Modelling Credit Risk using Updated IFRS9 — Transcript

Webinar on modelling credit risk with updated IFRS 9 standards, discussing challenges and regulatory perspectives globally.

Key Takeaways

IFRS 9 provides a more forward-looking and timely approach to credit loss recognition than IAS 39.
Banks must focus on objective evidence rather than manipulating profit through loss provisioning.
The standard is globally adopted, making shared learning and discussion valuable for practitioners.
Understanding credit risk cycles is crucial for accurate modelling and regulatory compliance.
Interactive discussions help clarify complex aspects of IFRS 9 and improve practical implementation.

Summary

The webinar is hosted from Tanzania with participants joining globally, highlighting the worldwide relevance of IFRS 9.
IFRS 9 is a global regulatory standard for credit risk modelling, replacing the older IAS 39 standard.
IAS 39 allowed forward-looking provisions but was often misused by banks to manipulate profits.
IFRS 9 introduces expected credit loss (ECL) models requiring earlier recognition of credit losses compared to the incurred loss model under IAS 39.
The discussion emphasizes the importance of objective evidence in credit loss recognition to protect minority investors.
The webinar aims to foster interactive discussion and knowledge sharing among financial sector professionals worldwide.
The speaker explains the cyclical nature of credit risk and how IFRS 9 addresses issues seen during financial crises.
The session covers technical aspects of credit risk measurement, including portfolio equilibrium and loss estimation.
Challenges in implementing IFRS 9, such as forward-looking assessments and their impact on profit reporting, are discussed.
The webinar encourages questions and active participation to deepen understanding of IFRS 9 applications.

Full Transcript — Download SRT & Markdown

Speaker A

Good evening from Tanzania, East Africa. I know for some of you it will be morning, and for some it will be afternoon. Welcome to our webinar session. We're just going to wait for a few minutes for our moderator and the main speaker to log in.

Speaker A

speaker to log in. Uh and in the meantime, I would love it if you would uh interact through the chat and tell us you're logging in from which country so that we get to know each other. Welcome.

Speaker A

In the meantime, I would love it if you would interact through the chat and tell us you're logging in from which country so that we get to know each other. Welcome.

Speaker A

looking very much forward to uh to to the session today. And I can say a little bit idea will be that I will try to give some sort of perspective on the on the standard. Um but definitely also

Speaker A

And in a few minutes, we should be starting. Thank you. In the meanwhile, I can say we see a lot of people now assembling from all over the world. Very, very inspiring to see, and I'm looking very much forward to the session today.

Speaker A

same regulatory standard. You know all throughout the world this sort of IFS is a is a worldwide standard. Um so we all struggle with it. on the other hand you know that also means the potential for us to learn from each other.

Speaker A

I can say a little bit of the idea will be that I will try to give some sort of perspective on the standard. But definitely also the idea here is to have a bit of a conversation and a bit of back and forth because, you know, the really cool thing is that all of us are working within the context of the same regulatory standard.

Speaker A

host this webinar. Uh it is a timely and very relevant topic and as you can see it is uh we are all working in the financial se sector throughout the world. We have a number, you can get it

Speaker A

You know, all throughout the world, this sort of IFRS is a worldwide standard. So we all struggle with it. On the other hand, that also means the potential for us to learn from each other.

Speaker A

Uh Jacob, you have already been given permission to share your screen. Oh, excellent. Excellent.

Speaker A

So looking very much forward to the discussion. Yes.

Speaker A

I think it's very nice to bring the questions as they come along. Oh, sure. Sure. I will do that.

Speaker A

Welcome everybody to our webinar. Thank you. First, I'd like to extend my sincere appreciation to Jacob for accepting our invitation from the University of Dislam in Tanzania to host this webinar.

Speaker A

this talk by giving a bit of a perspective on like what is you know why did we get IFRS9 as a concept to start with. So first and foremost let me see if I can still yeah here to take myself

Speaker A

It is a timely and very relevant topic, and as you can see, we are all working in the financial sector throughout the world. We have a number—you can get it from the chat—a number of people from India, from Zimbabwe, from Pakistan, and the like.

Speaker A

beginning of the of the 20th00 an interesting thing was that the standard that preceded IS-39 allowed you to make forward-looking provisions.

Speaker A

So I'm also looking forward to a very invigorating and lively discussion.

Speaker A

we don't want banks to make these forward-looking assessments. We don't want bank to to calculate this because what we see is the risk is that the bank sort of you know they try to trick around with their profits. They try to

Speaker A

Jacob, you have already been given permission to share your screen. Oh, excellent. Excellent.

Speaker A

what is the actual object objective and evidence. Now the problem though was that if you think about what the consequences of this was was that if you have a simple depth instruments you know in this incurred loss model

Speaker A

Good. Really excellent. I should also say that I'm at the moment unable to see the chat. So if some interesting questions come up in the chat, feel free to bring them out because I think it's very nice to bring the questions as they come along.

Speaker A

then apply the the the the losses. and you would write down the value and the the issue then was that you know first and formost you should say if you are in equilibrium so if you're operating on you know a portfolio total equilibrium

Speaker A

Oh, sure. Sure. I will do that.

Speaker A

you have Or you could imagine a standard like what we do in the US CFO that you take the ECL right at the very start and you take the very the very lifetime. And from a P&L perspective, it's not going

Speaker A

Yes. Let me also move this one. Okay. Finally, again, my dearest apologies that it took a little bit of time to get this working, but now we are on the road.

Speaker A

from uh the from from either sorry again. Well, sorry. So, it will really matter from from two different perspectives.

Speaker A

So let's actually start this talk by giving a bit of a perspective on why we got IFRS 9 as a concept to start with.

Speaker A

comes in. will really matter from a cyclical perspective because what we see and you see sort of you know in the in the picture I've shown here what you what you can see is the uh the the orange line shows uh the growth in the

Speaker A

So first and foremost, let me see if I can still take myself forward.

Speaker A

when the financial crisis came this all inverted and the problem a lot of people realized was that when you have good times and if you don't have to make any loss estimation because the portfolio is not yet default okay that means that you

Speaker A

So, you know, we used to have an old standard called IAS 39, and the IAS 39 standard was developed back in the late 1900s and was applied in the beginning of the 2000s.

Speaker A

all of the losses but then you might have a different management which was then what was originally the ones that enforced it and I think here is sort of part of the of the issue that you know you don't want the credit cycle to be

Speaker A

An interesting thing was that the standard that preceded IAS 39 allowed you to make forward-looking provisions.

Speaker A

this old loss recognition method that you would have sort of you know u that that you would have this sort of you know cyclical behavior that wouldn't really work in a in in a good way. So if we move on to then talking about you

Speaker A

But the problem a lot of people within IASB felt was that this was misused by the banks. So it was actually, you could almost say, a little bit of a reversal to what we have today.

Speaker A

lot and the problem then was that a lot of people realized that you know that was sitting like you know small investors that were sitting with bank stocks they realized that okay look a bank stock is much more risky than I had realized and

Speaker A

So they said, no, no, we don't want banks to make these forward-looking assessments. We don't want banks to calculate this because what we see is the risk that the banks try to trick around with their profits. They try to orient like this.

Speaker A

is it good because the stuff in it has not yet defaulted? It's just a timing question. Eventually it will go bad or is it that rather the reverse that you know what is happening here is that the underlying quality is good and there was

Speaker A

This is not something that we want the banks to be doing or operating with. Instead, what we want the banks to be doing is to really focus on what is the actual objective evidence.

Speaker A

minority investors because remember majority investors they sit on the board they will know what the credit quality is on the on the inside but the minority investors they will not be aware of what is uh what is happening within the bank.

Speaker A

Now, the problem though was that if you think about the consequences of this, if you have a simple debt instrument, in this incurred loss model, what you would do is you would only apply losses on it if it goes beyond 90 days delinquent or this classical default definition. That would be the only place where you would then apply the losses.

Speaker A

gone bad that we know from IS-39 the incured loss standard but you could also ask what future losses is to be expected and in the old standard you had no information about this and IB then did a number the international accounting

Speaker A

And you would write down the value. The issue then was that, first and foremost, you should say if you are in equilibrium, so if you're operating on a portfolio total equilibrium from a P&L perspective, it doesn't really matter what type of incurred loss standard you have.

Speaker A

this you know how much loans are it going to go bad and it would be much better instead of you know the bank sitting and g or the investor sitting and guessing that the bank would disclose information. So this was sort

Speaker A

So you could imagine a standard where you only basically take things upon write-off, and that's the only loss impairments that you have. Or you could imagine a standard like what we do in the US GAAP, that you take the ECL right at the very start and you take the lifetime.

Speaker A

value accounting is is a very sensible thing to be looking at here because imagine that you are for example a bond trader. So you buy for example sovereign bonds and you trade with them back and forth. I mean then you know if you are

Speaker A

And from a P&L perspective, it's not going to matter if the portfolio is in equilibrium. In equilibrium, this is all going to be the same, and it's not really going to make that much of a difference.

Speaker A

if they deteriorate the bond is going to lose value when the market realize like oh these guys are not going to be able to pay us back then the bond value is going to go down. And notice very

Speaker A

But it will really matter from two different perspectives.

Speaker A

you know they they don't you know they seem to have actually some real problems they could actually be in a trouble where they might not be able to pay back the bond value is already now going to be affected so the bond value is very

Speaker A

One perspective is that it will still matter from a balance sheet perspective. So if you make very early loss estimations, you will basically increase your capital requirement.

Speaker A

fair value accounting. However, that did not really work. And the problem is that because in fair value accounting, first of you, you know, most you know, a bond has a liquid market. So that's where you get the price from. But most of the things that

Speaker A

And it will also matter, and here is where first line in particular comes in. It will really matter from a cyclical perspective because what we see—and you see in the picture I've shown here—is that the orange line shows the growth in the credit portfolios, and the blue line shows the losses.

Speaker A

you know, during the financial crisis, there were companies that had certain loans under for value accounting that actually had to give up because when the when the crisis really hit, they no longer were an active market for these

Speaker A

So what you really see is that you had a lot of years with large growth and pretty low losses, and then when the financial crisis came, this all inverted.

Speaker A

in at in in Oxford and it was interesting to discuss with a lot of acade academics there because a lot of people there argued quite strongly for fair value accounting and said that you know this this gives the most reasonable

Speaker A

The problem a lot of people realized was that when you have good times and you don't have to make any loss estimation because the portfolio is not yet default, that means you have many good years and you just basically push out loans into the market.

Speaker A

quarter and the CFO he gets he gets really scared he said okay why are we booking losses and I said no because the market is in very turmoil now and nobody wants to buy our mortgages so the the

Speaker A

And then once these loans are out in the market, at some later point, suddenly the economy turns bad and then you get all of the losses.

Speaker A

going to sell this. I mean, you're not we're not going to sell our mortgages.

Speaker A

But then you might have a different management than the ones that originally enforced it.

Speaker A

impact a company. The credit risk side, this fact that, you know, if a counterparty is of worse credit quality, I will take losses. That's a real effect. But this fact that the market is less willing to pay for mortgages

Speaker A

And I think here is part of the issue that you don't want the credit cycle to be boom and bust. You would really like it to be a very stable cycle that it would operate on.

