PD Estimation: Pooled Method & Transition

Matrix Approach

Dr. Arindam Bandyopadhyay

Professor, Finance Area

National Institute of Bank Management

Why PD computation?

 By providing a PD for an obligor, one is providing a forecast of the

likelihood of default over the specified horizon (e.g., 1 year)

 This is true even if the historical default experience is used

 Saying: next year will look like the average of the past

 This purely historical view is a useful starting point and benchmark

for more formal models
Approaches for Default Prediction

 Expert judgement based Rating Models (Judgmental)

 Credit Scoring Systems for Corporates, SMEs and Retail Exposures
(Statistical: MDA, Logit, Probit, Tobit type)
 Artificial Neural Network Approach (ANN)
 Structural Models/Option Theoretic Models/ Market based Models
(MKMV type)
 Hybrid Credit Scoring Model (Like Moody’s)
Scoring Systems to Estimate PD

 Qualitative (subjective)
 Quantitative univariate (accounting/market measures)
 Quantitative multivariate (accounting/market measures)
 Discriminant analysis, logit & probit models, non-linear models,
such as neural networks
 Discriminant and logit models in use:
 Consumer models - Fair Isaacs (FICO)
 Manufacturing - Z-score (5 variables) based on MDA
 Industrials - ZETA score (7 variables)
 Emerging market industrial - EM score (4 variables)
 Logistic, Probit Models
 Structural Models/Option Theoretic Models/ Market based

Models (MKV type)

 Moody’s Private firm models (Risk Calc, Z” score-hybrid model)
 Other - bank specialized systems (Expert System)
Vendor-Provided PDs

 Based on historical defaults:

 S&P CreditPro
 Moody’s Credit Research Database
 Fitch Risk Management
 Based on balance sheet data:
 Altman’s Z-score (www.creditguru.com)
 S&P CreditModel
 Moody’s KMV RiskCalc
 Based on equity data / market data:
 Moody’s KMV EDFs from CreditMonitor
 Kamakura KRM-cr
 Hybrid models:
 Moody’s KMV Private Firm Model
 Fitch CRS Model Suite
 Kamakura Private Firm Model
 CreditSights BondScore
 Tasks for Computing PDs

Rate the obligor using one of the following:

- Expert judgment with given criteria
- Score from a quantitative model
- PD from a quantitative model directly estimating PD

Bucket the obligor into one of seven or more grades

Estimate PD for each bucket using one of the

following:
- External historical loss data such as rating agencies
- Internal historical loss data for obligors in each bucket
- Average modeled PDs for obligors in each bucket
IFRS 9 PD

 Two types of PDs are used for calculating ECLs:

 12 –month PDs –
 This is the estimated probability of default occurring within the next 12
months (or over the remaining life of the financial instrument if that is
less than 12 months).
 Lifetime PDs–
 This is the estimated probability of a default occurring over the remaining
life of the financial instrument.
 This is used to calculate lifetime ECLs for ‘stage 2’.
 PDs may be broken down further into marginal default probabilities (MPDs)
for sub -periods within the remaining life.
 To determine lifetime PDs, bank can builds from the 12-month PD model. For
this, it develops lifetime PD curves or term structures to reflect expected
movements in default risk over the lifetime of the exposure.
Expected Credit Loss (ECL) Model

Recognition of expected credit losses

12 month expected Lifetime expected Lifetime expected

credit losses credit losses credit losses

Interest revenue
Effective interest on
Effective interest on gross Effective interest on gross amortised cost carrying
carrying amount carrying amount amount
(i.e. net of credit allowance)

Stage 1 Stage 2 Stage 3

Underperforming Non-performing
Performing (Assets with significant (Objective evidence of
increase in credit risk
(Initial recognition) credit impaired
since initial
recognition) assets)
 IFRS 9 vs. Basel IRB PD
 IFRS 9 clearly emphasizes all the information used should be forward
looking must incorporate current economic conditions. It also states
historical information should be used, but adjusted to reflect current
conditions.
 Hence it is clear that PD must reflect current economic conditions and
need not be TTC or adjusted for a downturn, as in the case of Basel IRB
rules.
 According to Basel IRB regulations, PDs used for calculating regulatory
capital should be stable throughout the economic cycle, whereas ECL
model requires PIT PDs which are highly sensitive to changes in economic
conditions.
 PDs used under Basel IRB framework are through-the-cycle (TTC) PDs,
whereas under the IFRS 9 framework, we need point-in-time (PIT) PDs.
 By accounting for the current state of the credit cycle, PIT measures track
closely the variations in default and loss rates over time.
 For calculation of capital buffer against unexpected losses, the through-
the-cycle PD (unconditional of the states of economic cycle, PD) should be
used in the RWA formulas.
 Basel vs. IFRS 9

• The 12 month PD used for regulatory purposes (Basel) should be

adjusted for IFRS 9 usage.
• PD used for IFRS 9 should be point in time (PIT) probabilities (i.e., prob.
of default in current economic condition).
• Under Basel IRB approach, PD is calculated through the cycle and
estimates are less sensitive to changes in economic conditions.
• Therefore, Basel PD reflect long term trends in PD behavior as opposed
to IFRS 9 PIT PD.
• As a result, during benign credit environment, IFRS 9 PIT PD will be
lower than the regulatory PD (TTC), while the adjustment will be
opposite during a financial crisis.
Computing PDs

