Behavioural modelling – prepayment of

term loans

Anirban Naskar, Genpact

Arunima Banerjee, Genpact

2017
Table of contents

1. INTRODUCTION ...................................................................................................... 4

2. LITERATURE SURVEY ........................................................................................... 7

3. ANALYSIS ............................................................................................................. 12

4. CONCLUDING REMARKS .................................................................................... 25

ABOUT THE AUTHORS............................................................................................... 26

Page 2 of 26
The views expressed here are those of the authors and do not necessarily represent or reflect the views of their
employers.

Abstract

Banks for International Settlement (BIS) and various banking regulators mandate banks to
factor prepayments for fixed interest rate loans for assessment of liquidity risk and interest rate
risk in banking book. There are various methods of modelling prepayments of such loans.
Broadly, these methods are categorised under static model, the model which does not consider
financial incentive of prepayment, and the dynamic model, the model which considers financial
incentive of prepayment. In this paper, we have simulated loan data based on various real-life
factors and used the data to demonstrate various static and dynamic modelling approaches.

Acknowledgements

We would like to thank Kaushik Das, a risk professional, and, our colleagues from Genpact,
especially Sidharth Reddy and Sadanand Tutakne. We would also like to thank many other
industry experts; with whom we have discussed this subject at various occasions.

Key-Words: prepayment, foreclosure, behavioural modelling, term loan modelling, interest rate
risk in banking book, IRRBB, liquidity risk, asset and liability management, balance sheet
management, loan book management, risk management

Page 3 of 26
1. INTRODUCTION
The core activities of a bank are to raise funds through deposits and market borrowings, and
deploy the same through loans and advances and various investments. Banks tend to take
advantage of upward sloping yield curve by sourcing funds in short term and deploying these
funds in long term. For some of the assets and liabilities, the actual maturities vary from the
contractual maturities. One such item is term loans. Although terms loans are disbursed with
fixed tenure, borrowers have the options to voluntarily prepay, in full or part. Customers tend to
prepay when they have surplus liquidity (their equity) which can be used to reduce their debt
burden, hence, interest expenses. Sometimes borrowers can refinance the loans from other
sources at lower rate of interest. Such interest benefit, net of prepayment penalty, can trigger
prepayment events. From lenders’ perspective, prepayment assumptions impact earnings,
valuations and risk management planning. For example, during prepayment process, treasury
managers are left with additional cash balance than they would have anticipated. This additional
cash can be deployed in economic usages, thus increasing revenues or reducing expenses.
Reinvestment risk, earnings volatility, valuation volatility and liquidity risk are some of the risks
which are resultant of prepayment.

In such a scenario, measurement of prepayment risk becomes important, and, is also required
for regulatory compliance in banks. It is important to capture prepayment risks in interest rate
risk from banking book (IRRBB) and liquidity risk measurements, along with other regulatory
requirements. In the latest IRRBB guideline1 from Basel Committee, the Committee expected
banks to follow nine principles for managing IRRBB. The principle number five expects, “In
measuring IRRBB, key behavioural and modelling assumptions should be fully understood,
conceptually sound and documented. Such assumptions should be rigorously tested and
aligned with the bank’s business strategies.” In the guideline, it also mentioned the Committee’s
expectations around the implementation of prepayments. Banks must determine or supervisors
must prescribe the baseline prepayment assumptions, that is, conditional prepayment rate
(CPR). This must be assessed differently for loans in different currencies. We also expect
prepayments to vary with macro-economic scenarios that are required for the purpose of
scenario analysis and stress testing. Similarly, the liquidity risk guideline2 from Basel Committee
stated, “A bank should have a robust liquidity risk management framework providing
prospective, dynamic cash flow forecasts that include assumptions on the likely behavioural

1Standards, Interest rate risk in the banking book, Bank for International Settlement, April 2016
2Principles for Sound Liquidity Risk Management and Supervision, Bank for International Settlement
September 2008

Page 4 of 26
responses of key counterparties to changes in conditions and are carried out at a sufficiently
granular level.” Accordingly, banks are to measure impact of prepayments in forecasting the
future cash flows for term loans.