Speaker A

So basically what we are seeking for is a solution that we're which is sort of you know it should give us relevant information. We should be able to compare the credit quality between companies. The standard should faithfully represent the economic

Speaker A

This is part of the issues that was created with this old loss recognition method, that you would have this cyclical behavior that wouldn't really work in a good way.

Speaker A

So then the next idea that was thought okay so this first you know pure fair value calculation it's not going to work. ISAB then had their next idea and and you might say okay why do you go through sort of an entire history we

Speaker A

So if we move on to talking about when the financial crisis came—let me see if this is the right way—I should type.

Speaker A

losses so they said let's do a discounted cash flow calculation and I think quite a few of you here have have seen a discounted cash flow calculation so what you do here is You take the credit losses that you take the the

Speaker A

So when the financial crisis came, the default rates were going up quite a lot.

Speaker A

discount this with a discount rate and this is the fair value of the asset. Now if you then it's a little bit tricky and said look the fair value of the asset when I originate should really be you

Speaker A

The problem then was that a lot of people realized that small investors sitting with bank stocks realized that a bank stock is much riskier than they had realized.

Speaker A

losses goes up what is going to happen this expression if the losses goes up well if the losses goes up the value of the of the asset will go down. So here I have a pretty nice way of representing

Speaker A

In particular, you had some banks that started showing big losses, but you had some banks that had not yet shown too many losses at all.

Speaker A

very upset uh about this method uh and straight out said that you know okay it will be way too complicated for us because we will have to do this very complicated equation solving uh where we will have to include both the losses and

Speaker A

A lot of investors started asking, "Okay, look, this bank looks good now, but is it good because the stuff in it has not yet defaulted? It's just a timing question. Eventually, it will go bad."

Speaker A

was what we really wanted and it's still to this day if you read what's called basis for conclusion the material that sort of gives you the rationale for IFR9 it is still what is right is the theoretically most correct value but

Speaker A

Or is it rather the reverse, that what is happening here is that the underlying quality is good?

Speaker A

normal sort of equity investments. Let me yeah to make a short go shortly go back here and say you know the DCF calculation you typically use that for equity valuations. Now if you compare an equity to a depth instrument how are

Speaker A

There was really no way to tell. There were too few disclosures to tell investors if what is in the bank's balance sheet is good or not.

Speaker A

large risks because if just one of these companies are successful, they, you know, that stock is going to go through the roof and that's going to pay back for all your losers. So, you have these sort of, you know, venture capital

Speaker A

And this is the big motivation behind IFRS 9. The idea is that we want to be able to tell our minority investors.

Speaker A

reversal because on depth, you know, let's say that I lend to 10 companies, nine of them fails, one of them is a very successful company. Is that going to work out in a good way for me? No. If

Speaker A

that's going to be really really terrible because the debt uh the company that you know that went very well, I mean they just need to pay me back what they owe me contractually. You know, I charge them like a five or 10% interest.

Speaker A

Okay, so they pay me back five or 10% interest and that's all I get. But on all these guys that failed there, I might risk lose quite a lot. So in depth investments, the upside is often very small but the downside is very very

Speaker A

large. And ISB said, "Look, that's a pretty interesting idea that if you have these investments that are that are based around this philosophy, what you could do is that you can have a different type of accounting method that

Speaker A

more reflect this, you know, limited downside method." Um and the idea was then to to say that you know okay look we segregate these out and for these we will use what we then show is this first 9 method. Um but for

Speaker A

there are actually one special type of investments for which you still use this discounted cash flow calculation that I showed before and that is for distress depth. So if I buy distress depth, I still to this day will use this uh

Speaker A

discounted cash flow calculation and it's called the pocky uh uh purchase or origination of credit impaired asset because if you buy distress debt, you often pay very little. You maybe pay 10% for the distress depth, but your upside

Speaker A

is huge. You know, if you get the customer paying, you can actually collect everything on it.

Speaker A

So we will say you know we will put those investments to the side because that they are more of a special type of IFS9 that's not so super common but it's important already to to know that you know the reason now we're talking about

Speaker A

a very special accounting method here is because we say we have this small upside very large sort of downside. Um and shortly you know we also need for this all to work we need to have very simple and very understandable cash flows. If

Speaker A

not you know you have to go back to this fair value accounting and I will show here you know sort of two interesting examples for this SPI test. SPI stand for uh for um can I show this if I go a step back? Um

Speaker A

so it says you know solely payments of principle and interest. Uh and the the SPI test here I show sort of you know if you you can have sort of you know a number of different alternatives. One is

Speaker A

I show an example where you can have some some countries is very popular with bonds where the coupons are linked to the inflation index in the country and you can say okay is is this an instrument that we can uh that that we

Speaker A

are allowed to use if 9 you know the amvertised cost for yes because the logic is that inflation is often tied very strongly to you know time value of money so it passes the test because there's no sort of you know strange

Speaker A

exposure or something um I give another example you could have a mortgage where the payments depend on the income of the borrower that would not be permitted because in that case it would almost be more like an equity invested into the

Speaker A

person that you have borrowed. So that would not be be permitted for the for the FBI context. So you know what was the solution then that was selected you know now we have seen we have seen the fair value solution we have seen the the

Speaker A

solution of this discounted cash flow model how are we now going to operate uh on you know normal loan investments and the solution was a two-step model. So you do step one you do to determine the effective interest rate. So this is the

Speaker A

first equation that you can see here. What you do is you look at the contractual cash flows that you have.

Speaker A

You then take and say those at origination if I discount them back to today they should be equal to the balance of the loan unless I make some very special element or something in the loan but they should be equal to the

Speaker A

fair value which is typically equal to the balance or a normal loan with with no sort of strange features in it and then I can solve for the effective interest rate. So I from the first part of the equation I get the effective

Speaker A

interest rate. In the first part I assume there is no credit losses at all.

Speaker A

There's no risk at all. So I calculate with the instrument as risk-f free. And the idea why that is a reasonable approximation is as I said because because we typically have you know very little upside and very large downside.

Speaker A

You only tend to give loans when the risk is not too large. So this is also the logic why this sort of method sort of works. And it also means that if you give out more risky loans, your

Speaker A

accounting will be more sort of skewed and more and harder to understand. You still have to unfortunately calculate this way, but it will be a more skewed accounting. Now then second step is that you have to calculate the credit losses.

Speaker A

So how do you calculate the credit losses? Well, you take the losses and then you discount them with a discount rate. Now if we sort of you know go back to the method we had had previously.

Speaker A

Notice that you know you still had credit losses. You had you had contractual cash flows minus credit losses. But notice that you com you calculate a joint discount rate so that all of this together became equal to the

Speaker A

balance. But in this case you instead have that you know the first part is just equal to the balance. So what it means is that when you take the gross value minus the ECL that's going to be lower to the balance. This is the day

Speaker A

one effect. So this is fundamentally why you have that already at origination you take a loss and the loss will be larger the more risky the credit is. Now that is fundamentally sort of not really correct from if if

Speaker A

you think about the valuation point of view and this is something point out because if I give out a pretty risky loan and I do that with sort of good information I will typically charge a higher effective interest rate and if I

Speaker A

charge a higher effective interest rate typically then I will will get be able to uh you know cover some of the losses using it and this will be reflected in the discounted cash flow method but it's not correctly reflected here and this is

Speaker A

why we get the day one effect so that already day one you take a bit of a loss already this is the reason why you get in that issue so you know having then set this so now we sort of you know we have been able to

Speaker A

obtain sort of you know the the expected credit losses here uh now how should we then be think about you know what is an ECL really as a concept well there are two ways you can think about Yeah, one

Speaker A

way is to view it a little bit like a form of piggy bank almost. So what you do is if you think about your equity of the bank the what you really say is you take aside a part of the equity you

Speaker A

basically lock it up and say this part of the equity cannot be touched. It cannot be distributed as dividends. It cannot be used to fulfill my capital requirements. this equity is locked because I know there will be some future

Speaker A

losses and I will use this little equity reserve in order to pay for those future losses because your losses always comes out of the equity because if your losses comes out of the the depths that you hold as a bank you go bankrupt because

Speaker A

you should always pay back your depth holders. So that's one way you view it almost a little bit of as you sort of equity piggy bank that you that you set aside as loss reserves.

Speaker A

Another way of viewing it is that that you instead view it as that you know you take this risk asset and you say okay I have an equivalent risk-free asset that is a smaller value. So I have this big

Speaker A

risk asset and I take away the sort of you know unhealthy chunk of it and I take and I have the sort of healthy part that remains. So you almost view your each loan you view as two loans one that

Speaker A

is bad quality and one that is good quality. So you make this split in the loan pile as as two.

Speaker A

Okay. So now we have set a little bit of the of the stage uh here and the idea now is to let's actually move in and start talking about how do you calculate these uh these impairments. So if you

Speaker A

think about you know the formula that we derived here what we are supposed to do and what the standard tells us to do is to take and calculate the losses the shortfalls sort of from what we are owed

Speaker A

contractually. So you take each of these shortfalls and you to discount them using the effective interest rate and that's it. So, you know, you could say, okay, here's the formula. You know, no need to continue the presentation. All

Speaker A

all is well. Of course, I think you all know that there's going to be more to this story than than just writing down this formula and me, you know, logging off the call because at the end of the day, you know,

Speaker A

this formula stares you a little bit in the face. Okay, so losses, where where do I get losses from? Okay. So if you think that you have a pool of loans for example to obtain these this loss number

Speaker A

you would have to estimate each month I have today this pool of loans each month how much losses will I take on on each of them or each lot what is the expected value of losses I think I will take each

Speaker A

month that that is a really hard number just to ballpark I mean it's very hard to get an intuition for it and the first idea to realize that it's pretty tricky is that you're realizing that If you run

Speaker A

a sensible lending apparatus and you are quite responsible as a bank as you should be most of your loans will never ever take a loss on. So most of the loans that you give out customer will just pay you. You

Speaker A

will get the monthly interest from them. Everything will be no sunshine and then you will have a small pool of loans that turn bad. This is a very important insight and and again it is very different from the equity case. In the

Speaker A

equity case, you know, most of your if you are, for example, in speculative stocks, most of your stocks might underperform and then a few massively overperform. But in depth investment, it's a reverse. Since you have very little upside, most of your loans need

Speaker A

to perform really well and then only a few of them can perform quite crappy. So what you realize is it's a minor minority on your loans where you even take any losses at all.