 There are broadly two data sources used to compute obligor PD:
 Internal data and model/ internal ratings
 External reference data/ external ratings

 Banks will likely use a bit of both:

 For middle market & small business, no public ratings, so
internal ratings only
 For large corporate & large middle market, public ratings
available for benchmarking / calibration of internal ratings.
Transition Matrix Approach of Measuring PD

 The credit migration or transition matrix describes the likelihood of

migrating to any other state, including default, for a given rating.
 Based on Markov Chain process (discrete time stochastic process).
 Can be constructed through Mortality analysis/Cohort analysis.
 Cohort is a most popular method for estimating transition matrix.
 Transition Probability

• Default is not an abrupt process

• A firm’s credit worthiness and asset quality declines gradually
=>credit migration
• Transition Probability is the chance of credit quality of a firm
improving or worsening
• When credit quality worsens, the probability of future default
increases
• Default is an extreme event of credit migration
Uses of Transition Matrix

 Provide cardinal assessment of risk

 Used for estimating credit risk premium and pricing of credit
derivatives
 Valuation of a bond or loan portfolio which might be used by a
portfolio or risk manager (Credit MetricsTM approach of JP morgan)
 Tool for loan portfolio monitoring
 Used for estimation of Historical PD
 Model Benchmarking & Calibration
 Stress testing capital
Transition Matrix

Grade at T+1
AAA AA ... B CCC D
AAA

AA

Grade at T .
.
.
.

CCC
Key Issues in Calculating Historical PDs

 Sample set
 Time period: length of time
 Time period: capture of economic cycles
 Clear differences across business cycle
 Ratings consistency over time
 Academic research has shown that rating definitions do seem
to vary over time
 Application Score Card Development

Example Score Distribution Table

Bad Rate/
Score band Total Goods Bads Default Rate Odds
Upto 170 351 161 131 37.32% 1.2 Cut off score is taken such that
170 – 194 2205 1345 545 24.72% 2.5 default rate is contained at 20%
195 – 201 1404 938 270 19.23% 3.5 or less
202 – 211 2682 1894 420 15.66% 4.5 Cut off Score = 195
212 – 221 3663 2729 444 12.12% 6.1 All applications scoring 195 or
222 – 239 7866 6111 583 7.41% 10.5 above would be accepted
240 – 264 11466 9120 471 4.11% 19.4 All applications scoring below
265 – 279 5573 4415 113 2.03% 39.1 195 would be rejected even if
280 and up 6845 4279 71 1.04% 60.3 they meet the eligibility criteria
Total 42055 31988 3048 7.25% 10.5
Rating Use Test
Rating Stability-Example:
IPL T20 Team Standing Points Table
2008 2009

2010
Key Issues in Calculating Historical PDs

 Sample set
 Time period: length of time
 Time period: capture of economic cycles
 Clear differences across business cycle
 Ratings consistency over time
 Academic research has shown that rating definitions do seem
to vary over time
Estimation of PD through Rating Transitions: Cohort
Approach

• PD can be estimated by analyzing rating transitions over time

• Methodology - mortality rate analysis of one-year cohorts of
companies to find the number of firms in each rating class (or
industry) in each cohort moving towards default (D)
 Data - Banks’ firm wise internal rating transition data covering at least
one economic cycle (Minimum 5 years)
• The year-wise PDs for different rating grades can be estimated by
counting the number of defaulting companies in a yearly transition
for each grade and dividing by the total number of firms in that grade
at the beginning of the year
 Thus, one year default probability of obligors from the rating grade
(say i) or industry is given by Ti,D

Ni
Estimation of PD through Rating Transitions

• However, in order to obtain a through the cycle stressed PD,

one has to take a weighted average of these marginal PDs
over the entire sample period
• Even then, one has to ensure that the rating process has
considered stressed macroeconomic scenarios
 Average one-year default probabilities are obtained by
weighted average where the weight represents the relative
importance of each year as given by the cohort size for each
rating grade at the beginning of each year
Estimation of PD: Discrete Time Measure

 The one-year default frequency (DF) of the i-th rating grade (or
industry) is given by: T i,D
DFi 
Ni
 The average historical one-year default probability (also known as the
expected default probability or expected default frequency) for the i-th
rating grade or industry (PDi) is obtained by weighted average:

n Ti ,tD
Where PDi  wit
t 1 N it

N it
wit  n

N
s 1
i
s
 Rating Transition of 572 Corporate Bonds Rated
Externally by CRISIL

One Year Average Rating Transition Matrix for the Period 1992-2009 in %
Year T+1
AAA AA A BBB BB B CCC D
AAA 96.05% 3.95% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
Year T AA 2.81% 89.74% 6.20% 0.68% 0.39% 0.10% 0.00% 0.10%
A 0.00% 3.99% 83.71% 7.12% 2.70% 0.32% 0.54% 1.62%
BBB 0.00% 0.51% 5.09% 75.83% 10.69% 1.53% 2.80% 3.56%
BB 0.00% 0.70% 0.00% 1.41% 59.86% 3.52% 7.75% 26.76%
B 0.00% 0.00% 0.00% 7.41% 0.00% 40.74% 22.22% 29.63%
CCC 0.00% 0.00% 0.00% 2.13% 0.00% 0.00% 53.19% 44.68%