As banking regulators of various countries customise the BCBS’s guidelines to incorporate local
market practices and nature of risks, regulations on IRRBB and liquidity risks mandate
incorporation of prepayment risks in various risk metrics, and ultimately, in regulatory capital
calculations. In the USA, the Office of the Comptroller of Currency (OCC), at its liquidity booklet3
stated, “Effective management and control of the liquidity risk stemming from funding gaps
depends heavily on the use of operational cash flow projections and the reasonableness and
accuracy of the assumptions that are applied.” It also mentioned that “Highly volatile or
unpredictable asset amortization (prepayments)” is one of the many factors that affect the cash
flows. In the context of assessing risk of securitised products, the Fed stated4, “The prepayment
of assets underlying ABS may create prepayment risk for an investor in ABS. Prepayment risk
may not be adequately reflected in agency ratings of ABS. Examiners should determine that a
banking organization investing in ABS has analysed the prepayment risk of ABS issues in its
portfolio.” It continues, “Prepayment risk for ABS should be incorporated into an organization's
"net income at risk" model if such a model is used.” In fact, for some of the credit risk models,
the life-time recoverable include expected prepayments by borrowers.

The following figure summarises the Basel IRRBB guidelines pertaining to the treatment of
prepayment.

3 Comptroller’s Handbook, Safety and Soundness, Liquidity, June 2012, OCC

4 Examination Guidelines for Asset Securitization, SR letter 9016a1, Federal Reserve System

Page 5 of 26
Exhibit 1: Basel guidelines for treatment of prepayments

Fixed rate loans

No
Prepayment not Is customer a
required retail entity?

Yes

Use CPR No Capable of

Introduce time
provided by baseline
buckets in CPR
regulator CPR
modelling?

Yes Yes

AND… No
Capable of
Model baseline
adding Use baseline CPR
CPR
complexity?

This paper intends to address these regulatory requirements through appropriate prepayment
modelling techniques. We have studied various prepayment models and summarised our
findings in section 2 of this paper. Subsequently, we have narrowed down some of the
approaches which would comply with regulatory requirements for banks with varied size and
complexity. Those approaches will be provided in section 3. We have demonstrated how the
selected methodologies can be applied on the simulated data. The simulation procedure will be
given in section 3.2. The application methodology and the results will be produced in section
3.3. We will present our concluding remarks in section 4.

The intention behind this paper is to demonstrate some of the regulatory-compliant

methodologies on a given dataset. The specific results, produced here, may not be relevant for
a given bank, as the nature of loan books will vary from bank to bank. However, the
methodologies will be useful for developing prepayment models which will be compliant with
regulatory expectations.

Page 6 of 26
2. LITERATURE SURVEY
Under the stated objectives in mind, we have studied various prepayment modelling
methodologies. Some of these methods are obsolete and some are presently being used in
industry. Prepayment models are broadly of two types, one in which refinances incentives are
considered, and the other in which refinances incentives are not considered. The former kind is
called ‘dynamic model’ and the later ‘static model’.

2.1 Static Prepayment Models

Static prepayment approach, in spite of its limitations, is still considered as a simple approach to
describe prepayments. We have explored the following four common static models, some of
these are used in the industry and some are only referred to for historical significance. Some of
the notable dynamic models for prepayment are described below.

2.1.1 12-year life

This is one of the earliest static models built during the 1970s. The approach5 assumes that
there is no prepayment for first 12 years and then the entire outstanding is prepaid. The major
shortcoming of this model is that the mortgage market has experienced regime shifts since that
period, and, average behaviour of mortgage has changed since then. This model does not
consider various factors like loan characteristics, prepayment incentives, etc. Accordingly, this
model is too simplistic to be used in current period.