Speaker A

So first of all just using calculate loss for every loan is quite misleading. So we will need a method to first determine in this pool of loans which loans do I think actually is going to become those bad loans where I take

Speaker A

losses on that. That's the first thing we will have to to think about. The second thing to think about is that if you think about what's a what a loss is it's actually pretty complicated. Loss does not not mean that the customer

Speaker A

stops paying you. Because even if a customer stops paying you today, they might start paying you in the future. So you cannot just say oh loss is when for example a customer goes delinquent or stops paying me. No a loss needs to be

Speaker A

when you say okay eventually you know this money I will never ever get it back. You know there there's no future recovery. So you in the loss you need to have subtracted off all of the recoveries. And this is quite complex

Speaker A

that you would both have you know the customer needs to stop paying me but then I need to figure out what amount they will eventually I will get the recovery. It's pretty messy to do all of these things at once. So what what we

Speaker A

will also try to do is that we will apart from first we will want to determine you know what is the pool of these bad loans and then on these bad loans we need to determine the recoveries. So it would be much better

Speaker A

to divide this into pieces to better understand the problem which is exactly where we are going now. And so I started talking about the first concept that you need to identify which are these bad loans. So you know you have a pool of

Speaker A

loans today. They're all good. What you need to find out is you know first what is the probability that a loan goes bad on you? Because if it doesn't go bad on you you will not take any losses and the

Speaker A

ECL will be zero. So you need to determine this probability of the loan going bad and then we will later look at you know if it goes bad what happens then. So we need to really define this concept of bad like I I've talked about

Speaker A

it quite you know heristically how to think about it. So if you for example if you know often a reasonable way to start from is delinquency. You know once I talked to a Saudi bank and they said oh

Speaker A

we have some customers that are a thousand days delinquent but we think they will pay me soon. Yeah these customers are you know they are pretty that that's probably pretty bad. You know that that probably should be classified as you know a you know a

Speaker A

problematic customer. On the other hand we also have a lot of customers that might be one day late. You know there might be some problem with the payment system. There might be just some little bit of thing like one day would be way

Speaker A

too early. So you know somewhere between one day and a thousand days is where we can really say you know a customer really have trouble for real. And a common backs stop that is being used is to use 90 days. Now 90 days you can say

Speaker A

where why is 90 days reasonable? Well, 90 days is a pretty good measure because after all 90 days is three months and sometimes I have people I I discuss with and they say no no but I mean in our

Speaker A

case it's not weird that the customer is 90 days late and then then they stop paying or start paying us but then I often count it back okay but they are 90 days late like what attempt did you do

Speaker A

to contact them you know you had three months to contact the customer and really check out what is going on it's pretty I mean if you have not gotten any money from three And if you don't know anything about the customer, you are,

Speaker A

you know, this customer probably is pretty distressed. So this is a little bit where the 90 days come from. But remember that it's also supposed to be a backs stop. So what a good example is if you do

Speaker A

corporate lending. So let's do you say you do corporate lending and the corporate had have paid you and a few weeks later the corporation goes into default.

Speaker A

So let's say that they also do quarterly payments just for simplicity. So they do quarterly payments, they've just paid you, a few weeks later they go into default or they go into bankruptcy and you know there's nothing left. I mean

Speaker A

the company is empty. I mean you will not be paid in in let's say two months or something like that where the next qu you know coupon is to be due. I mean that's not going to come.

Speaker A

And sure you can wait 2 months and then you can wait on another 3 months for them to go 90 DPD until you set them in default. But you already know today that you know they are toast. They are in

Speaker A

serious problem. And sure you might eventually get something from the estate when they when you know when they when they try to restructure something but that's far far into the into the future from from now. So really you know you

Speaker A

should in this case already place them now in default and we often call this concept unlikeliness to pay indicators.

Speaker A

Is there an indicator that the company or the individual might really be unable to pay?

Speaker A

So then you know we have then this default concept and typically you should also try to align the default concept with it. So what I first says is that you should align your default concept within the corporation.

Speaker A

So that you you should not really have you know one default concept for capital and one default concept for credit origination and one default concept for risk management and one concept for for your 9 solution. You should really have

Speaker A

a try to have a consistent risk management default definition that you then consistently use. And it's also good business practice because I can tell you that if you have multiple default definitions within the company, it's going to be incredibly confusing

Speaker A

for the business line that maybe not work with credit risk so much to understand what is uh what is going on there.

Speaker A

So we now have identified the concept of default and I show here actually an example. It's actually not only companies or individuals that can default, but you can even have whole countries that default on their on their payments. And you know, in 2021, the

Speaker A

left picture shows projections on what was the default risk. So the probability of default, the probability that one year late, within one year, there will be a default event for different countries. And Sri Lanka at the time had

Speaker A

27.9% probability of defaulting and actually within one year as you can see from the right picture they were a default at the time within Sri Lanka. Now I have heard that you know they have they're doing much better now but it's a good

Speaker A

illustration of this principle of default. So the first part we want to identify this probability of default because we want to identify what is the bad book.

Speaker A

The second thing we want to work with is the exposure. So what we want what we mean by exposure is really to say like look I mean the company you know or the person or whoever I lent out to they've gone bad.

Speaker A

Okay, how exposed are we now if they have gone bad and and you know basically what is the maximum loss that we can take if we get no recovery at all. This is what we call exposure at default.

Speaker A

Now exposure at default actually is you know you can say oh but that's not so so hard to figure out you know and like I look at you know you you just take for example the balance of the

Speaker A

loan I mean that is my my exposure well well it turns out and I have a whole separate presentation about this topic you could talk for like an hour on the on this topic as well which we've unfortunately were not able to do but if

Speaker A

you think about it for most customers you really do expect them to give you some money before they default I Most of your customers should not default on their first payment or if you do that I think you have a rather a problem in

Speaker A

your business model. So most of your customers you actually expect to get back some money before they default. And if you get back some money your exposure is actually smaller than your balance because you're actually going to get

Speaker A

some money before they before they default on you. So what you really have to do is that in practice you have to run a forecast and you have to first use your PD model and you not only need to

Speaker A

determine what is the probability that the customer default but you also need to figure out what is the probability that the default month one is the default that the probability they default month two month three month four five six seven all the way up until

Speaker A

their contractual maturity. So you basically get sort of you know a probability a number of probabilities and if you divide it by the total PD you now will get the weight. So you will get okay what proportion of the default take

Speaker A

place month one what proportion of default takes place month two three four five six and seven and so on and now you can look at if the customer default on me month one how much is the balance of

Speaker A

the loan at if they default on month two what is the balance of the loan month three four five and then you determine so you have to make a conditional forecast you basically have to make a forecast when you condition on if you

Speaker A

the loan survives up until time T and then default. What is then the exposure there? But even that is not enough because you also have to bring in the discounting factor. And why is that? Let me give you a simple example. Assume we

Speaker A

have a loan that only have one single payment and that payment takes place after after um after a year and they have to pay us 5% of interest. So you give out 100 as a loan amount and then

Speaker A

one year later the customer is due to pay you back 105. So if you think about it now what is the amount that the customer is due on if they now default on you? Well they they are due 105

Speaker A

because 105 is what they owe you but it then become weird to say okay so my exposure is 105 but they only gave out a loan of 100. So, is my exposure larger than the loan amount I really gave the

Speaker A

customer? That feels really weird. I mean, it feels weird that you should lose more than you gave out. After all, I mean, what you lose there is really the compensation for for the risk. I mean, that's why it's 105 and it already

Speaker A

time has passed. This is why we need also to discount because what we need to do is we need to determine this 105.

Speaker A

what was it in present day's value and I mean if you discount the discount rate is going to be 5% and you discount it back it's 100 sure that's reasonable because you are you didn't get any payment so your exposure was 100%. But

Speaker A

imagine that the customer were given you a payment then actually it would mean that you know the balance would actually actually be be typically sort of smaller and then when you discount it it would be even smaller. And if you you can have

Speaker A

for example a good example is that you can have a lot of corporate bonds where you know you would say oh the company they don't pay back the principle they only pay me interest. Yeah, but then if they default on you and they have made a

Speaker A

number of print of interest payments actually because the EAD has stayed flat or the this EAD conditional EAD has stayed flat, you know, but they made that means that they have made you a number of payments that means that you

Speaker A

can also view it as these payments have been used to lower the nominal of the loan and pay back some of the principle and that means that your exposure is actually smaller. So it it's quite common to see that people just say no no

Speaker A

e it's just how much the loan balance is. But you see here that you really need to determine your you need to use your PD model. You need to forecast what is the probability to default each of the future months. Either you do 12

Speaker A

months if you are in stage one or you do lifetime if you're in stage two. We will come to staging a little bit later.

Speaker A

And then you need to look at you know the forecast of what the ED is and you need to discount it. And I have shown here just for example when Russia did their invasion of Ukraine you had a

Speaker A

number of banks that have lent out quite a lot that then you know all of these loans were effectively in default. So it showed sort of their their exposure.

Speaker A

So we now move on to you know the the the final stage here which is that if you think about it now we can think about it almost in in uh in how the loan cycle happens. What happens is that a

Speaker A

company or a customer will go into default and if they go into default we have a certain exposure and now we sit with this exposure and the key question is how much will we be able to recover and this proportion how much are you

Speaker A

able to recover what percentage are you able to recover this we call the loss given default or the LGD so here I hope that you see now why we get this model the PD the EAD then the LD model because the PD tells us what is

Speaker A

the good book versus the bad book the EAD tells you if there is a default what is my exposure ergo how much money have I not been paid back so it's not just the balance it's how much money sort of

Speaker A

remains still in the system when the default happens and finally the loss given default it tells me okay what proportion of the money am I unable to recover back upon and if I and these three parameters they were already well

Speaker A

known and and you had them in the in the IRB standard and they often call the the boss of parameters. Now it's important to notice that you know the IRB um PDE LGD is something different than IFS9 PD

Speaker A

but it's still IRB where it sort of comes from even though the interpretations and the the exact values are different for the for the PD ED and LD in an IFR9 context but it's still sort of the origin of this wave because

Speaker A

it's important to say that neither PD nor ED nor LD is mentioned in the first standards if talks a little bit about risk of default. Um, so you effectively they are talking about the PD, but they're talking about this somewhat

Speaker A

informally and they even stated that you don't strictly need to use probabilities. But I would say, you know, if you don't use this model, you often quickly get into much more complexities. Most of the auditors only see it on this way. Most of the

Speaker A

regulators never see something else than this. So this is really the form that you have to use. But it's not because the accounting standard requires it because there's already such a strong risk practice around PDA and LGD.

Speaker A

And now we have sort of set up uh sort of you know the the the the whole system here. And you know one sort of you know little bit of a fun way of showing you know how you can think about these

Speaker A

different parameters is that you know you can plot the PD on one axis and you can plot sort of the ED and LGD on sort of one axis and then you can say so let's for example say that you know the

Speaker A

probability is low the probability of default is low and you also have that you know the the impact is low as well I mean this would be sort of you know your prime everyday you know small good retail customers. You can then say that

Speaker A

you know let's say that we instead look at customers where you know the PD is actually it's actually pretty pretty high. Um but on the other hand it's still sort of you know a small exposure.

Speaker A

I mean a good example of this is uh for example subprime customers. So if you for example do mortgage lending to quite risky customers you know as long as a mortgage as long as the property is still worth quite a lot you know you're

Speaker A

not taking so large risk. Now of course there is an inherent risk in itself because if you have estimated LGD and the LGD then changes. Let's say that certainly for example property prices fall then the LGD might become much

Speaker A

larger because the collateral doesn't cover as much effectively this is what a lot of US banks did as their mistake during the financial crisis. Like what they saw was property prices keep going up. So even if someone default, that's

Speaker A

not a problem because I get a house that is worth more than the than the loan. I sell the house, I recover my investment.

Speaker A

Okay, that only worked as long as the housing market kept increasing. But instead, when the housing market crashed, suddenly the houses became worth less than the loans and you started taking quite material losses. So you should always be you know in this

Speaker A

category where you take an high PD and you say no no that's fine because I have good uh you know good collateral. Yeah.

Speaker A

Is the collateral good right now or will it be good when you really have to go and collect on it?

