explanation for this perception on rating timeliness is the through-the-cycle methodology that agencies use.
Through-the-cycle ratings are intended to measure default risk over long investment horizons and to respond only
to changes in the permanent component of credit quality. A second aspect of the through-the-cycle methodology
is the prudent migration policy. In a benchmark study with a financial ratio-based credit-scoring models--an
agency-rating prediction model and default-prediction models with various time horizons--they confirm the exclusive focus of agencies on the permanent component of credit quality and they model and quantify the agencies'
prudent migration policy. Batchelor and Manzoni (2006) examine the effect of rating revisions on sterling Eurobond yields using a panel model with conditional heteroskedasticity that controls for event-induced changes in
the variance of spreads. Positive rating revisions are fully anticipated by the time the upgrade occurs. Negative
revisions are only partially anticipated, and spreads on downgraded bonds rise for some time after the downgrade
has been announced. This asymmetry is not apparent in a conventional event study model. All ratings announcements
are accompanied by a temporary fall in yield volatility. They attribute this to the resolution of uncertainty about
the true rating of the bond. Mahlmann (2008) analyzes the economic role of such rating publication rights. Using
a theoretical model, it was shown that an equilibrium with partial nondisclosure of low-quality ratings can
emerge whenever investors cannot be sure whether rating nondisclosure is due to the firm being not rated, or due
to the rating's adverse content. Moreover, since from an investors' perspective, strategically acting rated firms and
unrated firms are pooled, unrated firms' debt is always under-valued (compared to a situation in which investors
know that the firm is not rated), and the debt of firms concealing their rating is always over-valued. Niemann et
al (2008) introduces a new approach to improve the performance of rating prediction models for multinational
corporations. In this segment, the low number of defaults poses a challenge, as it prevents rating models to be
constructed for individual industry sectors or regions. They show that reducing group-level heterogeneity in financial
ratios results in a rating prediction model with better performance than both unadjusted models and models adjusted
by including industry dummies or other simpler procedures. Their approach fills a gap in cases where a limited
dataset does not permit the construction of separate models for individual industries or regions. Cheng and
Neamtiu (2009) find that the rating agencies not only improve rating timeliness, but also increase rating accuracy
and reduce rating volatility. Their findings support the criticism that, in the past, rating agencies did not avail
themselves of the best rating methodologies/efforts possible. When their market power is threatened by the possibility
of increased regulatory intervention and/or reputation concerns, rating agencies respond by improving their credit
analysis. This study belongs to this new and small stream, and focuses on the following points.
1. Are there any changes in rating structures for the period of credit crises starting from BNP Paribas shock in
August 2007?
2. What kind of changes occurred on rating structures of each agency under the crises?
The most important point of analysis is to check the time consistency of rating before and after the Lehman
Shock and verify the validity of appraisal and doubts on the method of determining the rating that occurred after
the Lehman Shock.
This study defines the rating structures by five parameters, inspired by Altman’s Z-Score Model (Altman
1968). Procedure of estimating rating structures are consists of three steps. First, rating information are transformed
into published actual default probability in five years. Next, Default probability is transformed into Z-values by
probit function. Then, rating structures are estimated as non-linear functions by Altman’s five parameters using
Artificial Neural Network (ANN).
Journal of Applied Operational Research Vol. 5, No. 2
44
We used an ANN modeling software, NeuralWorks Predict (NWP) by NeuralWare Inc. (http://www.
neuralware.com/). NWP implements Cascade-Correlation learning algorithm by Fahlman and Lebiere (1991),
which is known at least 10 times faster than standard Back-propagation and is an algorithm that the network
determines its own size and topologies. NWP undertakes some non-linear transformations for each input variables
and variable selection by Genetic Algorithms in advance of training process.
The remainder of the paper is organized as follows. In the following sections, the data used are explained and
the results of the analysis are revealed, followed by a discussion. Finally, findings are summarized in the last section.