Transition Matrices act as indicators of the likely path of a given

credit at a given horizon
Corporate Rating Transition-Recent Years (2011-2020)
S&P Global Transition Matrix

Standard & Poor’s data, 1981-2021, all (global) corporate obligors

CCC/
AAA AA A BBB BB B C Default NR
AAA 87.09% 9.05% 0.53% 0.05% 0.11% 0.03% 0.05% 0.00% 3.10%
AA 0.48% 87.32% 7.72% 0.46% 0.05% 0.06% 0.02% 0.02% 3.88%
A 0.02% 1.56% 88.73% 4.97% 0.25% 0.11% 0.01% 0.05% 4.29%
BBB 0.00% 0.08% 3.19% 86.72% 3.48% 0.42% 0.09% 0.15% 5.86%
BB 0.01% 0.02% 0.10% 4.52% 78.12% 6.66% 0.53% 0.60% 9.43%
B 0.00% 0.0% 0.06% 0.15% 4.54% 74.73% 4.81% 3.18% 12.51%
CCC/C 0.00% 0.00% 0.09% 0.16% 0.49% 13.42% 43.91% 26.55% 15.39%

Source: S&P Annual Global Corporate Default Study

 Moody’s Global Average One-Year Transition Matrix,
1970-2015

From/
Aaa Aa A Baa Ba B Caa Ca-C WR Default
To
Aaa 87.48 8.135 0.59 0.058 0.024 0.003 0 0 3.709 0
Aa 0.833 85.151 8.448 0.438 0.064 0.036 0.017 0.001 4.991 0.021
A 0.056 2.572 86.601 5.366 0.51 0.113 0.043 0.005 4.679 0.056
Baa 0.036 0.159 4.296 85.442 3.744 0.694 0.163 0.021 5.261 0.183
Ba 0.006 0.044 0.466 6.174 76.172 7.173 0.679 0.124 8.246 0.916
B 0.008 0.032 0.149 0.449 4.784 73.515 6.486 0.562 10.604 3.412
Caa 0 0.009 0.027 0.108 0.416 7.021 66.772 2.806 14.321 8.521
Ca-C 0 0 0.056 0 0.623 2.461 9.468 39.589 23.714 24.089

in % units

Migration rates are calculated using monthly cohort frequency

Corporate Rating Pool Migration Statistics

SG MG RG D

SG
96.67% 3.07% 0.00% 0.26%
MG
3.08% 91.81% 2.53% 2.58%
RG
0.51% 12.05% 73.08% 14.36%
D
0% 0% 0% 100%

Source: Author’s own using CRISIL data

EL=EAD×PD×LGD
UL=EAD×LGD×SQRT(PD*(1-PD))
Industry PDs (%) for Different Loan Grades, 2008-15
SL# Industry Name IG NIG Total
1 Auto & Parts 0.00 26.67 3.74
2 Chemical 1.92 12.50 3.91
3 Construction & Engineering 1.89 14.29 6.17
4 Electronics & Electricals 0.00 14.63 4.72
5 Food & pdcts & FMCG 0.00 12.07 4.32
6 Gems & Jewellery 2.56 2.78 2.67
7 Infrastructure 2.29 30.43 6.49
8 IT 1.35 15.38 3.45
9 Leather & products 0.00 12.50 6.25
10 Metal & products 1.18 13.51 4.92
11 Non Metal 2.17 8.33 3.45
12 Other 0.00 25.00 6.45
13 Paper & products 0.00 18.00 7.02
14 Plastic & products 0.00 3.45 1.41
15 Rubber & products 0.00 8.33 3.03
16 Services 0.00 16.00 2.56
17 Textiles 3.39 8.33 4.82
Overall 0.94 13.90 4.47
Source: CRISIL Rating Data
Mapping Z-score with S&P Rating-TTC to PIT mapping
Mapping of Rating Scales and to S&P Rating
Rating Transition Matrix of Commercial Loan Pool
(Exposure >Rs. 5 Cr.) of a Mid Sized bank from 2005-10
(PIT Rating?)
Table 7: Overall One Year Transition Matrix: 2005-
2010

A++ A+ A B+ B C D NPA

A++ 50.24% 23.67% 14.98% 7.73% 1.45% 0.97% 0.48% 0.48%

A+ 6.00% 53.60% 26.62% 10.43% 1.80% 0.72% 0.24% 0.60%
A 3.21% 15.22% 59.18% 16.79% 4.25% 0.37% 0.22% 0.75%
B+ 1.52% 7.48% 17.76% 62.62% 7.13% 1.17% 0.47% 1.87%
B 1.98% 5.53% 12.65% 33.20% 40.71% 3.16% 0.40% 2.37%
C 0.00% 6.67% 6.67% 26.67% 13.33% 33.33% 6.67% 6.67%
D 0.00% 0.00% 6.67% 13.33% 6.67% 6.67% 60.00% 6.67%

Taking dynamic pool of 2702 borrowers across 30 regions and 23 industries, we

compute 5 year average migration matrix from 2005-06 till 2009-10 as documented
above.
Rating model has discriminatory power. However, we notice that stability of upper
grades (especially A++, & A+) is very low!
Transition Matrices