Speaker A

Up in the upper corner we will have so you know the the one with low PD and high uh ED and LD. It really shows you know the the the single when we have sort of you know you know maybe very

Speaker A

safe customers but we could actually be quite risky. A good example here is something like a government bond. I mean government bonds are very low risk. I mean the PD is very very low but it's actually not so clear like what you know

Speaker A

what legal ability you have to recover on them if a company were or if a uh if a country were to to default on their their depth in particular if they have denominated the depth in their own currency you know then it might be that

Speaker A

they just declare you know they just inflate it away they might declare sort of you know that your bond effectively become uh become worthless so here it's more almost this sort of you know single concentration one and then finally sort

Speaker A

of high PD, high EDLDD. I mean, this is typically sort of high risk that you are not uh not sort of willing to take.

Speaker A

So, so Jacob, um maybe we'll take a little bit of intermission and put into the interaction uh in the chat.

Speaker A

I mean, that's an excellent point to do. So, do you see that I have a little bit uh thought it now before we before we go? Thank you so much for for interrupting here.

Speaker A

Okay. Okay. So there was an interesting discussion on the charts uh about the parameters and I know you have just talked about it a few minutes ago uh the collateral the value or rather the function of the collateral

Speaker A

uh into the in the calculation of ECL particularly the LGD um in developing countries including my own with Tanzania in banking system and even micro finances you can't take a loan without having a collap collateral which is of

Speaker A

equal value to that loan. So how do they affect the value of the LGD when you're doing the actual calculation and a number of uh contributors here Ahmed and Nino pointed out that the type of the collateral and

Speaker A

the quality of collateral has a direct uh effect on the calculation of ECL. What do you think about that?

Speaker A

Super super good points raised up and and and really important uh questions. So I I will start with a little bit of an example and I hope that that will illustrate some of the points I will make here. So I once gave a basically a

Speaker A

version of this presentation to a Pakistani audience and there was a guy there from a Pakistani bank and he said oh yes we give out quite a lot of loans to farmers to buy goats. So the farmer he buys goats and the goats are

Speaker A

collateral for the loan and you know the goats have to be of you know equal value to to the loan as as you said here. So you know you can feel that oh this is a pretty good business model but what we

Speaker A

have noticed is that if the farmer doesn't pay us and we have to go and foreclose on the farmer and we have to take the goats because the farmer have you know struggled financially the farmers might not have had money to pay

Speaker A

for medicine for the goats. you might not have have food for the goats. So what we get is very very uh malnourished goats and they are not feeling well and their value is actually very low. So we even though on paper they you know we

Speaker A

had equal value of the collateral to the loan actually whenever we foreclose on someone we actually take a pretty sizable loss because the collateral value is lower. And my point here is that we often this you actually see for

Speaker A

example similar in the mortgage market here in the Nordics as well. So it's not uncommon if I lend to a property you know the property looks a very nice property and uh you know some people buy it and they get for example here in the

Speaker A

north in Sweden you can get an LTV of 90%. At most so you know then you say you you might say oh we are pretty pretty golden here. Look, I mean the collateral is worth, you know, almost like, you know, 10% more than the loan

Speaker A

that uh that they take. Okay, so we're going to have no problem because, you know, the property markets is not going to go down by 10%. I mean, that's completely unfathomable. And then when you have to foreclose on someone and you

Speaker A

come in and you check the property, what you might discover is that they started renovating the property and they ran out of money. and you come down into the basement and the basement looks like total crap and there are wires hanging

Speaker A

from the roof and the the house is in terrible shape. And of course, when you go out now and sell the house, you're going to sell that for much less than what you what you really thought that it

Speaker A

was worth. And that means that you're going to take some uh some real losses.

Speaker A

And and it ties back to this, you know, you talked about the concept of sort of, you know, collateral of equal value.

Speaker A

When we talk about value collateral, what is the value we are talking about? And I think a common mistake we make as bankers is that we too much tend to focus on the fair the current fair value of the collateral. You have a

Speaker A

house right now. What could you sell that on an organized market? You have these goats now. What could you sell the goats for right now? That's not the question you should be asking yourself.

Speaker A

Sure, we might have some regulatory requirements on LTV and so on. And for these reasons, it might be pretty nice to, you know, calculate with fair value.

Speaker A

But and remember, if everything goes well, you will never ever have to think about the value of either the goats or the house because if everything goes well, it's the customers going to keep the house. But it's only when things go

Speaker A

bad that the value of the collateral matter. And that might be when the value is also distressed. So when you do LGD calculations and when you include your your calculate your LGD you really need to think about distressed valuation. So

Speaker A

you need to think about you know could it be so that when I go and I foreclose that also the collateral could be in worse shape than I thought normal. And this is really important to think because this could be a channel where

Speaker A

you can make quite some losses on because you have only thought about you know you lent and it was sort of you know nice weather and you said oh yes I I I lend to this beautiful house it is

Speaker A

valued at this very high price. Great. And then when you go and foreclose on the person that might be the moment where you know you discover that there is actual issues and the house might have deteriorated and and so on because

Speaker A

there can be a fair amount of correlation between that someone defaulting might also not really be taking care of their of their house. So sometimes we you know a simple solution for this is that we introduce for example haircuts and so on. But you but

Speaker A

you can also you know collect the data um is really the best recommendation where you have done distressed sales and you look at what did you sell things for and you compare that with what the value was that you know the sort of normal

Speaker A

fair value and that's really sort of something that you need to bring into your LGD calculation.

Speaker A

Yeah. So in in short you're telling us that there's no one method cures it all at the end of the day. Uh ECL is forwardlooking. So you are not sure that if the value of collateral would remain the same up until the end of the very

Speaker A

that loan or at the time that you actually need the collateral. What I'm saying is more like don't forget that what you are when it comes to LGD and it comes to collateral. What you're concerned with is what is the

Speaker A

value of the collateral in a distressed situation. That's really what should concern you and that's really what you should bring the mentality. Similar with cars. I mean okay don't think about what is the value of the car right now think

Speaker A

about what is the value of the car from someone unable to pay the loans is that a person who takes care of their cars is that a person who have made sure that the car is in in prime shape probably

Speaker A

not this is the reason why you have to calculate these haircuts on the collateral value and you and you know the best way to do that is to try to collect data on this and really include this factor but otherwise at least use

Speaker A

some sort of of haircut multiples on the value of the collateral. Okay. Uh I hope that clears some of the air. And there's another question before we go you go on with your presentation from uh my apologies if I don't mention

Speaker A

it as it's supposed to be. Not Drom and N is asking how is EAD for credit facilities calculated. I'm assuming this is for micro finances and they like micro finance and banks. how is EOB calculated uh for these facilities?

Speaker A

I I will actually come come back to that. I almost I will save that question because I have a slide later on where I where I talk about that and I and I actually want to develop devote some

Speaker A

time. So I I I promise to this is not a copout for me. I really want to because this is a really really great question and it is one of the good examples where I would argue that the IRB and IFR9

Speaker A

standard is quite different from each other. So uh so so really really good that the question come up but uh let let's uh let's come back to it in just a little bit of a while where because what's now going to happen in the

Speaker A

presentation is that we're going to start looking at each of the parameters and look at how you estimate them. Did we have anything else interesting in the in the chat?

Speaker A

Yeah. Um I don't think so. There's a question about FRM. Uh this is a professional exam for financial risk management. The concept of uh coloration breakdown. Um I guess this is a general question and not related specifically for

Speaker A

but then we can then we can come back to it presentation. Okay. But really great questions uh really great questions. So uh so far so okay now uh now we move into you know the next phase which is let's look now

Speaker A

at the parameters you know we have set the scene we have the PD we have the ED we have the LGD. Now let's deep into how can you do the PD modeling. So first and foremost, you know, if I go a little bit

Speaker A

and and you know, if I go a little bit back in history, h and you know, we are all different countries here. Some of you guys have worked with I first 9 for a really long time. Some of you guys

Speaker A

have just now started to enter I first 9. Some of you have not even entered I first 9 yet. But I can tell myself from from experience that when I force 9 came it was quite a shock here in Europe and

Speaker A

a lot of institutions were not really really ready for it and and it was very very rushed in many institutions. I mean the irony was we were quite you know we had quite a lot of time because the

Speaker A

standard came in 2014 and they weren't supposed to apply until 2018. So if you started and you did your homework like you were supposed to do no problem. The issue was that many institutions didn't do this. They wait and they wait and

Speaker A

they wait and they wait and they waited and then they had to rush everything. So first generation of IFRS9 models were very very focused on thinking about you know sort of sort of binary questions because if you do loan origination for

Speaker A

example you you are to give a loan to a customer um then what is quite common to do is that what you basically say okay I mean either this customer defaults or it doesn't default so it's a very binary

Speaker A

thing and you might have for example you might take an 18month window or a 12-month window forward. So you say 12 months of the default have the customer defaulted or not very very common methodology similar in the IRB because

Speaker A

many banks here in Europe especially the larger banks were IRB certified so this you know capital model standard but but also in IRB you focus on a 12-month window you say will there be a default within 12 months yes or no binary

Speaker A

question and for binary questions logistic regression is sort of the real workhorse. Um, and it's often called, you know, you often call these scoring models and so on. I would argue fundamentally that scoring models uh and these logistic regression

Speaker A

models are not really appropriate for if 9 and we will we will see a little bit why. But but the key problem is that if you use these type of models in IFS9 you you will have to do quite a lot of

Speaker A

additional work because if we think back a little bit when I talked about the EAD I pointed out that you don't only need to know if a customer default or not.

Speaker A

It's not just a binary thing. You need to know when they default and this really matters. Imagine you know I I sometimes takes a very extreme example.

Speaker A

There is one loan that was given out back in the in the 1600s in the Netherlands. They wanted to build a flood wall protection. So they wanted to protect theirel from the floods. So they gave out a loan which was perpetual. So

Speaker A

it would pay a coupon every month every year 5% forever that they have been paying that loan now for over 300 years.

Speaker A

Now imagine that the customer you know imagine that you sit as the risk manager. you know they they come to you these people from the Netherlands and they want to borrow money from you and you say you know when you calculate your

Speaker A

EAD you say okay what if they default on me the first month or they default on me after 300 years I mean obviously there's going to be a huge difference between these two scenarios in both case you know you know they they owe you 105

Speaker A

because they owe you 100 principle plus five the interest but in the first case you have gotten no money at all from them and in the second case you have gotten you know thousands of thousands of money because you were able to get

Speaker A

all the interest payments. So of course you know in the second case there's very little of the loan actually effectively remaining and you know you have very good credit protection and in the first case you have not so much and this

Speaker A

really matters for because it means that when you build a PD model you not only need to figure out yes or no will it default you need to figure out when will it default if it default when will it

Speaker A

default because it matters for the size of the exposure that that you will get.

Speaker A

Um and that really as we say that that that really matters. So what you really need as we say here you not only need the PD you also need a time series you need to figure out what is the

Speaker A

probability of default the the first month the second month third month four months fifth month and so on and you also need to be be able to build sort of you know a quite sort of you know model

Speaker A

that can do some some updates and similar you want to also be able to extrapolate sort of you know long-term behavior and what you also would like to know is and here is also another important thing you know sometimes when

Speaker A

we talk about you know let's say that you do a window for example of 18 months sometimes you see that people just look at the end of the window and they say you know I check after 18 month has the

Speaker A

customer default or not or after 12 month have they defaulted or not and and that's what they determined but if 9 tells you that what you have to look at is that you need to look at is there any

Speaker A

default within the period because from a PD point of view what you care about is is there any default at all and as I show here in the picture it's not the same thing to have a customer paying

Speaker A

every month to have a customer paying defaulting curing and then paying again these are different from a PD perspective and the PD is different because in one of them you have a default and in the other you have not

Speaker A

and it will matter when we look at significant increase in credit default or significant increase in credit because it is dependent on the PD and not on the other parameters. So you need a you you need something that really

Speaker A

have a more of a longer perspective compared to just looking at this binary question.