Analysis
Data set
Data set includes long-term bond ratings given to companies listed on the Tokyo Stock Exchange from U.S. rating
agencies (S&P,Moody’s) and Japanese agencies (JCR,R&I) simultaneously at the end of March from 2005 to
2010. For companies with no long-term bond rating, we used the issuer credit rating. R&I does not publish longterm bond ratings and only publishes issuer credit ratings. JCR does not publish issuer credit ratings and only
publishes long-term bond ratings. S&P publishes only issuer credit ratings. Moody’s publishes both issuer credit
ratings and long-term bond ratings. However, in practice, for companies that receive both ratings, there are no
differences in their ratings. As a result, we collected 175 samples. After 2011, S&P and Moody’s have withdrawn
many ratings on Japanese companies and sample size is too small.
Table. 1. Number of samples.
2005
32
2006
34
2007
30
2008
31
2009
31
2010
17
Actual 5 year cumulative default probability 5YDP is collected from each agency’s site. Default probability
corresponding to each rating and Z-value corresponding to each default probability are shown below
~
~~
( P Z Z  Z  value  5YDP Z ～ N 0,1 Z-value corresponding to 5YDP = 0%, or 100% is assumed to be -6. or
6, respectively).


Table. 2. Actual 5 year default probability and Z-value.
S&P
AAA
AA
A
BBB
BB
B
CCC or lower
5Y DP
0.10%
0.30%
0.60%
3.00%
11.30%
25.40%
50.90%
Z-value
-3.09023
-2.74778
-2.51214
-1.88079
-1.21073
-0.66196
0.022562
Moody’s
5Y DP
Z-value
0.00%
-6.0000
0.00%
-6.0000
0.29%
-2.75888
0.69%
-2.46243
3.91%
-1.76123
8.64%
-1.36326
100.00%
6.0000
JCR
5Y DP
0.00%
0.00%
0.23%
1.01%
9.65%
36.59%
100.00%
R&I
Z-value
-6.0000
-6.0000
-2.83379
-2.32261
-1.30175
-0.34273
6.0000
5YDP
0.00%
0.04%
0.65%
1.12%
7.60%
Z-value
-6.0000
-3.35279
-2.48377
-2.28352
-1.4325
22.73%
-0.74777
Z-values above are output variables of ANN in this study and input variables are 5 financial ratio described below.
Financial data corresponding to each rating information is those of the nearest fiscal term collected from Nikkei
Needs. The choice of 5 input variables is inspired by the Z-Score model (Altman (1968)). Working Capital/Total
Assets (WC/TA) is a proxy for the fort term liquidity of the firm. Retained Earnings/Total Assets (RE/TA),
EBITDA/Total Assets (EBTD/TA), and Market Value of Equity/Total Liabilities (MV/TL) variables proxy for
historic, current and future profitability, respectively. Net Sales/Total Assets (NS/TA) is a turnover ratio of Total
Assets as a proxy for firm’s efficiency at using its total assets. The reason why Z-Score model is selected is that
Z-Score model is very popular among practitioners and gives clear financial implication. As future tasks, analyses
in this study can be expanded by newer models.
M Hagiwara et al
45
Robustness of rating structure before and under the credit crises
Explanatory power of ANN models on rating structure is measured by correlation between actual (Table 2) and
estimated Z-values, AIC and BIC. ANN models on actual Z-values in training period are estimated using data
subset randomly selected from training period, and Z-values in test period (i.e. credit crises, 2008~2010) are predicted
by the estimated models in training period. By repeating these procedures, transition of rating structures can be
recognized by decrease of average explanatory power during test periods. Averaged correlation and t-statistics on
decrease of correlation from training period are shown in Table.3 below.
Table. 3. Transition of rating structure under the credit crises 2007~2010.