 Transition Matrices act as indicators of the likely path of a given credit

at a given horizon.
 The rating transition matrices shown in the previous slides are useful
representations of the historical behaviour of ratings.
 The average one year transition matrices show the average one year
transition rates for annual cohorts over time (say from 1981-2016),
where annual cohort is weighted by the size of the cohort (no. of
issues).
 Study of Transition patterns reveal many interesting features of CRAs
rating system.
 ECL model component-PD
Method: Cohort Analysis (Markov Process)

CPDn=1-(1-p1)*(1-p2)*(1-pn)

Requires a long time series of historical rating data

 S&P Global Average Cumulative PDs (1981-2021) in %
Time Horizon (years)
Rating 1 2 3 4 5 6 7 8 9 10
AAA 0.00% 0.03% 0.13% 0.24% 0.34% 0.45% 0.50% 0.58% 0.64% 0.69%
AA 0.02% 0.06% 0.11% 0.20% 0.30% 0.40% 0.48% 0.55% 0.62% 0.68%
A 0.05% 0.13% 0.21% 0.32% 0.44% 0.57% 0.73% 0.87% 1.01% 1.15%
BBB 0.15% 0.41% 0.72% 1.09% 1.48% 1.85% 2.18% 2.50% 2.80% 3.10%
BB 0.60% 1.88% 3.35% 4.81% 6.19% 7.47% 8.57% 9.56% 10.45% 11.24%
B 3.18% 7.46% 11.26% 14.30% 16.67% 18.59% 20.10% 21.34% 22.45% 23.50%
CCC/C 26.55% 36.74% 41.80% 44.74% 46.91% 47.95% 49.08% 49.82% 50.48% 51.05%
IG 0.08% 0.23% 0.40% 0.61% 0.83% 1.05% 1.26% 1.45% 1.63% 1.81%
NIG 3.60% 6.97% 9.86% 12.23% 14.16% 15.75% 17.06% 18.16% 19.14% 20.04%
All rated 1.50% 2.93% 4.17% 5.22% 6.10% 6.83% 7.45% 7.97% 8.43% 8.86%

Cumulative PD captures the incremental default probability of a borrower

over a longer time horizon
Cumulative probability of default has uses in pricing long term loans and
in studying the risk behaviour of various grades over different maturity
horizon.
CRISIL Average Cumulative Default Rates, 1988-2015

Rating Issuer- 1-year 2-year 3-year

years
AAA 1431 0.00% 0.00% 0.00%
AA 3143 0.00% 0.18% 0.66%
A 4197 0.38% 2.06% 4.26%
BBB 9104 1.03% 2.68% 5.30%
BB 13855 4.07% 8.29% 12.55%
B 12277 8.08% 15.83% 21.78%
C 609 20.20% 33.62% 43.41%
Total 44616
Source: CRISIL

The survival rate up-to n years is as thus:

(1 - cn) = (1 - cn-1)(1 - dn)
where cn = cumulative default probability
dn = default probability
PD-Horizon-Matrix Multiplication

AAA AA A BBB BB B CCC D

AAA 93.66% 5.83% 0.40% 0.08% 0.03% 0% 0% 0%
AA 0.66% 91.72% 6.94% 0.49% 0.06% 0.09% 0.02% 0.01%
A 0.07% 2.25% 91.76% 5.19% 0.49% 0.20% 0.01% 0.04%
BBB 0.03% 0.25% 4.83% 89.26% 4.44% 0.81% 0.16% 0.22%
BB 0.03% 0.07% 0.44% 6.67% 83.31% 7.47% 1.05% 0.98%
B 0.00% 0.10% 0.33% 0.46% 5.76% 84.18% 3.87% 5.30%
CCC 0.16% 0.00% 0.31% 0.93% 2.00% 10.74% 63.95% 21.94%
D 0% 0% 0% 0% 0% 0% 0% 100%
Source: S&P Credit Week, January 2001

Movement from Grade

Pi to j A to D
1-year PD 0.04%
2-year PD 0.11%
3-year PD 0.20%
4-year PD 0.33%
5-year PD 0.50%
Forward Looking CPDs through Matrix Multiplication
(next 5-year horizon)
 Forward Looking Conditional PDs (next 5-year horizon)

Conditional PD is the probability that

it will default in the given year, given Conditional PD=
that it did not default in any of the
previous years
Forward Looking CPDs for a Bank (next 5-year horizon)
Forward Looking Conditional PDs for a Bank (next 5-
year horizon)

Conditional PD=
Alternate Approach-Estimation of Forward
Looking PD using forward rates

PSn= (1+fn-1)/(1+cn-1)

Follow my Excel Hands-on

 Lifetime PD through Matrix Multiplication (next 5-year
horizon)
Cumulative PD
65.%
60.%
55.%
50.%
45.%
40.%
35.%
30.%
25.%
20.%
15.%
10.%
5.%
0.%
0 1 2 3 4 5

AAA AA A BBB BB B CCC

PD Estimation through Bond Credit Spread-with LGD

 The required yield on one year corporate bond (r*) must compensate the
investor for credit risk if the investment in bond is to be preferred over
the credit risk free one-year Govt. Treasury securities (with r rate of
interest) .
 The FI manager would just be indifferent between corporate and Govt.
Treasury securities when:
PS (1  r * ) 1  r
1 r
 1 r* 
1  PD
• This analysis can be extended to the more realistic case assuming some
recovery in case corporate borrower defaults

