The definition of default is one of those things that sounds easy at a high level, but can get fuzzy when you get down to the details - like writing a computer program to build the default information.
Representing the high level we have Basel[452] quoted in part: “… The obligor is past due more than 90 days on any material credit obligation … Overdrafts will be considered as being past due once the customer has breached an advised limit…”
So default has a time dimension as well as an amount dimension. The “amount” is basically dollars but may perhaps be expressed as a percentage (of a dollar limit). Materiality considerations apply to the amount - credit obligations below some threshold are not material and would not trigger Basel-default no matter how long past due.
How should this materiality threshold be chosen? Need it have any relation to the level at which the bank would write off an account as uneconomic to pursue collection activities on? Presumably, no; one would choose a fairly low and stable default materiality for Basel-default purposes of, say, $100, to be as inclusive as possible and avert any argument that true defaults were being buried.
There should be no risk capital cost of having an inclusive default definition because, although it would lead to higher PDs than a definition with a $250 threshold, it should lead to corresponding lower EAD and LGD. (Although this point sounds right in principle, there might be technical objections to it depending on the maths of the formulas and how they work together.)
OTOH, the write-off level for collections is a movable feast depending on many factors - such as product type, collections technology and resources, stages of the collection/recovery process - that could not provide a stable baseline for a definition. Clearly also, the bank would want to recognise that category of defaults that was material enough to be considered a default but not material enough to go through the full collection/recovery processes.
The downside to making a default definition too inclusive - such that it flagged cases that were not “..on any material credit obligation ..” - is the dilution of the estimating and predicting power of the models for the risk components.
Basel[452] does not mention a materiality for the over-limit situation, although common sense would suggest that a materiality might apply. For example, if a customer with an $10000 overdraft reaches a balance of -$10050, must this start the DPD counter ticking? At NNB I found that modest changes in materiality in these cases made an enormous difference to the default analytics - changing the bad rates by significant factors. The real cause was that the bank didn’t credit-manage this particular product in a way concordant with Basel principles: the issue was resolved by changes to the product, which were in any case appropriate to align business practices with risk issues.
Using a percentage (of limit) as a materiality is a sensible idea on standard products, but can produce unexpected results on large datasets where there will often be some data oddities like limits of $1. Percentage supplemented by an absolute dollar minimum and maximum should provide belt & braces for these situations.




4 comments
Comments feed for this article
29 May, 2008 at 5:19 am
JonasR
Hehe, i had today a meeting about that. I’m extremly glad, that not only me and a few of my … friends has some serious problems with this. (Another controversal issue could be the IFRS and Basel incomplience…
About default definition, all you’ve written are true. I would prefer nominal values for materiality, only percentage version couse distorsion in provision on high amount loans, and makes the provision extreme sensible on mass retail quick cash loans denominated in foreign currency. (our currency has a quite big spread or volatility in short duration to the CHF, or EUR)
In my point of view there are some quite easy methods to figure out some usefull threshold. Some technical delays according to the known banking practice can be calculated (like “card sending fee”), that charched to an account, but usually not mentioned (only in the tiny-lettered part
) to the client, so the client forgets to pay it.
The same in mortgage loans in foreign currency, where the exchange rate changes makes the installments waving, so if the client simply “miss” the amount at due date, (or the installment reaches code-8 autmatic charghing process limit, or I don’t know how you call it… each bank has a dictionary
, and my english is bad… I worked for Cetelem/Paribas before, so my english is a little bit… rusty ) than simple gets into delay status. Not to mention the fact, that some clients are really on the edge of their budget, so that cases 90-day-delay comes and goes, they have the attitude to repay, but not the ability.
I would suggest to calculate the costs of collection procedures. If there are some … first unpayed score, or some scorecards, or rule-based expert-system, segmentation of soft and hard collection practice, then some homogen classes ought to be created (less spread in group, most spread between groups, SAS miner’s information value in autogrouping do the job) , but the base of calculation are theese cost numbers. After the realized EAD values has to be checked for some characteristics, its good to create a distribution and make other fields experts check it, maybe they can explane some elbows, or peaks. After make a special default definition, only for materiality modelling, long observation window, carefully censored data, and collection logs. After simulate different threshold values, I would prefer here neural network, but a simple logit could be enaugh to figure out a limit, where cured default (accoreding to IRB default definition) events are start to decrease more, while default events start to increase more, some kind of optimum, by same collection-cost-class.
That is the simpliest way. The best, that if you make categories by collection costs, you can easly create scorecards with this optimum calculations for application based on socio-demo meausres, and you can controll collection productivity.
Ok it was long… this blog is fine, I’m glad to find it. Now I can work, while I’m relaxing
thx
30 May, 2008 at 12:58 am
Clive
Delighted to be discussing - in any language - with someone who *actually does the work*! You clearly have done some advanced and interesting experimentation / investigation. I like the suggestion (if I have understood it) to “pull through” the collection considerations - using them for segmentation - and perhaps also for different default definition parameters (like threshold) in each segment.
At the beginning of NNB Basel program I did various experimentations with default definition parameters, seeking a sensible looking underlying behaviour in hazard curve shape and other measures like roll-rates. This is because the more clearly one can identify and measure a signal, the more effectively one may hope to model it - you have done this even further, considering EAD, LGD, collections.
At NNB however there tended to be a preference for exactly following the literal content of the Basel accord, which is a natural point of view in a compliance situation. The problem is, the devil is in the details, and the accord can’t and doesn’t cover all the details.
Thanks for your interesting contribution.