S&P
Moody’s
JCR
R&I
correlation
AIC
BIC
correlation
AIC
BIC
correlation
AIC
BIC
correlation
AIC
BIC
Training period 2005~2007
2005~2007
2008
2009
0.9844
0.7452
0.6515
-141.33
50.57
66.35
-133.18
51.34
67.12
0.9086
0.8066
0.4944
164.38
94.89
125.46
172.54
116.81
117.77
0.9824
0.2986
0.5389
138.59
160.50
143.33
146.74
161.27
144.10
0.9792
0.6276
0.4951
-154.78
44.57
52.07
-146.63
45.34
52.84
2010
0.8900
26.22
20.93
0.6361
78.80
75.62
0.2690
99.03
93.74
0.7968
30.37
30.37
Training period
2005~2008
0.9838
-196.73
-187.01
0.9754
75.22
84.95
0.9889
95.57
105.30
0.9782
-202.34
-192.62
2005~2008
2009
2010
0.6678
0.8238
65.39
33.51
66.16
28.22
0.5792
0.9291
117.57
56.25
118.34
50.96
0.6126
0.7473
141.77
78.65
142.54
73.36
0.5386
0.5976
62.19
41.06
62.96
35.77
Table.3 gives several implications below.
 As for Moody’s, explanatory power decreased significantly at 2008 and 2009, and improved at 2010. This
fact can be interpreted that rating structure experienced significant change in 2008 and 2009 and reversion
to the structure of training period in 2010. Reversion at 2010 is remarkable compared with Japanese agencies.
Moody’s is the most stable based on the transition of t-statistics.
 As for the other three agencies, explanatory power measured by correlation continued to decrease at 2008
and 2009, but recovered insignificantly at 2010. This fact can be interpreted that rating structure experienced
serious transition in 2009, but cannot revert to the structure of training period in 2010.
 Based on AIC and BIC, more plausible measure of nonlinear model like ANN, we get similar interpretation.
Sensitivity analysis
In order to analyze the more details on transition of rating structure, sensitivity analyses are carried out based on
ANN models estimated by all data in each year. Table.4 shows average exposures of each factor estimated by all
data in each year. ANN model is non-linear and so exposures have variance. And t-stats is related to the change
of average exposure at each test period (2008, 2009, 2010) from that in training period (2005～2007). Significant
change of average exposure implies time inconsistency of rating structure and changes of weight of each factor.
We recognize the points to be aware of is over-fitting problem accompanying ANN. To solve this problem, the
same analyses should be carried out on many models which were structured on randomly selected data subsets as
future tasks.
Journal of Applied Operational Research Vol. 5, No. 2
46
Table. 4. Sensitivity analysis (training period = 2005~2007; test period = 2008, 2009 and 2010).
2005
2006
2007
2005-07
WC/TA
0.0855
-0.0515
-0.3855
-0.0704
RE/TA
-0.4416
-0.0045
-0.1766
-0.1479
S&P
EBTD/TA
-0.0615
-0.2392
-0.0605
-0.4624
MV/TL
-0.3454
0.0576
-0.1133
-1.1939
NS/TA
-0.0659
-0.1496
-0.2358
-0.2835
WC/TA
-0.5786
0.6112
-0.2745
-0.2321
RE/TA
-0.1055
0.2565
-0.0331
-0.2596
Moody's
EBTD/TA
-0.8248
-0.1092
-0.4999
-0.4670
MV/TL
0.0948
0.2159
-0.1322
1.3233
NS/TA
-0.2009
0.6959
0.6267
0.1515
2008
(t-stats)
-0.3680
(-0.06)
-0.2062
(-0.15)
-0.3752
(-0.11)
0.2482
(0.29)
-0.0807
(0.10)
0.0275
(0.24)
-0.2220
(-0.05)
-0.4979
(-0.36)
-0.0632
(-0.12)
0.2671
(0.20)
2009
(t-stats)
0.3472
(0.26)
-0.2524
(-0.21)
-0.1403
(0.12)
-0.2741
(-0.17)
0.0890
-0.20)
-0.0985
(0.00)