(1  r )
r  r  
*
 (1  r )
RR  PS  PS RR
Limitations of Discrete Time Transition Matrix
Approach

 It ignores some potentially valuable information:

 When transitions take place during the calendar year
 How many changes (if any) have led to a given rating at the end of
the year (2 one notch upgrades from BBB to A and then to AA, for
example, yield the same results as 1 two notch upgrade from BBB
directly BBB directly to AA).
 They are also affected by the choice of observation times: the
beginning and end of each calendar year in the sample.
 Other Methods
 Continuous Time Measure-Using monthly data
 Duration (hazard) based Estimation-based on the times (durations) between rating
changes-i.e., the time spent by issuers in a given rating class (See Servigny and
Renault book, Chapter 2, pg. 54 and Lando and Skodeberg, 2002, working paper,
University of Copenhagen)
 These are based on the times (durations) between rating changes (i.e. time spent by
issuers in a given rating class). The simplest maximum likelihood estimator for the
constituents of the generator is: i, j
T ( h)
̂  h
 (s)ds
i
N
0
 Where T (h) is the no. of transitions from rating i to rating j during the period h
ij

and Ni(s) is the no. of firms in rating i at time s.

 Intuitively, this estimator counts the number of transitions from ith grade directly
to j grade during the period h and also considers the total time spent by firms in
a given rating class (i.e. counts the number of firms in rating i at time s). s is the
sub set of h.
 Thus for a horizon of 1 year, even if a firm spent only some of that time in transit,
say from AA to A before ending the year in BBB, that portion of time spent in A
will contribute to the estimation of the transition prob P AA to A
Other Approaches…adjusting transition matrices with
economic cycle

 The transition matrix changes with time, for example with

economic fluctuations-need to adjust the migration with
economic cycle (conditional migration approach)
 Generating the rating transition matrices under various
economic scenarios is another way to incorporate continuous
nature of transitions.
 Understanding the dynamics of rating migration-PIT (or
unstressed) PD vs. TTC (stressed) PD
 It is important to note that ignoring the continuous nature of
transitions an have substantial impact on credit portfolio risk
measures
 Stress testing and simulation are the methods to mitigate this
Problems with Internal Ratings

 Any internal rating approach raises lot of questions:

 The objectivity of qualitative judgments (consistency)
 The validity of the rating allocation process (calibration, risk mapping)
 The quality of forecast information embedded in ratings (validation
issues)
 The time horizon (point in time issue)
 The consistency with external ratings (benchmarking)
 There is a dilemma between PIT and TTC rating in Banks (will be
discussed later)
 Based on Moody’s database, Carrey and Hrycay (2001) estimate that a
historical sample between 11 and 18 years should be necessary in order
to test the validity and stability of internal ratings. Standard and Poor
(S&P) believes that a time period of 10 years should be considered as a
minimum for investment grade (IG) securities, while 5 years should be
enough to back-test non-investment grade (NIG) issues.
Pooled PD: Exposure Based Approach-Large Bank
Case
Addition in
Gross Std. Gross Fresh
Advances Assets NPA(Rs. Slippage
Year (Rs.Cr.) (Rs.Cr.) Cr.) (MPD)% MPD'
1998-99 12006.46
1999-00 14925.86 12082.86
2000-01 17061.00 13808.00 953 6.50%
2001-02 22784.00 19408.00 758 4.15% 5.02%
2002-03 24755.00 21512.00 699 3.25% 3.83%
2003-04 24500.00 21408.00 794 3.31% 3.82%
2004-05 29085.00 26464.00 788 3.02% 3.41%
2005-06 39195.00 36511.00 732 2.37% 2.60%
2006-07 53489.00 50917.00 807 1.99% 2.13%
2007-08 72287.00 69937.00 665 1.21% 1.27%
2008-09 86740.00 84423.00 873 1.23% 1.28%
2009-10 107118.06 104660.06 1033 1.16% 1.20%
2010-11 131407.00 129013.00 1409 1.30% 1.33%
LRPD 2.19% 2.32%
See the difference in 10 year LRPD estimates based on MPD and MPD’
NIBM
Pooled PD: Exposure Based Approach-Large Mid size
Bank Case

Closing
Gross Std. Opening Gross Additions
Period Advances Assets Gross NPA NPA in GNPA MPD MPD'
1997-98 5707.5
1998-99 6956.4 6458.36 397.23 498.04
1999-00 9325.53 8798.02 498.04 527.51 236.59 3.23%
2000-01 11076.41 10490.65 527.51 585.76 446.32 4.89% 5.20%
2001-02 14157.87 13206.08 585.76 951.79 601.02 5.22% 5.55%
2002-03 16524.29 15378.04 951.79 1146.25 630.28 4.53% 4.84%
2003-04 20612.3 19401.39 1146.25 1210.91 527.10 3.08% 3.30%
2004-05 25299.2 22806.93 1210.91 2492.27 1877.64 9.02% 9.78%
2005-06 35548.56 33432.25 2492.27 2116.31 372.97 1.37% 1.48%
2006-07 45394.68 43940.63 2116.31 1454.05 379.27 1.07% 1.14%
2007-08 54565.83 53285.72 1454.05 1280.11 579.75 1.28% 1.33%
2008-09 69064.72 68006.60 1280.1 1058.12 495.70 0.88% 0.90%
2009-10 84183.94 82715.19 1058.12 1468.75 1133.10 1.64% 1.67%
2010-11 96838.90 94918.36 1468.75 1921 1556 1.87% 1.90%
LRPD% 3.17% 3.37%

NIBM See the difference in 11 year LRPD estimates based on MPD and MPD’
Trend in Fresh Slippage Rate
 Pooled PD for Homogenous Buckets of Retail Exposures
(Frequency based Measure-Tracking the numbers)
 The historical default experience approach is most appropriate for quantifying pooled
PDs for point in time retail buckets. It will be most accurate when long-run average
default rates are calculated over a number of years.