Speaker A

And one very nice tool that I myself often tend to to advocate I mean there are many models but one very nice model to deal with this is the hazard model.

Speaker A

So the hazard model it focus on the intensity of risk. So what it really means is that per unit time what is the risk that the customer will default on you. So you really are thinking about you know almost like a pool of loans and

Speaker A

you're thinking a survival perspective and you say what is the probability that you keep surviving each small little incremental unit and what you then look at is how this hazard change over time and it's often this hazard it has to be

Speaker A

positive because there's always a positive risk. So one way to to be able to mathematically model it quite convenient is that we look at the log hazard because the logarithm of a positive number is a real number. So

Speaker A

then you will have this this z the log hazard and you can write then you know what you do then is that you calculate this integral. So e to the power of of z e to the power of of of

Speaker A

setta then over over time. So this you call the hazard model and the hazard is also a very nice scale to use. So what you can do is that you can transform uh to use the hazard scale is often much

Speaker A

better because the problem with PDS is that if you for example plot PD on the x-axis you know most of your loans will be very low PD. So it will only be a small small part of the of the axis

Speaker A

while the large part of the axis will be made up of these very high risk loans.

Speaker A

But the log hazard scale is quite nice because minus 10 here would represent half a basis point. That is typically almost as low as you you can estimate risk. This is incredibly low risk. So a basis point here just for those who

Speaker A

doesn't know a basis point is a hundreds of a percent. So this is half of a hundth of a percent if you have you know 0.45 basis points.

Speaker A

uh and on the other hand so so a good example just to get a good uh reference is that um around you know 0.3% or somewhere like that is typically what we tend to view as as investment grades.

Speaker A

So that would typically be around 30 basis points. So you know but then then that that's also a way of thinking about it that you know if 30 basis point is investment grade 0.45 45 basis points would be incredibly small.

Speaker A

But then if you go all the way to log hazard equal to zero then you have a PD of 63%. That is quite massive. You could of course continue to scale. You could have something that's smaller than minus

Speaker A

10 in log hazard and something that's greater than zero. But it's a very nice scale. I actually have this at my my desk. I have like a log hazard scale because I often think about risk in this sort of scale and moving you know one

Speaker A

unit in this scale is often a pretty hard pretty you know reasonable way to think about sort of notches in risk and so on and what you can do is that you can do what you often call the weight of

Speaker A

evidence model. So the idea is that you know how are you know you will have when you build your model you will need to have multiple different variables typically and you might then say okay so how should I put the variables together

Speaker A

one very nice approach is that what you do is that you build multiple models so what you do is you take each variable in isolation let's say that you have for example days delinquent and you have another variable that is utilization of

Speaker A

for example a credit line and so on so you have list of different variables and you then the idea would be to say look I mean I analyze each variable independently. I build the best model I can. I build the sort of survival model

Speaker A

and with this log hazard model for each of the of the different variables and then I say okay now I have basically built the best possible model with one variable and with another variable and another variable. Now, can I combine

Speaker A

these models and build one really nice joint model? And this is what you often call the the weight of evidence score.

Speaker A

So, what you do is that you do a pre-processing. You determine each individual model and then you weight these models together with the these omega these weights and then you determine the weights and perk. And this method is quite nice. I mean there are

Speaker A

many other ways that you can do things but it's a quite nice method because it means that you can look at each variable independently and really think about it and I used to sit as a model validator and I can say I really appreciate when

Speaker A

when modulers would come with this because they could show me one variable and we could talk about it and how it behaves and they would come with one variable and we could talk about it and how it behaves and so on. So when we

Speaker A

aggregated together things it wasn't really so complicated. Then we would just basically weight the models together and take sort of you know the best of of uh all these worlds and the weights then tell you basically how important are the different variables in

Speaker A

combi you know compared to each other. So sort of highest weight will be the most important variable and lowest weight would be less important variables and this is again a pretty nice way of sort of making it interpretable. I can

Speaker A

also tell you that you know if you run IFRS9 models in production what often will matter is that you know suddenly one month PD spikes it goes up ECL goes up your CFO comes running you know down through the floors and he says you know

Speaker A

okay like tell me what what's going on why is suddenly this going up I need an answer now and you are when you work with 9 I think those of you who work with it knows that you know time is of

Speaker A

essence it's common that you will get you know your data to calculate on maybe first or second day of the month fourth day maybe even fourth day before lunch CFO wants to book the numbers in there is no time to wait you need to very

Speaker A

quickly determine is there a problem in the data is there some new trend going on like what's really going on and then having built the model in this way where you have these subm models that you have weighted together is very convenient

Speaker A

because you can then very quickly run each of the subm models and check you know what is the score for each of the variables And if you see some of them that stands out. Okay, look, let's deep dive into

Speaker A

that variable, what's going on? So, so I mean interpretability, we often talk about this, you know, in modeling. Oh, how should we prioritize interpretability versus accuracy? I often say modularity is underestimated.

Speaker A

Modularity is this idea that I can very quickly break things into pieces, check each of the pieces, and then determine which is the faulty component and go from there. And this is really why this sort of you know uh waiting model works

Speaker A

quite nicely. Let's see. Yes. Another thing I really want to talk about is customer segregation because you can also ask yourself a little bit about you know okay like which you know which type of of group how and how you should sort of you know

Speaker A

segregate your your model. Um, at the end of the day, you know, there's always a balance. You know, if you put all customers together in one big huge model, that's not really going to often be be that good because customer might be

Speaker A

might be very very different from each other. And I say I used to work in the unsecure lending industry for I worked for my my current bank and there we would have two different channels basically that you could get into the

Speaker A

into the bank. So either you could come directly to us in the bank and you for example either called us up or went to the web and said I would like to apply for a loan. These customers that came to

Speaker A

us directly they were of pretty good quality because they had made the their due diligence and they said this is the bank I like. I would like to to go here.

Speaker A

But we also had a lot of loan brokers. So a loan broker is basically a comp an aggregator. So you you as a customer go to the loan broker. Loan brokers will send out offers to many banks and then

Speaker A

the winner bank will be the ones who offer the the lowest interest rates. And what we noticed was that the loan broker customers were often of much lower quality. And the reason for this was often that these guys were you know they

Speaker A

were much more sort of you know out there in the market trying to find find sort of loans. They were much more sort of you know risktaking. And you know since these customers were so different, you could argue sure you could have put

Speaker A

everything together but we had quite a lot of good data on both of these groups and then it really made sense to make it into different models. And this is really my advice like what sure you should not build you know a billion

Speaker A

models but you should really be careful on putting together groups where the business treat them differently. And often you know sometimes I often get this or you know the people say oh how do I know which model how to segment and

Speaker A

I say the absolute best way to know how to segment is to go into your credit department and ask how are you seeing difference between customers because in this case you know there was a whole different department that dealt with

Speaker A

these two sets of customers because people said oh the direct the the you know the customer that came to us the D2C customers oh they are so different from these broker customers. Yeah, sure.

Speaker A

But if we know that as a business, we probably should have them as separate models.

Speaker A

Another thing that really matters and I say the I talk about is the delinquency behavior because if you look at you know for a behavioral model what is the most predictive variable typically the most predictive variable is how delinquent is

Speaker A

the customer. If you have customers that are delinquent the more or less the only question that matters is how delinquent are they? Most of the explanatory power can more or less be explained by you know number of days delinquent. Are they

Speaker A

30 days delinquent? Are they 60 days delinquent? Are they 70% 70 days delinquent? Okay, they are about to fall over the edge now because remember 90 days delinquent and they are toast. They go into default. Uh so the then you know

Speaker A

the the goal here would be that you know delinquency days gives you quite a lot of information. Now imagine that the customer up to date. Okay. What then?

Speaker A

You know the days delinquency is zero. So is all customers up to date the same?

Speaker A

No, they're not because typically delinquency information matter here as well. But it matters in a very very different way because what matters is how delinquent have customers been in the past. So did they were they delinquent last month? Were they

Speaker A

delinquent three months ago? Were they delinquent 12 months ago? Have they been delinquent multiple times the past 12 months? These are pretty pretty important questions to ask because it again is a very very predictive variable. So for customers up to date

Speaker A

past delinquencies is also a very predictive variable. Now notice the difference. So if I put these two groups you know the the delinquent customers and the non- delinquent customers if I put them in the same model you know I

Speaker A

will put a lot of emphasis on being delinquent or not delinquent and how many days delinquent. Oh that's a really important variable. Yeah, but that variable tells you zero for the groups that are not delinquent. And it and it

Speaker A

means that it's much harder then to group these two together. And if you have the data, it might make sense to have different models for delinquent and non- delinquent customers. Now, if you look at non- delinquent customers, I

Speaker A

said to you, you know, past delinquencies. Okay, that's really informative. Yeah, but okay, what if you have customers that have either never been delinquent or at least they were delinquent really, really long time ago?

Speaker A

Okay. So, you know, are all of these customers identical? No, they're not identical. They also are different. But here we have to here it's much harder.

Speaker A

Here we have to often look at a lot of indirect information. We have to look at things like utilizations. We can also look at what did we originate the customers at? So, at what risk level, at what PD level did we originate them?

Speaker A

Because when we brought the customer into the bank, we typically made a PD assessment there as well. And you can use that PD as a starting point for the PD to use when they're within your bank.

Speaker A

Because after all, you know, day one, the starting PD that you had when you originate the or the the origination PD should be the first 9 PD if they have the same definitions when you start off the system.

Speaker A

So and and this consistency mean that you know it might make sense to actually have different models for you know currently delinquent, past delinquent, never delinquent customers. I'm not saying that makes sense in every bank.

Speaker A

It depends on how much data you have. But I think it's really valuable to distinguish between, you know, what variables should I use and are these variables relevant for all groups? And if they're not val, you know, valuable

Speaker A

for all groups, maybe these groups needs to be dealt with separately. And again, I think it really makes sense to look at how you deal with it in the business.

Speaker A

Another thing that is really valuable to think about is the combination of the question of attrition.

Speaker A

So especially if we talk fintexs and these sort of businesses what you tend to see is that um that that for a for a fintech typically customers leave pretty quickly. You know you might have a quick high turnover of customers. Same thing

Speaker A

can you have for certain products in in an ordinary bank. So default is actually not the only way that the customer can leave you. They can actually leave you by basically closing down their position. Now in that sense from a risk

Speaker A

point of view that's pretty nice. You know sure that's not good from a business point of view but at least that lowers the losses for us. But attrition is important to bring into calculation because if you think about it attrition

Speaker A

lowers the PB because if the customer closes on you they will not default on you. And sure you can say no but I don't need to make a explicit attrition model because if I just take the PD history

Speaker A

and I just look at where there are loss and where they're not a loss and if the customer attract on me I will not see a loss number. So you know if then I don't need to bring it in. Yeah but there is a

Speaker A

real value in including attrition explicitly for for sort of two reasons. One reason is that it's important for you to understand is the PD low because the customers are really lowrisk customers or is the PD low because the

Speaker A

customer keeps leaving you before you see the problem. And I can tell you an example. I I used to work with lending in Norway and in Norway for a long time they had a very very aggressive loan market. So what happened was that the

Speaker A

customer would be with one bank and they would take a 15-year loan and then they would stay 12 months, sometimes even shorter, and then they would take a loan in another bank. So they would move the loan. So they would move the loan back

Speaker A

and forth, back and forth, back and forth, and eventually they they might default. But if they move it back and forth, back and forth, back and forth, you know, if you and most of your customers keep doing that, most of your

Speaker A

customers would look like they never default on you. So actually the customer looked pretty safe and and it was you know you gave them pretty good prices because you know I will take very little losses because you know it's a little

Speaker A

bit like a hot potato. I will always be able to you know it will leave my balance sheet before the bomb explodes.