-0.3547
(-0.29)
-0.2656
(-0.09)
-0.1370
(-0.22)
1.0266
(0.78)
2010
(t-stats)
-0.0680
(-0.01)
-0.2178
(-0.18)
0.0734
(0.16)
-0.0119
(0.20)
0.1245
(0.27)
-0.5461
(-0.41)
0.3652
(0.25)
WC/TA
-0.3665
-0.0925
-1.1233
-0.1076
-0.9031
(-0.58)
RE/TA
-0.2477
-0.4186
-0.1582
-0.4261
0.0469
(0.12)
MV/TL
-0.2476
-0.2677
-0.8288
0.2465
-0.1685
(-0.17)
NS/TA
-0.2519
0.1769
0.5515
0.0771
0.3039
(0.22)
WC/TA
0.0351
-0.2796
-0.0351
-0.2339
-0.7199
(-0.40)
RE/TA
-0.6838
-0.0629
0.1372
-0.4035
-0.1762
(-0.07)
-0.8394
(-0.42)
R&I
EBTD/TA
-0.5549
-0.6270
-0.3961
-0.4083
-0.2788
(-0.05)
0.5862
(-0.04)
2005
2006
2007
2005-07
2008
(t-stats)
-0.2659
(0.02)
JCR
EBTD/TA
-0.5465
-0.2031
-0.4796
-0.0887
-0.2342
(-0.11)
MV/TL
0.1738
0.5071
-0.2569
0.5743
-0.3915
(-0.36)
NS/TA
-0.2363
0.3314
0.4451
0.7663
-0.0272
(-0.22)
2009
(t-stats)
-0.7149
(-0.43)
-0.5166
(-0.23)
0.0044
(0.03)
-0.3235
(-0.32)
-0.0666
(-0.08)
-0.2757
(-0.07)
0.0017
(0.08)
-0.0239
(0.11)
-0.8594
(-0.70)
-0.0270
(-0.22)
2010
(t-stats)
-0.8916
(-0.45)
-1.2259
(-0.63)
0.8318
(0.57)
-1.1481
(-0.25)
0.7755
(0.44)
-0.7219
(-0.27)
-0.8740
(-0.26)
0.4542
(0.21)
-0.9861
(-0.26)
0.2189
(-0.28)
Table.4 gives several implications below.
 Despite the significant decrease of explanatory power of the rating structure in test periods, change of
exposures are insignificant concerning all factors. All agencies experienced insignificant change in exposure in
2009. These facts imply that there possibly be factors which are not included by ANN models estimated in
this study but have significant effects on rating structures. In order to confirm this point, the same analyses
should be repeated using many kinds of factors included in newer credit scoring model. Or by repeating
analyses on data subset ,for example low and high rating sample, significant facts can be found. Although
change of exposures are insignificant, several implications below can be derived.
 As for S&P, exposures of all factors revert to that in training periods in 2010. This is the facts inconsistent
with the facts implied in table 3. It is implied that factors which are not included in rating structure causes
the decrease of explanatory power in 2010. Sign of exposure of MV/TL change from negative in training
period into positive in 2010. This implies S&P attached less importance to future profitability after credit crises.
 As for Moody’s, despite the remarkable reversion of explanatory power in 2010 as shown in table 3, exposure
of WC/TA, RE/TA and EBIT/TA continued to change in 2010. Sign of exposure of WC/TA change from
negative in training period into positive in 2010. This implies Moody’s attached less importance to liquidity
of the firm after credit crises.
 As for JCR and R&I, exposure of many factors get more and more different in 2010 from that in training
period. This is the facts consistent with the facts implied in table 3. An exposure of EBTD/TA increased
and the exposures of other factors decreased. Sign of exposure of EBTD/TA change from negative in training
period into positive in 2010 and sign of exposure of MV/TL change from positive in training period into
negative in 2010. This implies these agencies attached greater importance to future profitability than current
profitability after credit crises.
