PECDC Predicting Bank Loan Recoveries Agenda Challenges and - - PowerPoint PPT Presentation

PECDC

Predicting Bank Loan Recoveries

Agenda

 Challenges and importance of modeling LGD  How to construct a database  Preparation steps for modeling  Modeling approaches  Increasing the dataset size  Conclusion

2

Modeling in a qualitative world

 Wines are the best when grapes are ripe (lower acidity) and

the juice is concentrated

 In hot summers the grapes get ripe and with low rain fall

they get more concentrated

 Regression analysis by Orley Ashenfelter:

 Wine quality (auction price) = 12.145 + 0.00117 winter

rainfall + 0.0614 average growing season temperature – 0.00386 harvest rain

 Reaction by the most famous wine critics:

 “Somewhere between violent and hysterical”  “An absolute sham”  “It’s really a Neanderthal way of looking at wine”  “I’d hate to be invited to his house to drink wine”

3

LGD modeling is important

 LGD’s are equally as important as defaults  The supervisor requires LGD and downturn LGD (Very

Challenging)

 There is more uncertainty around LGD’s than PD’s

4

Uncertainty about LGD estimates

Based on an European SME securitization end of 2009.

Comparison Basel 2 estimates and CRA estimates

• 1,00

2,00 3,00 4,00 5,00 6,00 Avg Pd (1 year) Avg LGD Avg Loss (1 year) CRA 1 CRA 2

5

To model LGD you need

Predictor variables Model Para- meters LGD

• bser-

vations

6

12.145 + 0.00117 winter rainfall + 0.0614 average growing season temperature – 0.00386 harvest rain = Wine quality (auction price)

EAD D & LGD GD da data taba base str truc ucture

• rigin

inatio ion 1 ye 1 year ar prio ior t r to defau ault De Default lt Resoluti tion

• n

118 d different nt d data fields ds p per defaul aulted o d obligor Covering 4 4 snap apshots i in t time Inform

• rmation
• n o
• n
• Borrow

rrower l r level,

• Loan

an l level an and d

• coll

llate teral l l level

Borro rrower: Reason of default, Financial, Country, Business, Legal structure, Jurisdiction, Operating, Pd rating etc. Lo Loan Le Level: Facility type, syndicated, repayment structure, Seniority, Maturity date, Limit,

• utstanding amount, LGD rating

Collatera ral l Level: l: Collateral type, Book value, Market Value, Valuation date, Ranking, Jurisdiction, Ship/Aircraft / Project type , Contingent

• bligations

Cash flow

• w i

infor

• rma

mati tion: Amount, date, currency, cash flow type, source of cash flow, liquidated collateral id

7

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 31,803 7,253 1,244 507 493 287 218 195 55 26 22

Asse Asset C Class ss

Small/Medium Enterprise (SME) Large Corporate Aircraft Finance Banks Project Finance Real Estate Finance Shipping Finance Commodities Finance Private Banking

Variety of asset classes

8

10,890 8,876 5,753 3,172 2,856 2,127 1,691 1,250 1,122 795 727 657 615

Overdraft Term Loan Amortized without Balloon Term Loan Bullet Payment Guarantee and Stand By LC's Revolver/Line > 1 year Receivables Financing Operating Lease Revolver/Term Loan Demand Loan Revolver/Line < 1 year Term Loan Amortized with Balloon Other/Unknown Trade Related Documentary Credit Capital Lease

Many facility types

9

PECDC LGD observations

10

Predictor variables

 Bank loans have many different appearances:

 cash and non cash  utilization  repayment  seniority  collateral  jurisdiction

 Predictor variables selection:

 Literature  Observations in the database

11

LGD: Two stochastic variables

 LGD = EAD -/ - recovery cash flows  EAD is driven by

 facility type  covenants  bank behavior

 managing down the exposure  triggering the default

 Recovery cash flows

 quality of the obligor  Seniority  jurisdiction  collateral

12

Interpretation of the database

 Are secured loans the same as unsecured loans plus

collateral?

 What is the impact of covenants?  What is driving the recovery and where does the cash come

from?

 Is there a difference in default registration between i.e.

secured and subordinated loans?

13

Are some loans more in default?

14

Time Company Value

Restructuring for subordinated Restructuring for secured loans Bankruptcy

Modeling approach

 LGD modeling for

 before default  after default

 Approach

 Linear model violates normal and homoskedastic residuals  Log-linear model (NIBC)  Dealing with the Bimodal distribution of LGD’s (NIBC)

 Two-step approach

• Logistic model
• Loss Given Loss Model

 Survival Analysis in Time (Danske Bank)

 Alternatives

 Assuming a Beta distribution

15

Two step approach

16

Loss | Default Prob Loss | Loss Prob

Modeling steps

 Manipulating the predictor variables

 Perform calculations  Relabeling/ grouping of data fields  Changing data fields into a reusable formats (empty or zero)  Capping LGD’s between 0% and 150%  Continuous variables, discrete variables and dummy variables

 Creation of a Reference Dataset

 Exclusion of small exposures  Exclusion of over and underpayments  Exclusion of biased years

 Regressions



Backward elimination based on lack of significance and multi collineairity



Analysis of explanatory power and correlation with observed LGD’s

17

Variables

18

Variables Pre- Default Post Default Signi- ficance Continuous Variables ln Default Amount X X S Guarantor Amount Percentage X X S Collateral Amount Perecentage X X S ln Entity Sales X X

• Utilized Percentage

X X

• Principal Advances Percentage

X S Discrete Variables Number of Facilities X X S Age of Facility in Years X X

• Dummy Variables

Country group X X S Industry group X X S Seniority Level X X S Guaranteed X X S Collateralized X X S Multiple Facilities X X S Public or Private (or SPV) X X S Operating Company X X S Syndicated Loan X X

• Reason for Default

X S Advancements after Default X S

S Significant

• Insignificant

Using unresolved defaults to increase the data base

19

Conclusion

 Modeling LGD is difficult  What is needed?

 Proper understanding of the lending business  Representative, large, unbiased and detailed datasets  Increase the dataset by using unresolved defaults

 We need more research and data collection to get to the

LGD models we ideally want.

20

The best moment to plant a tree is 20 years ago, the second best is today

21