2.1.2 FHA Experience

This is another model which is no more in use but has historical significance. This model was
developed using data maintained by Federal Housing Administration (FHA), USA since 1930s.
According to Lawrence Rosen6, the model was built using FHA data on 30-year FHA insured
mortgages and is derived from a probability of survival table of the FHA. For the first time,
seasoning (age) of loans was introduced in prepayment modelling. Frank J. Fabozzi7 observed

5 Philippe Priaulet and Lionel Martellini, “Fixed-Income Securities: Valuation, Risk Management and
Portfolio Strategies”, Page 601, Wiley publications, May 2003
6 Lawrence Rosen, “The McGraw-Hill Handbook of Interest, Yields, and Returns”, page 127, McGraw-Hill,

1995
7 Frank J. Fabozzi, “Collateralized Mortgage Obligations: Structures and Analysis”, page 22, Wiley

publications, April 1993

Page 7 of 26
that the since prepayment rates are closely linked to interest rate cycles, using the average
prepayment rates over various cycles as estimates for prepayments will not be effective. Also,
since FHA tables were published periodically, there was ambiguity in terms of the identifying the
correct FHA table that should be used in the calculation of prepayments. Also, the model was
developed using mortgage data and hence, this could not be used for other term loans, like auto
loans.

2.1.3 CPR and PSA

Conditional prepayment rate (CPR), also referred to as constant prepayment rate, model is
most widely used prepayment model for regulatory compliance, especially for IRRBB and
liquidity risk compliance. CPR is an annualised of prepayment compared to principal
outstanding. CPR is defined as,

𝐶𝑃𝑅 = 1 − (1 − 𝑆𝑀𝑀)12 , 𝑤ℎ𝑒𝑟𝑒, 𝑠𝑖𝑛𝑔𝑙𝑒 𝑚𝑜𝑛𝑡ℎ 𝑚𝑜𝑟𝑡𝑎𝑙𝑖𝑡𝑦 (𝑆𝑀𝑀) 𝑖𝑠 𝑑𝑒𝑓𝑖𝑛𝑒𝑑 𝑎𝑠,

𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑝𝑟𝑖𝑛𝑐𝑖𝑝𝑎𝑙 𝑝𝑟𝑒𝑝𝑎𝑦𝑚𝑒𝑛𝑡 𝑖𝑛 𝑜𝑛𝑒 𝑚𝑜𝑛𝑡ℎ

𝑆𝑀𝑀 =
𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔 𝑝𝑟𝑖𝑛𝑐𝑖𝑝𝑎𝑙 𝑎𝑡 𝑡ℎ𝑒 𝑒𝑛𝑑 𝑜𝑓 𝑡ℎ𝑒 𝑚𝑜𝑛𝑡ℎ

Under Public Securities Association (PSA), the major participants in the mortgage securities
market had agreed8 on a standardised method to calculate yield of collateralised mortgage
obligations. This model eliminated the confusion that arose out of nonstandard prepayment
assumptions made by different market participants. In this model, the rate of prepayments (for
mortgages) would linearly increase from 0% to 6% in 30 months, that is, at the rate of 0.2% per
month, and then be constant at 6%. This base assumption is also referred to as ‘100% PSA’.
The prepayment assumptions can be scaled up / down linearly. For example, ‘200% PSA’
means CPR increases from 0% to 12% in 30 months, which is, at the rate of 0.4% per month,
and then be constant at 12%.

2.2 Dynamic Prepayment Models

The dynamic prepayment models have increasingly gained importance as these models are
effective for cash flow modelling and also capture the sensitivity of prepayments to market interest

8 The New York Times, “Calculation Standard Set on C.M.O. Yields”, June 14, 1985

Page 8 of 26
rate, thus dynamic in nature. Most of the dynamic models consider refinancing incentive,
seasoning, seasonality and burnout effect9 to be model factors. Some of the notable dynamic
models for prepayment are described below.

2.2.1 Andrew Davidson & Co. (ADCO) Model

The ADCO Model considers the turnover and seasonality, refinance incentive, cash out effect and
credit cure effect as the factors that primarily guide prepayment.

Refinance incentive is the most important factor driving prepayments and is primarily driven by
the level of the current mortgage rate relative to the weighted average gross coupon (GWAC) of
the pool. Higher the difference between these two rates higher is the incentive to prepay.

Turnover and seasonality is considered to be the another important driver of prepayment and it
tends to be seasonal in nature with turnovers increasing during the summer.