Speaker A

I will just have them for a little bit of time and they will pay me some good interest. Yeah. But then in Norway they passed a law that made it much harder to move between banks. Basically the banks

Speaker A

were required to give the customer a equal or better offer than they had at another bank at the the previous bank in order to be allowed to move. So suddenly the customer froze in place. They they did not move anymore and suddenly a lot

Speaker A

of the banks realized that now these were high-risk credits and this the default rates just skyrocketed uh like like you would not believe it.

Speaker A

And had more of these banks separated attrition and default from each other. Then what you could have done is to said look I know that we're about to get into points where attrition is going to go down. Let me crank down the attrition

Speaker A

and see oh these are actually pretty risky customers if I look at just the raw PD component. So separating attrition and default is actually pretty important. And there's also a third factor I think is quite valuable to to

Speaker A

separate as well and that is if you have a customer that go in that u that that dies for real. So we you know we also talk about you know life death processes when you talk about default but you can

Speaker A

really have of course always tragic when customers you know pass away due to high age or or something like that and these customers actually you know the the estate then takes over the depth and that behaves actually quite differently.

Speaker A

So you you then have you know often in your systems and it's really good this is I would say if you work out in a bank a good exercise is go and check how is this record in your system is this

Speaker A

recorded like for example a default because if it is it might really obscure your statistics especially on low risk portfolios where you might have quite few normal credit defaults but you know you will always have one or two

Speaker A

customers that keep dying on you um because these are of course driven by very different factors. I mean the death is not driven by credit risk. So it's also good to separate it out from that perspective. These estates are also more

Speaker A

complex because at the end of the day you can ask you know is this an attrition or is it the default and it can almost be a little bit of in between. If you operate in a more lowrisk industry I would say you know if

Speaker A

you operate like mortgage industry typically uh you know a death it behaves much more like attrition event. It's basically just the customer terminating early on you. But if you operate in a pretty risky indust risky industry, it can be much more like a default event.

Speaker A

But the reason is that if you have, for example, an unsecure loan, you know, the estate might not have too many many assets and the estate might be un unable to to pay you. So you know, again, it's

Speaker A

important to check, you know, how material is it and what typically happens. Does it make sense to put it in the pile of the defaults? make it sense to put it in pile of attrition or should you rather you know it treat it as a

Speaker A

separate category also you know it was already mentioned here before you know this concept about forward-looking information I have a whole presentation on its own that talks about you know doing scenarios um and and you know you can talk a lot about

Speaker A

that I think this is a really really fascinating topic I sometimes also think this is a quite misunderstood topics when it comes to forward-looking nest Because forwardlooking, you know, often people spend a really really a lot of time and say, I need to build really

Speaker A

advanced macro models and so on. I would say the most important thing when it comes to forward-lookingness is to avoid being too backwardlooking.

Speaker A

So if you think about for example, let's say that you build your models and you take you know some data set and you are really proud of yourself. You spend a lot of time, you built this very nice

Speaker A

model, used five years of data and you put the model in production and you let it run and it runs year in and year out and you're, you know, you're very happy about it and you understand it. Sure.

Speaker A

But, you know, if you just let the model run forever and you never sort of recalibrate it or change it or anything, the model might drift away from reality.

Speaker A

And this is a lot what what forward-lookingness is is about. It's basically about that you cannot be overly reliant on backward data but you really have to think about you know what what is going to happen in the future

Speaker A

and a lot of the future you can learn from just looking at it right now but of course what you also would want to reflect is let's say that you've had data that you've collected in a lowrisk period what you would like to do is that

Speaker A

you would like to say okay if I think actually there's going to be a high-risk period incoming that should probably be reflected already Now in my losses and I go back to the original example I gave on the bond trader that you know

Speaker A

you had your portfolio bonds that were traded in the market. I mean if you know when uh if if the world goes into a worse state you know the value of these bonds are going to go down right away

Speaker A

and you are supposed to reflect the same thing in your ETHL model. This is why you need the forward-looking information.

Speaker A

Um and you know you also another thing I can say is that imagine that you run a pretty lowrisk um credit market. So you you give out for example mortgages and and you know there's not so much risk in

Speaker A

these mortgages. You're pretty low you know LTVs. They are very good customers. So then you might say no but I'm not going to take a loss. My LTV is for example 40%. So the loan to value is 40%. you know, I mean, even if the house

Speaker A

is in worse quality and even if the property market goes down, I mean, I Yeah, I'm not going to lose uh lose that. Yeah, but that's why we need scenarios because you you then basically need to include at least one scenario

Speaker A

that is so bad that you actually take some form of losses. So, scenarios and using multiple scenarios and waiting together also matter because you need to properly reflect the full probability distribution. And this is why we use uh

Speaker A

the scenario analysis. It is to have wide enough scenarios so that we see the full losses but also the full upside of the portfolio. And as I said, this is a whole talk on on its own. If someone is

Speaker A

is more interested in learning about this, you know, feel free to, you know, ping me on LinkedIn and similar happy to send you, as I said, have a whole presentation on just this topic alone and how you can do weight calculations

Speaker A

and and so on. So, so it's a whole chapter in itself. But if we now take a little bit of a step back, you know, okay, what have we been able to do here?

Speaker A

Uh, we have this, you know, competing risk model where we look at, you know, the default risk, the attrition risk and this sort of, you know, death death of natural causes. And what we also do typically is that we then, you know,

Speaker A

look at, you know, the the probability part comes from taking the default contribution. and also we include some of these deaths that lead to actual uh defaults and the PD then comes from from integrating this profile. Now in

Speaker A

practice you tend to have that you know you do deal with things on a monthly levels you approximate this integral with a sum instead but this is I would argue a pretty nice framework to operate in. There are of course other frameworks

Speaker A

you can use for example transition matrices and so on. H but I would argue that they are often a little bit they they they make too strong of theoretical assumption. So I myself is pretty fond of these um survival models and I would

Speaker A

also argue that it often mirrors a lot if you go to your business analytics people you can often see that they draw these sort of you know cumulative vintage curves and so on. So it tends to mirror quite a lot of what the business

Speaker A

is actually already doing. And the nice part with this model is that we will get both the overall PD and we will also get the timing of the default. We will get an estimate on where are the defaults

Speaker A

taking place just as we needed. So we move on to exposure at default. And the first part when it comes to loans that's pretty straightforward. Um you know what you typically do is that you tend to make what we call the sudden default

Speaker A

approximation. So you basically assume that the customer keeps paying you and then three months before the default they just stop paying you and that's really what you do your calculation. So when you when you forecast your EAD you say okay you know

Speaker A

remember you always condition on the time of default because this is what you got previously if we look you know sort of you know back in the previous slide you know you get this you know monthly default probability. So when you do the

Speaker A

EAD, you condition on and say, "Okay, I now want to calculate an EID based on that they will default on me month 12 from now." Okay, if the customer is currently up to date on you, that means that they need to stop paying you month

Speaker A

9 because then they have month 9, 10, 11, and then the 12 month they uh they fall over or a little bit how you set up your your counting. But but then it's quite easy for you. you calculate you

Speaker A

know the the the healthy months the customer pays you and then once they stop paying you you know they you know you instead just acrue interest on it and then you discount everything back.

Speaker A

So not really that that come now some people have sometimes say no no but what if the customer had been del you know they go delinquent somewhere in the middle period and then they they go they are delinquent and then they start

Speaker A

paying you back and and so on. Sure, I mean that can absolutely happen, but it turns out in practice that it isn't really so much of a of a problem.

Speaker A

Another thing sometimes people bring up is what about over amortization? So, we're not talking about attrition here.

Speaker A

We talked about attrition previously. We're talking about customers that, you know, they're still on your books, but they just paid you back faster than you had expected. They they throw in a little bit of extra money. And of

Speaker A

course, that's going to lower your exposure. And if you just do this normal contractual uh calculation, you're not going to reflect that. Is that a problem? In practice, in my experience, you know, happy if someone has some different different views, but in in my

Speaker A

experience, often this is not so much of an issue because risky customers, the ones who default on you, tends to be the ones less least likely to make these overpayments. So if you think about the ECL, you know, the ECL is very heavily

Speaker A

weighted toward bad customers because it's basically an estimate of, you know, bad depths and and most of the bad customers tends to be the one that are least likely to over advertise. So sure, there can absolutely be a part of your

Speaker A

book that overtise and it might be quite common, but it would be much more over represented on good customers which get a very low weight in your calculation because they have a low PD. So in practice I'm not necessarily super

Speaker A

concerned. Now we come to the question that was asked you know previously which is about credit facility and I love that that question came up because credit facilities is really tricky how to uh how to deal with because credit

Speaker A

facilities have a current balance and you know you know an on an onbalance part and they have an offbalance part and first and foremost a really interesting thing about a credit card is that you know because I would use credit

Speaker A

card as the as the main example but you have company revolving facilities and and things like that but effectively a lot of them works the works the same way. Now you know I I mentioned before that you know and that when you do I9

Speaker A

calculation you should base it on when you talk about lifetime you should base it on the contractual lifetime of the instrument.

Speaker A

Okay. But if you base it on the contractual lifetime of the instrument what is the contractual lifetime of a credit card? And I and I recommend to everyone you know go into the term if you work for a bank go or you know you

Speaker A

are a customer with a bank go into the terms of services and really dig into what it says about your card. In most cases, you will find that the company have reserved language that says that they reserve the right to cut the card

Speaker A

and to remove any offbalance exposure at at will with you know uh if they if they perceive that there is some issue or or something like that. So from a purely contractual point of view most banks have the right to instantly terminate

Speaker A

the card. So the contractual lifetime is you know what is it like one day or something like that. So it would mean that if we just followed the book by the letter we should really you know have a

Speaker A

really short horizon on this offbalance component and it wouldn't really be a be a problem. I say noticed this because it was discussed in the industry at the time and they said look that that's not really a faithful representation because

Speaker A

most banks would not really cut you know you know functioning cards. They would typically only do that once you're really distressed and then you're really taking on the exposure.

Speaker A

So what they introduced was a special exemption in the rules for revolving facilities where they say for a revolving facility you actually don't have any contractual time. You instead more or less have to to make your own assessment here. And what you typically

Speaker A

do is I mean one way to view it is that effectively view the credit card as perpetual and or at least whatever you know maturity time that you might have in your in uh in your terms of service.

