Home / Summaries / Knowledge Management and Business Intelligence / pd-dr-probabilities

Data Science Applications - Credit Risk Modeling

Q: What is the Rating Migration Analysis?

Its a transition matrix From/To of 1 period to the next. All probabilities are between 0 and 1. Sum is 1 ofc. Markov process, default state is absorbing state. Can find state in X months by using matrix multiplication Sometimes this process is questioned - downgrades tend to be more easily followed by further downgrades - duration dependence effect: the longer an obligor keeps same rating, the lower migration probability - migration probabilities tend to be correlated with business cycles

Q: What is the Rating Philosophy?

Is the term used to describe the assignment horizon of a borrower rating system: related to how quickly obligors are expected to migrate from one rating to another in response to economic cycles. PIT (Point in Time) - take into account obligor specific, (non) cyclical info - rating changes rapidly with macroeconomic situation - PD is best estimate of obligors default during next 12m - IFRS 9 Through the Cycle (TTC) - only take into account non cyclical info - rating robust with respect to macro economic situation - PD is best estimate of obligors default during credit cycle - Basel

Q: What PD calibrations are there?

Calibration deals with 3 sources of variability - sampling variability - external economic conditions - internal effects Basel requires availability of 5 years of data for calibration Calibrated PD = best estimate PD + conservative add on Uncorrected cohort DR - all customers still at risk at beginning of year t (Nt) - # of customers from Nt that default during year t (Dt) - DR = Dt / Nt Withdrawal corrected cohort DR - withdrawals: 50% - DR = Dt / (Nt-0.5Nw) PD calibration: average, maximum, upper percentile or use bionomial itnerval to account for sample variability.

Q: What rating philosophies are there?

PIT: IFRS compliant TTC: Basel compliant Both when it comes to calibratio

5 important questions on Data Science Applications - Credit Risk Modeling

What is Cumulative Logistich Regression?

Is an extension of logistic regression to deal with ordinal multiclass targets.

log(P[C<=R]/(1-P[C<=R]) = theta + B1x1 + B2x2 + .....
Parameters are estimated using maximum likelihood.

For each observation you can calculate the probabilities for all ratings.

What is the Rating Migration Analysis?

Its a transition matrix From/To of 1 period to the next. All probabilities are between 0 and 1. Sum is 1 ofc.

Markov process, default state is absorbing state. Can find state in X months by using matrix multiplication

Sometimes this process is questioned
- downgrades tend to be more easily followed by further downgrades
- duration dependence effect: the longer an obligor keeps same rating, the lower migration probability
- migration probabilities tend to be correlated with business cycles

What is the Rating Philosophy?

Is the term used to describe the assignment horizon of a borrower rating system: related to how quickly obligors are expected to migrate from one rating to another in response to economic cycles.

PIT (Point in Time)
- take into account obligor specific, (non) cyclical info
- rating changes rapidly with macroeconomic situation
- PD is best estimate of obligors default during next 12m
- IFRS 9

Through the Cycle (TTC)
- only take into account non cyclical info
- rating robust with respect to macro economic situation
- PD is best estimate of obligors default during credit cycle
- Basel

What PD calibrations are there?

Calibration deals with 3 sources of variability
- sampling variability
- external economic conditions
- internal effects
Basel requires availability of 5 years of data for calibration
Calibrated PD = best estimate PD + conservative add on

Uncorrected cohort DR
- all customers still at risk at beginning of year t (Nt)
- # of customers from Nt that default during year t (Dt)
- DR = Dt / Nt

Withdrawal corrected cohort DR
- withdrawals: 50%
- DR = Dt / (Nt-0.5Nw)

PD calibration: average, maximum, upper percentile or use bionomial itnerval to account for sample variability.