Cash out effect considers the steady appreciation of home value in the near past and implies
the equity building up in the home. This factor depends on current ageing of the loan,
refinancing incentive and the magnitude of home price appreciation. This factor can cause
prepayments even by borrowers having no interest rate refinancing incentive.

The credit cure effect takes into account the spread of borrower ratings before and after credit
curing. A greater spread10 implies a greater likelihood to refinance post credit curing.

9 “A period of slowing mortgage prepayment within a mortgage backed security (MBS). This usually
occurs after the mortgages start to mature. When some percentages of the underlying loans fail to prepay
after an interest rate cycle, this is known as burnout. Those borrowers who did not refinance during the
first interest rate cycle are less like to do so if interest rates drop again.”, Investopedia LLC.,
[Link]
10 When borrowers have poor credit worthiness, the borrowing rate is considerably higher than the

prevailing market rate. However, with improvements in their credit quality, these borrowers become
eligible for lower rates which are closer to the prevailing market rates, thereby increasing the probability of
refinancing.

Page 9 of 26
2.2.2 Bloomberg Prepayment Model (BPM)
The BPM approach considers three independent components of which housing turnover and
refinancing components are similar to those in the ADCO method explained in section2.2.1. The
third component is the default and curtailment component unique to the BPM methodology.
Default model captures the prepayment that occurs due to sell off of the property post a loan
defaults. Curtailment shortens the maturity of the loans and reduces the WAL of the pool. Full
payoff (foreclosure) normally occurs during last few years of the tenure.

Apart from the components cited above, BPM also incorporates the loan level components like
loan size, credit score, loan to value ratio and occupancy type as instrumental variables in
modelling prepayments.

2.2.3 Solomon-Smith-Barney (SSB) Model11 12 13

Similar to the ADCO model, the SSB model uses refinance incentive and housing turnover as
important components affecting prepayment. Additionally, the model uses two other
components, a defaulter model and a curtailment and payoff model, which is similar to those in
the BPM model.

2.2.4 Wharton Model

In 1992, Zenios and Kang came up with the Wharton prepayment model14. Though, this was a
mortgage prepayment model, it could be useful for other term loans after customisations. The
model considered, refinancing incentive, seasonality effect, ageing effect or seasoning, and
burnout effect.

11 First published in Salomon Smith Barney’s proprietary analytical system, The Yield Book, in August
2000, this is a fully loan level prepayment model
12 Lakhbir Hayre, “Salomon Smith Barney Guide to Mortgage-Backed and Asset-Backed Securities”, page

549, Wiley publishers, April 2001

13 Lakhbir Hayre and Arvind Rajan, “Anatomy of Prepayments: The Salomon Brothers Prepayment

Model”, Salomon Brothers, June 1995

14 Kang and Zenios, “Complete prepayment models for mortgage backed securities”, 1992

Page 10 of 26
2.2.5 Goldman-Sachs (Richard-Roll) Model
This model is very similar to the Wharton model. The CPR is determined as a function of market
mortgage rate and contract rate, mortgagor costs, age of the loan, month of the year and the
interaction between these variables. The seasonality and burnout factor are derived as a
function of age of the loan and the refinancing incentive. The refinancing incentive is derived as
a function of the ratio of contract rate and market mortgage rate.

Page 11 of 26
3. ANALYSIS
3.1 Approach
Since the CPR prepayment methodology is the most widely used prepayment model for
regulatory compliance, especially for IRRBB and liquidity risk compliance, we focused on this
methodology, for our analysis, in this paper. However, based on the systems and infrastructure
available, banks can adopt any of the other techniques of prepayment modelling identified
above. We use simulated portfolio data to model prepayment rates. Though simulations have
been used here, we incorporated real life factors that affect prepayments and cash flows.
Therefore the simulated scenarios used here are similar to actual real life scenarios. We have
provided details of the simulation process in section 3.2.

We used the simulated data to fit a model that can be used to predict CPR numbers. The
methodologies followed for model fitting include one-step and two-step regression15 models.
Moreover, depending on the explanatory factors included in the model, we used static and
dynamic approaches where the static model considers seasoning as the only explanatory
variable and the dynamic model considers refinance incentive and seasonality along with
seasoning as explanatory variables. The details of the modelling approach and the alternate
models explored have been provided in section 3.3.