Speaker A

But it's quite common that there's actually no no real ending time because you know the company can terminate it at will. You know some I've seen one bank who said oh but the the lifetime I use in the first 9 is based on the lifetime

Speaker A

on the physical plastic card. Yeah, but you will send a replacement card once that time rounds out. So that that's not going to fly. But what you can do is that you can basically determine sort of, you know, how long, you know, over

Speaker A

how long time do the company do the customer typically stay with you and over how long time into the future are losses typically taking place and you make a sort of, you know, loss emergence curve and that's really what you what

Speaker A

you base the time on. And you typically here attrition will for example matter. So it's good to this is also a good example why it's good to do attrition modeling because you can then use attrition to determine how long the

Speaker A

customer stays. But then you might now ask okay how to include this offbalance part and a very common model is to have that you say that the exposure default is equal to the current balance plus a portion of

Speaker A

this offbalance. You know the most extreme assumption would be you know on one hand you can say okay the customer you know they have a credit card of 100,000 they've used 40,000 um they have 60,000 remaining okay you know the

Speaker A

balance is 40,000 I will just assume that they use no more money if they default okay that that's pretty aggressive that's probably not such a good approximation but on the other hand you know you could also say okay the

Speaker A

customer have 40,000 they have 60,000 remaining I will always assume they use everything in the in the run up to default they will always spend on every single single you know dollar remaining on the card. So I always use a h 100,000

Speaker A

that's also overconervative and I first tells us to be unbiased. So okay then a model that's common from to see from the IRB is that you say okay I will do something in between I will say you know

Speaker A

some portion let's say that this portion is 50%. So the CCF I set that to 50%.

Speaker A

What I will do is I will set that CCF to the to I will then combine it. So I will take 40,000 plus 50% of 60,000 that is you know 30,000. So I will use 70,000 then and this is the CCF and this is

Speaker A

quite common to see that you know you you estimate these proportions but from an IRB point of view this really makes sense um because if you think about it from an IRB point of view you only already try to be too

Speaker A

conservative. So you say okay um my what what my exposure will be is my current balance plus I take some portion of the offbalance but from an IFR spying point of view you try to be unbiased and this

Speaker A

model will run you into trouble if you just have the onbalance plus CCF times the offbalance because if you think about this model the EA is always equal to or larger than the onbalance but remember what we talked in the previous

Speaker A

section we talked about that the E can actually be smaller smaller than the current balance because the customer will pay you back. That's not going to be reflected here. So, you know, you what if and I I used to work for a

Speaker A

credit card company, so I experienced this quite a lot. It's actually not so uncommon that customers will default on a smaller amount than they currently have because sometimes you see that you have customers that have drawn down pretty almost everything and then they

Speaker A

pay you back quite a lot and then they start drawing back again and then they default. But if you have this I mean effectively the customer have gone down in balance that's not going to be able to be handled if you just have the CCF.

Speaker A

Therefore I argue that you need to add an additional parameter the CRF the credit retention factor to stabilize the regression. So what you really can do is that when you calculate these you can make a regression between sort of you

Speaker A

know between your actual EAD your onbalance and your offbalance exposure and then so you you have both of these as variables and then you estimate these regression coefficients. So instead of having one coefficient the CCF you really need to

Speaker A

have two coefficients. So this is not not so often so seen but I would argue it's not seen so often because people too often don't give enough attention to the fact that if 9 and IRB are quite different from each other to the nature

Speaker A

okay we come now to the final parameter the loss given default so you know we are now in uh in default okay what is next in the estimation and you know then we one of the things that's also important to focus on is

Speaker A

actually that the EAB might be somewhat different from the customer balance. I think a really good example is if you buy distress debt as I as I mentioned before or or similar that you might have originated loans at discounts or

Speaker A

premiums uh because let's say that you have distress depth that you bought you bought the the the depth at 10%.

Speaker A

Okay. um to to nominal value. I mean the EAD that's not going to be a 100. That's going to be what you bought the loan for. The EAD is is going to be set based on what your balance was to start with.

Speaker A

And the balance is or or based on what your purchase price were to start with.

Speaker A

So you have to be a little bit tricky when you when you originate loans or when you get loans that that do not sort of behave sort of normally and might have these discounts and premiums because it should really be the original

Speaker A

fair value that you start the EAD calculations from. And similar I mean another favorite example of mine is that if you have uh if you have these brokers loan brokers I mentioned this before um and I show actually one here which have a uh very

Speaker A

Swedish one which is called lendo which is pretty dominant when it comes to unsecure lending in the Swedish market.

Speaker A

So you know they have you know they have one bank and they target one other bank and they have 35 more. So when you that is do the Swedish one h you go to this lendo they actually send your offer to

Speaker A

multiple ones. Now lendo will charge around 5% for this if you get a loan through them. So if I lend out a H 100,000 to a customer through Lendo, actually it costs me 105,000.

Speaker A

And these 5,000 I'm not allowed to just write them off as an immediate cost because really I was prepared to pay 105 to get the right to the customer's cash flows. So I really need to write the loan balance as 105. But if the customer

Speaker A

default on me, I only have the ability to go after the customer's 100. I mean, I cannot come to the customer and said, "Look, I mean, because you default on me, I need you to pay the fee me also

Speaker A

the fees I paid to this broker." That's not going to fly. So, you need to be a little bit careful when you do LGD calculation because often in LGD, it makes sense to look at LGD as a proportion of balance. But if you then

Speaker A

have the EAD and the balance is different, you actually have to adjust for that. Another thing that you have to be slightly careful about is that you have to distinguish between a PD and LGD. So I used to work something for a

Speaker A

previous bank with the leasing and then it was quite common that when we did leasing agreements we did them with quite large companies and then the some of the risk managers they would say oh I mean the LGD on this agreement oh it's

Speaker A

very very low the LGD is very low because this is a very good customer.

Speaker A

It's a very good customer and they often and they have a lot of collateral and they have a lot of assets. Sure, they're not tied to us specifically, but you know, if they ever were way to go default, you know, oh, they would have

Speaker A

so much assets that we could settle off, we're always going to get their money back. And I was supposed to say, yeah, but okay, if they have a ton of assets and they're really liquid and they really have a nice business model while

Speaker A

working, I mean, they're not going to go default on you. They're only going to go default on you if they're really bad.

Speaker A

So, they're only going to go default if things have turned bad. And you have to imagine what that world looked like.

Speaker A

It's a bit like the goat example we gave before. If the you know you say oh but the farmer will always feed his goats.

Speaker A

Yeah but that's if the farmer have money and if the farmer have money it's also going to pay you and it's not going to default on you. But what if the farmer has no money to pay you then he might

Speaker A

not also have had no money to to uh to pay the goats and hence the collateral might be of worse quality.

Speaker A

Another thing before you even go to to collateral, one part important part to uh to consider is really to consider the cure rates because most of your customers that go default, you might actually be able to get back from the

Speaker A

default position uh and to and to be able to recover quite a lot from them h or or to to get them back into performing loan. And if you get them to a performing loan, I mean effectively the you know I mean if they stay

Speaker A

performing forever I mean things are are then pretty good from you from uh from that perspective. Um so you actually have to include this when you do the LUD. Um and one quite common model is that you look at this cure rate. So you

Speaker A

will estimate that and if the customer do not cure back I mean then you will have what you sometimes call loss given loss or loss given sale because basically you then you will have to go on and try to sell their collateral and

Speaker A

so on but but if they but even if they cure it's important to remember that you still have a loss exposure because if you have a customer that go into default and the customer cure and then they are

Speaker A

out you The customer doesn't have zero risk. I mean in fact this customer typically have higher risk than they had to start with. They can default on you again and you can have multiple defaults and you can get a risk again and you have to

Speaker A

typically compensate for for this. Now it depends often how risky the how risky the portfolio is if you have to do these corrections. But this is why I've included this loss given cure LGC. So it's good to and what you can do is a

Speaker A

way to estimate this is that you look in your data and you look at customers that just cured and you look at what is their ECL divided by their balance that's a pretty good proxy on you know how large

Speaker A

is this effect and then you know okay so we have talked about you know we've talked about the LGD concept we talked about we've talked the PD con we talked about the LGD concept we come to you know the final

Speaker A

part which is the staging Um and remember that you know I talked about that you know had this first idea that would they would do this discounted cash flow calculation. Um that you know they they were were ruled out on that.

Speaker A

Uh but a nice part of uh but but it would have been nice for them to to keep that method and we can see now why and one important thing in accounting is the matching principle. So the matching

Speaker A

principle is that expenses should be reported at the same time as revenues are. So there there should be this this lineup between these two and if you think about an incured loss model. So incured loss model is basically as we

Speaker A

said that's when you get when you take you know you take no losses before you they go into default that violates the matching principle because as long as the customer keeps paying you you take in income and then when they stop paying

Speaker A

you okay you take all the losses. So you really did not match losses uh the expenses with income. If you take this credit adjusted effective interest rate method, this discounted cash flow that really works here because remember that

Speaker A

you know you adjusted the interest rate the discount factor you adjusted that for the credit losses. So you actually acrew less interest over time and you build up sort of a buffer and basically that buffer is then sort of released if

Speaker A

they then were to were to go into default. So in this case you fulfill the matching principle. If you take the expected credit loss model and we look at lifetime losses now that violates the matching principle but now it's wrong

Speaker A

way the other way because you actually take all of the losses first or first because you take a lifetime loss and then you take all of the income afterwards. So again you are wrong now on the matching principle but you're

Speaker A

wrong the opposite direction. I saw that there was someone here from from US and you guys work with the CSL. The CSL standard use you know it has no stage one and stage two. It does everything on lifetime and I and the plan was

Speaker A

originally for uh FASP and ISAB the two you know accounting agencies to really align on a common standard but US decided to go their own way and said look we're going to go lifetime for everything. ISAB said no we really care

Speaker A

about this matching principle so we have to come up with some form of own system because they argue that you know you get two large day one losses if you take everything up front and they were then you could even have you know you can

Speaker A

even get really weird situations where like a company would have like negative equity because basically the the assets are just too lowly valued.

Speaker A

So ISB tried to create a compromise and compromise you often said you know nobody will be happy and their solution was to say okay look to reduce this day one effect. What you will do is that you will for if the loan just stay on the

Speaker A

same risk level as you took in the loan effectively you know day one you know the the the day one effect should really be zero but we will proxy it by using only 12 months of losses. So you will

Speaker A

look at the the PD12 so the probability of default default within 12 months and then you will look at all the losses connected to that instead that's called stage one and the argument for that you can do this exemption is that it's a

Speaker A

matter of risk pricing if I know that the loan is pretty risky I'm also going to charge a much higher interest rate.

Speaker A

So as long as the loan just stay on the same risk level as I thought it would be, I don't really have an real economic loss and then I would just gradually acrue the the risk over uh the the

Speaker A

losses over time. But on the other hand, if it turns out that you know I'm completely mispriced the risk then I need to make a quite big sort of adjustment up front to really recognize this. And you can compare it to for

Speaker A

those of you that know is IS 37 the standard there you have for example if you are a building company and you build let's say you build a house and you know you have some expenses but when you eventually sell the house in the end

Speaker A

you will get some income what you're allowed to do is that you are then allowed to recognize you know income over time basically because you you are basically constructing an asset on your balance sheet so you recognize income

Speaker A

and you match it over to expense. But imagine that you now realize that, oh no, this house is going to be very expensive. It's going to be much more expensive than it than the income I will get finally when I sell it. So I'm

Speaker A

actually in a lossmaking contract or honorous contract that we sometimes call it. Then you really have to go and make an upfront provision for this loss right away. And it's pretty much this logic that we have in stage one and stage two.