Pooled PD: Exposure Method (Personal Loan Pool)

Units in Rs.
Year 2015 2016 2017 2018 2019 2020 2021 2022
Standard Advances 270 280 2850 3200 4750 5505 6206 7290
Incremental GNPA 14 18 90 200 188 240
MPD 1.24% 0.85% 2.50% 4.46% 3.43% 3.79%
LRPD=pooled PD 2.71%
Pooled PD: Frequency Method
Units in
Year 2016 2017 2018 2019 2020 2021 2022
Total no. of accounts Outstanding 9000 12000 14600 15000 16000 20700 21000
No. of accounts defaulted out of 100 126 360 680 540 590
the no. of accounts Outstanding in
the previous year
Yearly Default Probability (PDt) 1.11% 1.05% 2.47% 4.53% 3.38% 2.85%
LRPD=pooled PD 2.56%
Pooled PD Approach

 The Bank may compare Jan 2016 with Jan 2017 or may follow
financial year March 2016-March 2017 for tracking default numbers
and no. of accounts outstanding.
 Make adjustments every next year because portfolio will change.
 For example, exclude those accounts outstanding in 2016 maturing in
2017), exclude those accounts of 2016 which have defaulted in 2016.
Include those accounts of year 2016 which have not matured or
defaulted and may still be outstanding in year 2017. Include the new
accounts which have been sanctioned in 2016.
 A Bank can also track NPA exposure (in Rs. Amount) movements
(ratio of GNPA additions to 3 year average gross advances) if no.
data is missing.
 However, one should use a more conservative method
 A minimum 5 year average is required to estimate the stable PD
Retail Risk Transition Matrix
Time Period:
T+1

Super Prime Prime Near Sub Down Upgra

prime plus prime prime grade de
Super 62.6% 21.5% 11.3% 3.5% 1.2% 37.4%
prime
Prime 11.5% 48.5% 30.9% 6.4% 2.7% 40.0% 11.5%
plus
Time Prime 3.6% 20.5% 53.1% 15.6% 7.2% 22.8% 24.1%
Period
:T Near 0.8% 7.0% 36.4% 36.2% 19.7% 19.7% 44.1%
prime
Sub 0.1% 0.9% 7.9% 22.3% 68.8% 31.2%
prime

Source: TransUnion CIBIL Reports

 Pooled Cohort PD Approach: Transition Rate (%)
(New Obligors)
Initial State Termin <650 651- 701- 716- 726- Closed Default
al State 700 715 725
1-12M
<650 51.0 22.3 8.1 3.1 2.0 5.8 7.6
651-700 18.2 35.6 24.2 9.3 8.7 3.2 0.8
7001-715 8.1 15.9 30.5 17.8 25.6 2.7 0.5
716-725 4.5 8.2 14.7 21.4 48.6 2.2 0.3
726- 1.8 3.0 5.7 7.6 79.3 2.3 0.2

Source: Based on Markov Chain Analysis of a Major UK Bank Retail Data

for Credit Card Consumers
 Pooled Cohort PD Approach: Transition Rate (%)
(Mature Obligors)
Initial State Termin <650 651- 701- 716- 726- Closed Default
al State 700 715 725
1-12M
<650 44.1 23.5 11.3 4.9 7.0 4.0 5.3
651-700 13.6 32.5 25.6 10.7 14.4 2.5 0.6
701-715 4.7 11.8 37.2 18.8 24.8 2.5 0.3
716-725 2.1 5.0 14.9 30.4 44.7 2.6 0.3
726- 0.4 0.9 2.1 3.7 90.4 2.4 0.2

Source: Based on Markov Chain Analysis of a Major UK Bank Retail Data

for Credit Card Consumers, International Journal of Forecasting 28, 2012
 Retail PDs-System Level

YoY
PD-2 PD-3 Growth%
QLY Annual SICR-30 CPD-2 year CPD-3 year (July-
SL# Retail Loan Type PD% PD% DPD year Forward year Forward Sep'19)
1 Auto Loan 5.07% 16.68% 37.45% 39.18% 27.01% 49.04% 16.20% 10.50%
2 Bankcards 0.82% 2.76% 8.95% 5.97% 3.30% 9.30% 3.54% 29.50%
3 Loan against Property 2.51% 9.65% 9.50% 16.44% 7.51% 25.57% 10.92% 22.00%
4 Mortgage/Home Loan 2.81% 9.34% 26.73% 19.55% 11.26% 28.56% 11.20% 9.00%
5 Personal Loan 1.10% 2.95% 12.08% 7.10% 4.27% 12.48% 5.79% 46.20%