30 May, 2008 at 4:12 am
JonasR
Yeap, I belive the best a bank can do, to base the main risk managing systems on collection datas/procedures. The biggest experince I’d learned in Cetelem/Paribas, (cetelem is a retail only bank, mass-mass retail, no debit activity, just personal/comsum loaning, 7 application in 10 accepted/refused without human interaction!) that collection brings the “moment of truth”, the time when probability collapses into event. With some cinizm I can question, that if you kick all bad clients, you would not be able to figure out how to recognize a bad client, or even know what bad really means
The more sofisticated and logged the collection procedure, the best application/behave scroring can be approached. Unfortunatelly some banks has package based selling out on delaying exposure… No collection data, low ROC for application scorecard… Not even talking about some meanfull LGD. (the most brutally-elegant, neoprimitiv LGD modell. Selling out the outstanding on 40% after 90 day soft collection, no hard collection LGD = 0.65…. and we were prudent
) Symptoms : High volatility in provision allocation (Profit/Loss method, because of high selling out), low accept rates on applications, low separation power on scorecards, as I see it.
In my most pervert dreams, I’d whis to play with some datasets, where exists:
real good stuffs )
- application datas with some rule-based-expert-system logs, scorecards (challenger champion, or probability based… o.k. our legal system does not “support” that, but for fraud detection it is acceptable, so … our legal system is crap, that’s why every hungarian bank chilling because of Basel audit, but every of them will pass it, although they are not prepared,… can not be prepared in fact. )
- behave score, based on some prophazard modell, (there are some step-by-step How-to s, made by Stepanova
- collection scorecard on collection step-ins, first-unpayment-score
- challenger - champion collection soft-hard with proper event-logging.
- collection products, and datasets on collection product events (like loan-freezing, plan different schedule, and so. The point to handle different from regular products, to not to mess up your behave score)
In this circumstances you can make professional IRB components, and sharp modells, and then you start to figure out how to handle PD cycle, or PD LGD correlations, and some serious issues, and do it with success.
But frankly, in most cases, these problems are just another layers in the mist.
btw I would be very interesting what you think about a problem I simple can not handle.
Indirect sales, agent activity in loan granting made serious expand here, in mortgage loans of retail segment. The 20-30 % of granted loans are existing loan refinancing, because of market fights between banks, and dumbness of citizens in finance( I am sure that at least 30% - 40% of clients can be confuest with the percentige of a percentige).
The problem is given, there are no default events. All right there are some, but 90 day delay on a 30 Mill HUF ( 200.000 USD) loan, does not mean lost client at all, and most of the problematic clients look for another bank, and with a soften condition (and more money sometimes) simple does not pay for 40 days, when he knows, that he will repay the loan from another loan!
You simple dont know, why a portfolio leaves, and how good they were, how does it infulence your models, you have to censore them out, in 30-60-90 DPD modelling, because they leaves the portfolio, before default event, or after default event (no cure possibility) , but not good clients at all, most of them delaying with 1 installment so they make some …”noise” between 0-90, but that’s all, censored events, that “would be defaulter, if there were not so easy to get another loan from another bank”
Quite hard to tell anyting about the PD pools, couse you realize some kind of inverted rejection inference, or expelling inference
that a problematic deal can default without reaching 90 day, but you dont know which leaving deal leaves because of better conditions, or which start to be a defaulter.
(to say, that leaving portfolio is not the bank’s problem is obviously not enaugh. The outstands through agents are quite contraselective, the most risk-awareness banks gets the worst clients, because the agents seeks for fast, and sure money, so agents ought to be payed by portfolio’s risks, they get. But you can not make a sperativ payment if there is a 2% monthly fluctuation in loans, mainly generated by the expanding agents activity !
devil’s circle. And its quite hard to say anything about LGD too! most of the clients repay the whole loan, from another loan. And of course we sell loans too, to refinance other loans!!!!! so somehow, we take risk, and we should aware that how good portfolio we let go, and not offer better conditions!)
I have one clue, but I am really interested in, that you think about this.
31 May, 2008 at 12:39 am
Clive
You make many interesting points that I hope to find time to think about.
The problem/question you pose last has an alarming setting - so much churn that the default situation is hard to observe and model. As you say, one may not know whether the propensity to churn is positively correlated with the propensity to default (if “leaving because start to be a defaulter”
or negatively correlated (if “leaving because of being able to secure better conditions”).
Speaking from an Australian point of view, there are quite significant costs to re-financing a mortgage loan or personal loan (but not credit card). The debate about mortgage brokers (”agents”
and their value lies close to the surface here - perhaps other readers would like to comment.
In the situation you describe, where are the defaults? Can a borrower really re-finance a loan elsewhere even if they are 60 or 90DPD with their current bank? If so, one model for this situation is the children’s game called “pass the parcel” - when the music stops, whoever is holding the parcel… this is not perhaps against a background of rising real estate (collateral) values?
Or is it the case that your data history does not reach back very far, and that many of your accounts are censored by re-financing, further limiting the scope to model?
In the latter case, as we have both commented, the analytical difficulty comes from the interplay (correlation) between the two propensities (hazards): default and churn. Perhaps one could try to model them jointly (I have not tried this). A simpler approach would be to try and segment, into say 2*2=4 broad segments: high churn & high default, high churn & low default, low churn & high default, low churn & low default. Different predictors, models, and objectives may apply in each segment.
That segues into a discussion about modelling for profitability rather than for individual components (like PD). Whilst bad PD customers will be unprofitable, perversely also the very good PD customers may also be unprofitable: being “transactors” and early repayers of loans, they minimise their interest payments.
Will need to come back to your points when time allows; thanks for the contributions.