3.2 Data simulation for modelling

We simulate a portfolio of 1000 term loans. These loans are between 15 to 20 years. In the
portfolio, 50% of the loans have a flexible tenure (a prepayment alters the loan maturity while
keeping the EMI unchanged) and 50% of the loans have a flexible EMI (a prepayment alters the
loan EMI while keeping the loan tenure unchanged). For each loan, we start with a loan amount,
term and start date. For this simulation, we freeze the portfolio at certain date and observe its
behaviour for the analysis period. This also ensures than we do not consider incremental book.
The details of the simulation parameters are provided in Exhibit 2, Exhibit 3 and Exhibit 4.

Exhibit 2: Prepayment probabilities

15Various kinds of linear regressions have been used, algorithm minimizes sum of squared errors in all
cases.

Page 12 of 26
 Loan rate - prevailing rate * <1% 1 – 3% 3 – 6% > 6%

Seasoning < 5 years Low Medium

Zero Low
> 5 years Medium High

* Simulated from 10Y GSEC + Spread

Exhibit 3: Consequence of prepayment

Outcome Associated probability

Maturity remains constant, EMI reduced 50%

EMI remains constant, maturity advanced 50%

Exhibit 4: Amount of prepayment

Outstanding / loan amount 100-10% 10-0%

Low* 0 to 25% 25 to 50%

Medium 25 to 50% 50 to 90%

High 50 to 90% 90 to 100%

*As derived from Exhibit 2

Given these, we simulate the cash flows for each loan as described below.

Page 13 of 26
Exhibit 5: Procedure to simulate cash flows

For each loan i

i=0

Monthly loan payment,

EMI0A

Yes Is loan No
No Is POS>
tenure
EMI0?
variable?
Yes
Calculate principal Calculate principal Calculate principal
payment (PPMTi) E payment (PPMTi) D payment (PPMTi)B

Calculate EMIiF EMIi=EMI0 Calculate EMIiC

Calculate residual
maturityG

Calculate prevailing
refinance rateH

Calculate refinance
incentiveI

Calculate prepayment
probabilityJ

Calculate prepayment
amountK

Calculate principal
outstanding (POSi) L

No
End Is POS> 0?

Yes
i=i+1

Page 14 of 26
3.3 Modelling methodology and results

3.3.1 Data treatments

The objective is to fit a model using the simulated data. We begin with cleaning the data to
remove/replace observations that are not likely to get repeated in future.

The following figure shows the CPR rates from the simulated portfolio.

Exhibit 6: Actual CPR from loan portfolio (for 237 months)

We observe from the above figure that the CPR numbers beyond the 170th observation are not
in line with the general CPR behaviour. This occurs due to outstanding balances that are low
due to previous prepayments and loan closures. Thus, we consider observations till the 170th
period to model the CPR rates.

Moreover, we see that there is a spike between observations 100 and 125. This is due to the
underlying 10 year US GSEC yields around this period. This observation is also outside the
general behaviour of CPR rates and thus, we cap the values around this period to the 90th
percentile value of observations 51 to 170.

The final dataset of CPR numbers that is used for modelling purposes is given in the following
figure.

Page 15 of 26
Exhibit 7: After outlier treatment (170 months)

The following table also summarizes the descriptive statistics of the final CPR rates series.

Exhibit 8: Descriptive statistics (in %)

Min. 1st Qu. Median Mean 3rd Qu. Max. [Link]
0.30% 7.71% 8.19% 7.12% 8.60% 8.97% 2.45%

3.3.2 CPR Modelling

We model the CPR series using static and dynamic approaches. We use ordinary least
squares (OLS) regression to model the CPR numbers as a function of different explanatory
variables.