Source: various CIBIL reports

Retail PD Positions-All Product Categories

Period PSB PVT NBFC/HFC FinTech

Mar-21 4.90% 2.01% 3.05% 3.13%
Apr-21 4.94% 2.04% 3.96% 3.57%
May-21 5.71% 2.48% 5.08% 4.69%
Jun-21 5.52% 2.63% 4.57% 3.69%
Jul-21 5.29% 2.76% 4.59% 4.71%
Aug-21 5.35% 2.63% 4.20% 4.68%
Sep-21 4.87% 2.22% 3.64% 4.83%
Oct-21 5.12% 1.89% 3.96% 4.61%
Nov-21 5.07% 1.78% 3.47% 4.53%
Dec-21 4.85% 1.97% 3.23% 3.58%
Jan-22 4.97% 2.07% 2.97% 3.29%
Feb-22 4.78% 1.85% 3.00% 2.81%
Mar-22 4.45% 1.40% 2.34% 2.26%
Avg. PD 5.06% 2.13% 3.70% 3.88%
Source: FSR, RBI (June 2022) based on TransUnion CIBIL Data
Retail PD Trend

Consumer Loan Default Rates

7.00%

6.00%

5.00%

4.00%

3.00%

2.00%

1.00%

0.00%

PSB PVB NBFC FinTech

 Delinquency Performance Report-Roll Rate Analysis
Score Accounts Active % Current % 1-29 % 30- % 60- % 90+ % Write % Bank %
Days 59 89 -off rupt
Days Days

0-169 200 198 99% 127 64 22 10.9 15 7.50 12 6.10 9 4.50 6 3.00 8 4.00
% % % % % % %

170-179 348 300 86% 207 69 36 12.0 20 6.80 9 3.00 9 3.10 9 3.10 9 3.00
% % % % % % %

180-189 435 367 84% 264 72 48 13.0 15 4.00 12 3.20 10 2.80 8 2.20 10 2.80
% % % % % % %

190-199 466 387 83% 286 74 43 11.0 21 5.50 11 2.80 8 1.94 10 2.56 9 2.20
% % % % % % %

200-209 2456 1876 76% 1482 79 225 12.0 43 2.30 47 2.48 18 0.96 39 2.10 22 1.16
% % % % % % %

210-219 41563 3600 79% 2952 82 342 9.55 93 2.58 83 2.30 23 0.65 67 1.87 40 1.10
% % % % % %

220-229 5678 4325 76% 3676 85 389 9.0% 93 2.15 67 1.54 14 0.32 51 1.18 35 0.80
% % % % % %

230-239 7658 4598 60% 4046 88 359 7.8% 87 1.90 41 0.90 18 0.40 28 0.60 18 0.40
% % % % % %

240-249 5786 3546 61% 3333 94 142 4.0% 35 1.00 18 0.50 -- 0.00 7 0.20 11 0.30
% % % % % %

250+ 4987 2176 44% 2089 96 44 2.0% 17 0.80 9 0.40 9 0.40 4 0.20 4 0.20
% % % % % %

Total 32577 21373 66% 18363 86 1648 7.71 441 2.06 307 1.44 118 0.55 230 1.08 165 0.77
% % % % % % %
 TTC vs. PIT PDs

 From a risk-mitigation standpoint, it is not only default risk for

today or tomorrow that has to be forecasted.
 For buy-and-hold strategies (follow-up & monitoring) what
matters is default risk any time until the horizon of the
underlying credit instruments.
 Hence, an appropriate credit assessment should in theory not
just be limited to a PD at a given horizon, but also reflects its
variability (transition & volatility) through time and its
sensitivity to changes in the major factors affecting a given
company.
 Therefore, one needs to consider not only a short-term PD, but
also the estimated trajectory of this PD over a longer horizon.

NIBM www.nibmindia.org
 Hypothetical Stressed & Unstressed PD for a single
obligor over a business cycle

NIBM
Dynamics of Rating: PIT PD vs. TTC PD
 Regardless of whether a rating system relies on expert judgment or
statistical models, one can think of a risk bucket as a collection of
obligors with similar PDs.
 A point-in-time (PIT) normally groups obligors according to one-
period-ahead predicted default frequencies. It assesses the likelihood
of default over a future period, most often the period one year from
now but can also be over next two, three, five or ten years forward.
 An accurate PIT PD describes an expectation of the future and
incorporates all idiosyncratic effects and all relevant cyclical changes.
 Thus, a PIT PD corresponds to the usual meaning of ‘probability of
default’ and is unconditional with respect to unpredictable factors.
 An obligor’s PIT PD can be expected to change rapidly as its
economic prospects change.
 During the economic expansion the unstressed PD declines and the
obligor receives a higher rating. During the economic recession the
unstressed PD increases and the obligor receives a lower rating.
 PIT PDs Can Generate Large Swings in Required Capital.
PIT Rating
PIT PD vs. TTC PD…
 A through-the-cycle (TTC) credit risk measure primarily reflects a
borrower’s long-run, enduring credit risk trend. Compared to PIT risk
measures, TTC risk measures display much less volatility and pro-
cyclicality over the cycle.
 TTC PDs, in contrast to PIT PDs, reflect circumstances anticipated over
an extremely long period in which effects of the credit cycle would
average close to zero.
 Many view Basel Committee calls for the use of TTC PDs since this
implies stable estimates of capital requirements.
 The Internal Rating Based Approach (IRB) under pillar1 of the capital
adequacy framework advocates the use of risk measures with through-
the-cycle orientation to satisfy regulatory capital requirements.
 Stability is seen as a desirable attribute of strategic capital reserve.
 A through-the-cycle (TTC) philosophy groups obligors according to
stress-scenario or long run default probabilities.
 The difference:
 Broadly, point-in-time systems attempt to produce ratings that are responsive
to changes in current business conditions while through-the-cycle systems
attempt to produce ordinal rankings of obligors that tend not to change over
the business cycle.
TTC Rating
 Approach for Extracting TTC PD from PIT PD