M Hagiwara et al
47
Conclusions
This study attempts to know the details of transition of rating structure under recent credit crises starting from
BNP Paribas shock in August 2007 using ANN. As results, rating structures of all agencies before crises (20052007) lost explanatory power under the credit crises (2008). Only Moody’s seem to have recovered explanatory
power by 2010 and those of S&P, JCR and R&I have not. But exposures of five factors of all agencies have not
experienced significant change but slight change in exposures of future and current profitability or liquidity under
crises. It indicates that the risk of selecting the rating agencies might be enlarged and should be much concerned.
Insignificant change of exposures imply that there possibly be significant factors which are not assumed in this
study. As another implication, variance of exposures are large because of difference in exposures between high
and low rating group, or industrial effects. As future tasks, this study can be developed by analyzing based on
factors of new model, for example, Z-Score+ model just published in 2012 as newer model of Z-Score model or
on data subset selected by rating grade or industries.
The frameworks of study can be widely used in the cases where many agencies (evaluators) publish the rating
information in categorical data form. It can be a versatile method that allows us to make a quantitative comparison
and discuss validity of rating information with the aim of bringing about efficient investment or financial decisionmaking, and asset pricing.
Acknowledgments— This work is supported by Grant-in-Aid for Scientific Research(C)(23530449).
References
Altman E (1968). Financial Ratios, Discriminant Analysis
and the Prediction of Corporate Bankruptcy. Journal of
Finance 23: 189–209.
Altman E and Rijken HA (2005). The Impact of the Rating
Agencies' Through-the-Cycle Methodology on Rating
Dynamics. Economic Notes 34:127-54.
Barkman AI (1981). Testing the Markov Chain Approach
on Accounts Receivable. Management Accounting 62:
48–50.
Batchelor R and Manzoni K (2006). The Dynamics of
Bond Yield Spreads around Rating Revision Dates.
Journal of Financial Research 29:405-420.
Betancourt L (1999). Using Markov Chain to Estimate
Losses from a Portfolio of Mortgage. Review of Quantitative Finance and Accounting 12: 303–318.
Blume ME, Felix L and McKinly AC (1998). The Declining
Credit Quality of U.S. Corporate Debt: Myth or Reality?.
Journal of Finance 53: 1389–1413.
Carty LV and Jerome SF (1994). Measuring Changes in
Credit Quality. Journal of Fixed Income: 27–41.
Cheng M and Neamtiu M (2009). An Empirical Analysis of
Changes in Credit Rating Properties: Timeliness, Accuracy and Volatility. Journal of Accounting and Economics 47:108-130.
Cyert RM and Thompson GL (1968). Selecting a Portfolio
of Credit Risks by Markov Chains. Journal of Business
41: 39–46.
Fahlman SE and Lebiere C (1991). The Cascade-Correlation
Learning Architecture. August 29, CMU-CS-90-100,
School of Computer Science, CMU, Pittsburgh, PA.
Kaplan RS and Urwitz G (1979). Statistical Models of
Bond Ratings: Methodological Inquiry. Journal of
Business 52: 231–261.
Mahlmann T(2008). Rating Agencies and the Role of Rating Publication Rights. Journal of Banking and Finance
32:2412-22.
Niemann M, Schmidt JH and Neukirchen M(2008). Improving Performance of Corporate Rating Prediction
Models by Reducing Financial Ratio Heterogeneity.
Journal of Banking and Finance 32:434-446.
Rumelhart DE, Hinton GE and Williams RJ (1986). Learning
internal representations by error propagation. In D.E.
Rumelhart and J.L. McClelland, editors, Parallel Distributed Processing Vol.1, MIT Press, Cambridge, MA.
West R (1973). Bond Rating, bond yields and financial
regulation: some findings. Journal of Law and Economics 16: 159–168.