We calculate the model forecasting error as below:

∑𝑛 (𝐴𝑐𝑡𝑢𝑎𝑙 𝐶𝑃𝑅 − 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝐶𝑃𝑅)2

𝑒𝑟𝑟𝑜𝑟 = √ 𝑖=1
𝑛

We have modelled CPR using both, static and dynamic CPR modelling methods. Here, we
summarize the results obtained, followed by descriptions of each model type.

Page 16 of 26
Exhibit 9: Summary of results from alternate methods
Sl. Model
Model Type Forecast trends S.S.E.
No. category

1 One CPR for all bucket 2.45%

2 One step linear model 2.60%

Static

Two step linear model

where CPR increases
3 from zero and 0.91%
stabilises beyond a
point

Two step linear model

where CPR increases
4 from zero till a point 0.80%
beyond which its
growth is minimal

5 Dynamic One step linear model 1.24%

Page 17 of 26
 Sl. Model
Model Type Forecast trends S.S.E.
No. category

Two step linear model

where CPR increases
6 from zero and 0.89%
stabilises beyond a
point

Two step linear model

where CPR increases
7 from zero till a point 0.69%
beyond which its
growth is minimal

Static CPR Modelling

[Link].1 One CPR for all buckets

In this approach, we determine a constant CPR number across all maturity buckets. This is
calculated as the simple average of the CPR series and results into a CPR number of 7.12%.
For this model, we get a prediction error of 2.45%.

Page 18 of 26
Exhibit 10: Static CPR Model – One CPR for all buckets

However, we see a constant CPR number, as given by the blue line, across all maturity buckets
that averages out the actual behaviour of CPR curve over time, and is particularly erroneous
during the initial periods of origination. Hence, we fit a linear model to the CPR series using
ordinary least squares approach.

[Link].2 One step linear model

In this approach, we use a linear model to predict CPR numbers over different levels of
seasoning or months into the loan. Thus, seasoning becomes the only explanatory variable in
our model. We use a simple linear regression model for this purpose. This model generates a
prediction error of 2.60%. The following figure shows the plot of the actual vs predicted CPR
numbers.

Page 19 of 26
Exhibit 11: Static CPR Model – One-step CPR model

In the above figure, positively sloped linear curve fit reflects the actual pattern in CPR numbers
during the initial stages of seasoning, the actual CPR numbers flatten out over time and follows
a constant trend after a point of time. This cannot be captured using a one-step linear
regression model and hence we move towards a two-step regression model.

[Link].3 Two step linear model

For this approach, we split the actual CPR numbers series in two segments - early and
seasoned and fit two separate models to each segment. However, since the CPR rates are
continuous in nature, we forcibly ensured that the two segments meet at one point instead of
being completely discrete segments.

To identify the breakpoint, we identified the optimal point which minimizes the total squared
error in model prediction as obtained using the fitted models for the two segments.

While for the first segment, we used a positively sloped linear model with no constant, for the
second segment we have used alternate assumptions as below:

 Linear segment with zero slope, that is, constant

 Positively sloped linear model with non-zero constant, slope is different from that of
first segment

Page 20 of 26
[Link].3.1 Linear segment with zero constant

Using this approach, we fit the following models to the two segments:

 Segment 1: 𝑦 = 𝑏 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔
 Segment 2: 𝑦 = 𝑎

Using this approach, the 37th observation was identified as the break point. This minimised the
model error and the model prediction error came down to 0.91%.

Exhibit 12: Static CPR Model – Two-step CPR model variant 1

[Link].3.2 Linear segment with non-zero constant

Using this approach, we fitted the following models to the two segments:

 Segment 1: 𝑦 = 𝑏1 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔
 Segment 2: 𝑦 = 𝑎2 + 𝑏2 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔

Using this approach, the 37th observation was identified as the break point which minimised the
model error and the model prediction error came down to 0.80%.

Page 21 of 26
Exhibit 13: Static CPR Model – Two-step CPR model variant 2

Dynamic CPR Modelling

For the dynamic CPR modelling, we used the same model types as used for static CPR
modelling and also allowed for inclusion of other explanatory variables including refinance
incentive and seasonality.

[Link].4 One step linear model

In this approach, we use a linear model to predict CPR numbers as a function of seasoning,
seasonality and refinance incentive. We use a simple linear regression model for this purpose.
This model generates a prediction error of 1.24%. The following figure shows the plot of the
actual vs predicted CPR numbers.