Adjust Credit Bank

TTC PD PIT PD
Cycle-

Agency PD
Grades Mapping
Agency
PDs

Bank Adjust Credit

PIT PD Cycle +

Approaches: Regression Methods to incorporate macro conditions,

Migration Study for stability analyses, Mapping bank rating migration
with agency rating for cycle adjustment or adjustment through TTC PD
scalar etc.
PIT vs. TTC Adjustment
Mapping Z-score with S&P Rating
Mapping Z-score with S&P Rating with Different
Horizons
Forward Looking PIT-PD Methods: Z-index approach

 IFRS 9 does not provide any information on how to convert TTC PD to

PIT PD.
 However, many researchers/practitioners have prescribed applying
credit quality index (Z-index) shift factor to convert TTC PD to PIT PD.
 In this methodology, the TTC transition matrix is represented by a
normal distribution (see next figure) and by using Z-shift (right), the PIT
matrix can be obtained (JPMC Credit Metrics Approach).
 The Z-index is calculated such that once applied to each state in a
period’s transition matrix, it would minimize the difference between the
probabilities of historical PIT transition and TTC transition.
 Broadly speaking, Z index measures the “credit cycle”. In good years, z
will be +ve implying for each initial credit rating a lower than average
default rate and a higher than average ratio of upgrades to downgrades.
In bad years, the reverse will be true (i.e., z is –ve).
Rating Distribution
Extraction of Z-index: JPMC Credit Metrics Approach

• Once z is extracted, one can track its historical movement over economic cycle.
• It can also be linked to GDP growth rate using regression model to generate
grade-wise PD forecasts.
• Z is –ve during trough of the economic cycle and +ve in good time.
• The normal year may be considered as average transition (TTC).
Conversion of TTC PD to PIT PD: Approach 1 (apply z
adjustment)
 Using credit quality index (z-index)-The parameter z determines the systematic
influence on migration in year t and can be interpreted as a credit cycle
Conversion of Transition Matrix: (Markov Chain z-shift
Approach)

 The Markov property is that the probability of an object occupying any

given future state depends only on its current state.
 A transition matrix describes a set of transition probabilities that fulfil
this requirement.
 If V(0) is the vector of current state, then: V(t)=V(0)×M t
 In a more general form, M may change over time M(t)
 Markov Chain-z-shift (PIT-Good time, Z>0)
 Z-shift is a popular approach that describes a single parameter shift
that stresses a matrix to plausible forms of M(t):
Original Matrix (TTC) Cumulative Inverse Normal=z

shift (e.g.
Shifted Matrix (PIT) Cum.-Normalized add 0.10)

Z-shift is estimated by simultaneously solving Z scores and asset correlations

such that they actually predict historical default rates. (See my Excel Illustration)
The IFRS9 PD Model-Z Factor Model

 The Z factor Model

Point-in-Time Average TTC PD Z-score (or “credit

(PIT) PD cycle Indicator”)

   1 ( PDTTC )   Z (t ) 
PD PIT (t )   
 1   
 
Standard Asset Correlation
Normal
Function Common correlation between
different assets in a segment
Forward Looking Z & PD Term Structure

 The time series regression models (Multivariate) can be used to link Z index
with key macroeconomic factors or time factors.
 Example: Z=-0.10+0.45*GDPgr+0.85*Creditgr-0.40*UnempRt-0.35*Ex_vol
 One can also simply correlate Z with macroeconomic factors (like GDP growth
rate, Unemployment rate, Credit growth)
 One can also use trend fitting regression (or ARIMA) to project Z forecasts.
 As a next step, use the predicted Z indices to obtain forecasted PIT PDs.
 The cumulative PD term structure can be obtained through either matrix
multiplication (of PIT transition matrices) or using mortality analysis: CPD t=1-
(1-MPDt)×(1-MPDt-1)
 Note that the marginal PDs represent the probability of defaulting at time t
conditional on survival until t-1.
 MPDt=1-{(1-CPDt)/(1-CPDt-1)
Z-index (state of the economy)
Odd Ratio based Z index and Forward PD
Using Scenarios to project future PD

 Using these scenarios, one can model Downgrade, Upgrade and

Default Events & Re-estimate Transition Matrix
Multinomial Logit Regression-to link PIT PD with TTC
or Project PD (Predicting Rating Transitions)

100% 1
probabilities: p
1  e    x

prob of falling
into class AAA

0%
linear
transformation:
   xi

xi
e.g., profit
margin

Similarly, one can project various rating paths, from A

to A or A to A_up, A to A-down, A to Default
 Thank You

My Email: [email protected]