Speaker A

And it so it's important to remember that the point of stage two is not to be a waiting room for stage three. It's not to identify loans that will go to stage three. It is to identify places where

Speaker A

you have mispriced your risk. Um the problem is however that this concept is incredibly vague. We have sort of two outer boundaries. So if the co if the if the account is 30 days delinquent then we put them in stage two

Speaker A

because that's a sort of safety wall to not have them go all the way to 90 days uh delinquent. And on the other hand if the company or or the lender or the borrower is on um is treated as

Speaker A

investment grade then we don't need to do uh any any staging. We can just place them in stage one because that's called the low credit risk exemption. You have to calibrate that because you might say, "Okay, most of my counterparties are not

Speaker A

rated." But then you're allowed to determine an internal rating that is aligned with um uh with the investment grade. Uh but the problem is that there's still a lot of of sort of diversity between you know between these

Speaker A

two boundaries and those boundaries are pretty set but where to do in between it's very very unclear and I have a whole separate presentation again very happy to send if someone interested where I talk about you know what can you

Speaker A

do and and how to think about it but effectively already you know today I had you know I was in meeting earlier today and and u you know where where we were discussing this topic and effectively it is really hard what to do about staging

Speaker A

and I would say the industry as a whole really struggle with it. Um Jacob just uh you have about seven minutes.

Speaker A

Yeah, I was about to say we basically finished this one. So we we are more or less now in I I know that we are about to run out of time. So basically you know as we said we have now gone through

Speaker A

the standard talk about it how it really really developed and look about the parameters because now I really want us to get to the to the question side.

Speaker A

Yes. Uh you just let me know. Have you you still have a a slide to present?

Speaker A

Yeah. Yeah. Yeah. No, basically this just sort of summarizes what I've talked about previously. So So uh I I think we are a pretty good point to move on to questions.

Speaker A

Yeah. Um there are a number of questions and I have uh given in the chats whoever is in I've given a link to the post uh of this webinar so that if you can post your questions right there Jacob will be

Speaker A

able to see them and answer them right there because time is not really on our side. I'll just highlight a few uh questions. Uh there's one question from Rajab who says how do we account for customers who significantly increase the

Speaker A

UD uh their credit utilization shortly before default. Yeah and and and I would touch very good question I would say that's basically with the with the CCF model. So the CCF model is that if you have a history that

Speaker A

customer tend to do this okay then the CCF should be pretty high. So so this is really covered through the CCF model.

Speaker A

Sorry, can can you hear me? Yes, we can. Yes. Apologies. I So, but but yes, that is with this the CCF model. So, let's see if I can also get the questions uh and figure out. Perfect. Um I think you

Speaker A

should uh please leave your emails somewhere in the chat so that the materials can be sent to you after Jacob shares them uh with us. And may I Jacob please listen for a few questions. Jri you have a question a minute please and

Speaker A

make it short and clear. Jri your Yes. Jri, hi Dar, please copy the emails that are being sent in the chat for the sending of the materials. Uh Jes, there's another hand. Um yeah yeah yeah I saw I saw the question

Speaker A

at one uh place mentioned that for particular corative customers is default in accounts get cured in such case we should count as default and and and this is a really really good question yes you should and this is very important what

Speaker A

when you calculate the PD model you calculate time until your first default event so even if the customer cures back you still have to treat it as a default and then the fact that they cure that is being taken care of by your cure model

Speaker A

that's part of your LGD. So from from a from a sicker point of view where you typically looks at the PD ratio. Okay.

Speaker A

There you you still need to consider the fact that they did default on you. So so you really have to make this uh this split between them. Yes.

Speaker A

Thank you. Uh is there anybody else with a question? We remained with very few minutes. Perfect. Mago please.

Speaker A

Okay, my my question is what what does a typical day look like for someone who is into credit risk modeling looking at IRB models and if 9 models they went really really good so I would say it depends a little bit where you

Speaker A

are in the in the month so typically you know imagine that you sit on like an IFRS9 desk for example typically when you run the time where you run the models that is you know the beginning of the month that's very you know stressful

Speaker A

times you get all the data in, you try to get the numbers out. That's much more of a little bit of a machinery. Okay, all the data comes in, you use the machine, you're trying to get the the

Speaker A

numbers out. Um, and you do some analysis, why is it changing? What is different this month? Then you a little bit switch over to, you know, sort of the in between period where you, you know, during the month, so maybe first

Speaker A

week of the month that's, you know, getting the numbers out. The other three weeks of the month, what you then try to do is you try to maybe build the models better. You try to improve upon the models. um you look at you know

Speaker A

different different analysis and changes and try again to you know connect the model better to to uh to the business model. So what I would say is you know a lot of what it's about to work in quantitive credit risk is really to

Speaker A

understand credit risk. You spend quite a lot of time in really trying to understand the products that you're working with. I have been very fortunate that through my bank that they've been allowed me to go out into the branches

Speaker A

and really you know get to visit you know branch managers and talk to them and sit down with people and I really this is my recommendation to you guys like you know if you want to understand credit risk you really have to

Speaker A

understand the credit process. So really spend time with the people in your organization that manages the credit.

Speaker A

Sit on the with the people that you know manage the phones where customer calls in and for example when you have collection department that calls customer listen into the calls try to understand what happened. You hear the stories from the customer you hear the

Speaker A

issues that they having. Okay that gives you an indication of what variables and what things are are relevant.

Speaker A

Okay. Thank you so much David. Please one sorry navidid one minute please n David Barati hello yes uh this is me I I wanted to ask a question about uh so after a period of time you will uh start to

Speaker A

review the model that you have uh uh created before or uh doing some portfolio review in those cases in those uh situations what would you focus first?

Speaker A

Um really really good um question. So I would say let's let's say we start with for example a PD model like the absolute important thing first is that you determine if the calibration is correct.

Speaker A

So what you have what you then determine is let's say that it's for example a 12-month horizon you're operating with.

Speaker A

Okay then you basically would check you know okay you can make a time series you check what is the model projections 12 month ago what was the reality of the 12 months and you start to check how they

Speaker A

works. Um it's it's important to say this is a big difference from how we do in for example IRB or in credit origination where we will focus often quite a lot on if the if the risk differentiation is the correct thing you

Speaker A

know the genie so genie is not unimportant in IFS9 but it is uh it is much less important than the calibration and the way you can understand it like this imagine that you have perfect risk selectivity so you know exactly in what

Speaker A

order the customers are in risk. You can place the customers from lowest risk to highest risk but the problem is that your estimates of the risk is 10 times wrong. I mean then the ECL number will be 10 times wrong and it will be quite

Speaker A

disastrous. While on the other hand if you have you don't really know who the risk customer is but you have an overall understanding of what the overall PD level is you're still going to get a pretty decent ECL. This is therefore the

Speaker A

calibration level is the first thing that you need to check when you determine this and then secondary you start looking at the risk differentiation properties like the genie and similar but but very good question and and there's a lot of art I

Speaker A

would also say you know if I were to give one advice I used to work as a validator one advice when you do validation is don't be too stuck on test yes testing is very important but really try to think about you know is there

Speaker A

something in this model that looks off you know I run some simple calibration and look is there something that you know stand out to me and they say you know this variable choice seems weird this seems problematic okay dig into

Speaker A

that try to be you know the way that you should think as a validator is a little bit like a stress tester you go out you kick the tires you try to understand is there some issue here rather than being

Speaker A

too mechanical thank you we are really out of time see that can you make it very short uh so that Jacob can answer and we close this session see that Hi.

Speaker A

Yes. Yeah. Hi. So, uh Jacob, I'm trying to uh uh isolate the dollar impact of PD, LGD, EAD, any possible granular components of LGD on ECL. Uh do you have any recommendation on that?

Speaker A

Yes, really super good question and and I worked at the previous bank. I was part in building a system for this. So, in in because you are right like when you do a validation, this is really what you want because you want to be able to

Speaker A

tell management how much is the impact of each of them. uh here will come a very very speed version. Feel free to follow up on me on you want more details on it. But what you do is that you first

Speaker A

calculate your your modeled u ECL with everything just modeled. What you then do is that you replace the PD component with an empirical default rate. So you take PD and you substitute in the de the the the the you substitute in if the

Speaker A

customer defaulted or not when you follow it up. So you basically for everyone who defaulted in a period you calculate still the theoretical EAD in LGD but you now just include the defaulters in your sample. So you go

Speaker A

from sort of using a PD to just using the indicator function the default or not and you that first difference between these two that shows you the the sort of dollar amount because how correct the PD is. Now you do the second

Speaker A

step which is that you replace the ED with the actual default balance. So the difference between default rate and EAD versus default rate and observed balance. Okay, that gives you now the EAD uh effect in dollar amount. And

Speaker A

finally you compare you know the right of amount compare and and the right of losses versus the LGD and you substitute in that difference and you get sort of the LGD component side. So as I said this you basically take one parameter at

Speaker A

a time and you replace it with the observed values. This is really and and you look at the difference in the ECL value when you replace it with an observed difference. This is how you break down the dollar amount.

Speaker A

Thank you so much Jacob. As you can see there's uh it's still hot. The discussion is still very much hot and I concur with the the rest that in the chat that it has been a very very productive and very very nice uh

Speaker A

presentation. If you really enjoyed this, please feel free to show some love to Jacob by Yeah.

Speaker A

And I can say as well, you know, if someone have no more questions, we were not able to come here, feel very much free to reach out to me on on LinkedIn, similar. Always happy to talk about if 9

Speaker A

and I should express my my sincere gratitude for for being invited. It was a real pleasure to be here today.

Speaker A

Yes, I have posted Jacob's uh link in. You can uh go and interact with him.

Speaker A

I've also posted the association of the students. We can reach the students association as well as I myself so that we can meet. We have seen your email addresses and after Jacob sends over the materials we'll be able to send you the

Speaker A

materials for the revision. Jacob on behalf of the University of Dislam department of mathematics and the students thank you so much for accepting our invitation and thank you everyone in the call that contributed. The chat was very much alive in the middle of the uh

Speaker A

presentation and we kept on exchanging ideas and informations about the afs of the world. That's how you know that this is a timely and very much needed uh information around the world. Uh from Tanzania, good evening from everybody

Speaker A

and hope to keep on in touch with you guys. Thank you so much Jacob.

Speaker A

Thank you. Bye. Thank you everyone. Okay, goodbye everybody.

Topics:IFRS 9credit risk modellingfinancial regulationexpected credit lossIAS 39banking standardsfinancial crisisloss provisioningwebinarfinancial sector

Frequently Asked Questions

What is the main difference between IFRS 9 and the previous IAS 39 standard?

IFRS 9 requires banks to use an expected credit loss model that recognizes losses earlier and more forward-looking, whereas IAS 39 used an incurred loss model recognizing losses only after a default event.

Why was IAS 39 considered problematic in credit risk provisioning?

IAS 39 allowed forward-looking provisions but was often misused by banks to manipulate profits, leading to insufficient loss recognition during good times and sudden large losses during crises.

How does IFRS 9 help minority investors?

IFRS 9 promotes transparency by requiring objective evidence for credit losses, helping minority investors better understand the true credit risk and financial health of banks.

Get More with the Söz AI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries, speaker detection, and unlimited transcriptions.

App Store Google Play

Or transcribe another YouTube video here →