Page 22 of 26
Exhibit 14: Dynamic CPR Model – One-step CPR model

[Link].5 Two step linear model

We follow a similar approach to develop a dynamic two-step model as followed for the static
two-step model.

[Link].5.1 Linear segment with zero constant

Using this approach, we fit the following models to the two segments:

 Segment 1: 𝑦 = 𝑏 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔 + 𝑐 ∗ 𝑟𝑒𝑓𝑖𝑛𝑎𝑛𝑐𝑒 𝑖𝑛𝑐𝑒𝑛𝑡𝑖𝑣𝑒 + 𝑑 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑎𝑙𝑖𝑡𝑦

 Segment 2: 𝑦 = 𝑎

Using this approach, we identified the 37th observation as the break point. This minimised the
model error and the model prediction error came down to 0.89%.

Page 23 of 26
Exhibit 15: Dynamic CPR Model – Two-step CPR model variant 1

[Link].5.2 Linear segment with non-zero constant

Using this approach, we fit the following models to the two segments:

 Segment 1: 𝑦 = 𝑏1 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔 + 𝑐1 ∗ 𝑟𝑒𝑓𝑖𝑛𝑎𝑛𝑐𝑒 𝑖𝑛𝑐𝑒𝑛𝑡𝑖𝑣𝑒 + 𝑑1 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑎𝑙𝑖𝑡𝑦

 Segment 2: 𝑦 = 𝑎2 + 𝑏2 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑖𝑛𝑔 + 𝑐2 ∗ 𝑟𝑒𝑓𝑖𝑛𝑎𝑛𝑐𝑒 𝑖𝑛𝑐𝑒𝑛𝑡𝑖𝑣𝑒 + 𝑑2 ∗ 𝑠𝑒𝑎𝑠𝑜𝑛𝑎𝑙𝑖𝑡𝑦

Using this approach, we identified the 37th observation as the break point. This minimised the
model error and the model prediction error came down to 0.69%.

Exhibit 16: Dynamic CPR Model – Two-step CPR model variant 2

Page 24 of 26
4. CONCLUDING REMARKS
A wide variety of modelling techniques can be applied to model prepayment. The prepayment
modelling approach, used by an entity, depends on specific features of the portfolio. Hence no
single approach can be recommended to be optimal. Therefore, the identification of the final
modelling approach needs to take into account the portfolio structure and model performance
using the identified methodology. Validation and monitoring are thus crucial to such models.

Many other factors, which are not covered in this paper, also affect prepayment trends. Such
factors cannot be generalized and should be given due weightages. For example, we should
understand how loans are being marketed, sanctioned, priced and serviced while modelling.
Subprime mortgage markets grew after 1980 and collapsed during the 2008 crises. The
subprime products were structured and marketed differently than the prime products. Subprime
customers borrowed despite high prepayment penalties which created a barrier for refinancing.
Subsequently, post crises, the prepayment scenario was completely transformed under the new
interest rate regime where the rates were kept low by the central bank. Despite low interest
rates, banks had to struggle with defaults rather than prepayments which should have been
‘natural’ phenomenon of high prepayment in low interest rate periods.

Page 25 of 26
ABOUT THE AUTHORS
Anirban Naskar is a Senior Manager in Genpact. He works in asset and liability management
(ALM) and market risk management for banks and consulting organizations. He holds a
Bachelor of Technology (with honours) as well as a Master of Technology from the Indian
Institute of Technology, Kharagpur, India, and received his Master of Management (MBA
equivalent) from the Indian Institute of Technology, Bombay, India. He is also a National Talent
Scholar, awarded by the government of India.

Arunima Banerjee is an Assistant Manager in Genpact. She works in predictive statistical and
mathematical modelling in the area of risk management. By education, she is a Master of Arts
(Economics) from University of Delhi, India with specialisation in Game Theory and
Econometrics.

Anirban Naskar Arunima Banerjee

E: [Link]@[Link] E: [Link]@[Link]

Page 26 of 26