The ECL Laboratory Expected Credit Loss Simulator
Designed by Md Nazmus Sakib, CFA · Made with Claude
Start Here — The Story of a Loan
No finance background needed. We'll build up — slowly — to what Expected Credit Loss actually means. By the end of this tab, the rest of the simulator will make sense.
Part 1 — What is a loan, really?
  1. You need money today — maybe to buy a car, start a shop, or build a house.
  2. A bank gives you that money — let's say $10,000. This is called the principal.
  3. In exchange, you promise to pay it back over time — say, $200 every month for 5 years.
  4. Notice: $200 × 60 months = $12,000. You borrowed $10,000 but pay back $12,000. That extra $2,000 is the bank's interest — their fee for letting you use their money.
Think of it like renting an apartment.
The principal ($10,000) is like the apartment itself — you get to use it. The interest ($2,000) is like the rent you pay for using it. When you finish paying back the principal, the "rent" stops. The bank is essentially renting out money.
A loan in pictures
TODAY $10,000 Bank gives you 5 YEARS · 60 MONTHLY PAYMENTS you pay $200 every month AFTER 5 YEARS $12,000 You've paid back The extra $2,000 is the bank's interest — their fee for lending.
Part 2 — Three things can happen
  1. The good case (95% of the time): You pay every month, on time, until the loan is finished. The bank gets all $12,000. Everyone is happy.
  2. The shaky case: You start missing payments. Maybe one month, maybe two. The bank gets nervous — will you recover, or get worse?
  3. The bad case (we call this "default"): You stop paying entirely. Maybe you lost your job, your business failed, or something serious happened. The bank now has a problem: they expected $12,000 back, but they might never see most of it.
Three possible futures for any loan
✓
Performing
Borrower pays every month. Bank gets full $12,000 over 5 years. Most common.
~95% likely
!
Watch-list
Some missed payments. Borrower might recover, might worsen. Bank watches closely.
~4% likely
✗
Default
Borrower stops paying. Bank might recover only part of the loan — through collateral, courts, or settlements.
~1% likely
Think of it like lending $100 to a friend.
Most friends pay back. Some are flaky — they pay sometimes, miss sometimes. A few never pay you back. If you only had one friend, you might just accept whatever happens. But what if you lent money to 10,000 friends? You'd need to estimate, in advance, how much you might lose.
Part 3 — Why banks have to estimate losses in advance
  1. A bank doesn't make one loan — it makes thousands or millions. Even if 99% of borrowers pay back, the 1% who don't can add up to huge losses.
  2. Banks must put aside money — called a provision or loan loss reserve — to cover those expected losses. This is real money set aside on the balance sheet, not just an idea.
  3. Before 2018, banks could wait until a loss actually happened before recognizing it. This let problems hide. The 2008 financial crisis showed how dangerous that was.
  4. The new global rule (called IFRS 9) requires banks to estimate expected losses upfront — for every loan, every quarter, even before any borrower has missed a payment. This expected loss is called ECL.
The whole point of this simulator: show you how a bank answers the question "if we make this loan, how much do we expect to lose on it?" — before any loss actually happens. That estimated loss is called the Expected Credit Loss (ECL), and every bank in the world must calculate it for every loan they hold.
Part 4 — The big idea, in three pieces

Estimating expected loss on a loan needs answers to three simple questions:

  1. How likely is the borrower to stop paying? A small number, like 2% per year. We call this the Probability of Default, or PD.
  2. If they do stop paying, how much do we lose? Usually we recover some — through collateral, lawsuits, or repossession. If we lose 40 cents of every dollar owed, that's a Loss Given Default of 40%, or LGD.
  3. How much do they owe us right now? The outstanding loan balance — what the borrower currently owes. We call this Exposure at Default, or EAD.
The expected loss formula, drawn out
QUESTION 1 PD How likely to default? 2% × QUESTION 2 LGD How much lost if so? 40% × QUESTION 3 EAD How much owed now? $10,000 = ANSWER ECL Expected loss $80 2% chance of default × 40% loss-rate × $10,000 owed = $80 expected loss The bank sets aside $80 on its balance sheet — for this one loan — to cover the loss it expects.
Part 5 — Where it gets complicated (and why this simulator exists)
  1. That $80 number used three rough estimates. Real banks need precise ones. What's the actual probability of default? Based on what? How does it change if the economy worsens?
  2. A loan isn't just one year. It might be 5 or 30 years long. The chance of default in year 1 is different from year 5. So you need PD for every year, then combine them.
  3. Different loans are different. A loan to a healthy borrower has lower PD than a loan to a struggling one. The rules say banks must categorize loans into stages — performing, watch-list, defaulted — and treat each stage differently.
  4. Then there's the economy. If a recession is coming, PDs go up. If it's a boom, PDs go down. Banks must build in macro scenarios — best case, base case, worst case — and probability-weight them.
  5. Finally, banks don't just have one loan — they have thousands. The simulator builds up from one loan to a portfolio, then shows how the components interact.
What's in the rest of this simulator:
  • Primer (I): The official rules and definitions from IFRS 9.
  • Components (II): Play with PD, LGD, EAD individually. See how each affects ECL.
  • Staging (III): How banks decide if a loan is "healthy", "watch-list", or "defaulted".
  • Single Loan (IV): Walk through one full loan, month by month, year by year.
  • Portfolio (V): What happens when you have 250 different loans? The math scales.
  • Macro Overlay (VI): What if a recession hits? Probability-weight three scenarios.
  • Sensitivity (VII): Which input matters most? Test the model's vulnerabilities.
  • Workbench (VIII): The full machine — try advanced statistical methods used by real banks.
  • Glossary (IX): Plain-English definitions of every term in the simulator.
Quick check — before moving on
Q1. A bank makes a $50,000 loan. They estimate a 4% chance the borrower defaults, and if they do, the bank loses 30 cents on every dollar. What's the expected loss?
Answer: ECL = 4% × 30% × $50,000 = $600. The bank sets aside $600 as a provision for this one loan.
Q2. Why doesn't the bank just wait to see if the borrower actually defaults before recognizing any loss?
Answer: Because the 2008 financial crisis showed that "waiting" lets problems hide. The new rule (IFRS 9) makes banks recognize expected losses upfront, so their financial statements show risk honestly and in real-time.
Q3. What are the three questions a bank needs answers to, in order to calculate ECL?
Answer: (1) How likely is the borrower to default? (PD) (2) If they default, how much will we lose? (LGD) (3) How much do they owe us? (EAD).
Ready to see this in action?
Next, watch one real loan unfold from origination to default — and see how ECL changes at each life event.
One Loan's Story — From Origination to Resolution
Watch a single real-world loan unfold through five life events. At each event, see exactly how ECL changes — and why.
i
If you only remember three things from this tab
  • ECL is recalculated every reporting period — it's never a static number.
  • The biggest ECL jump happens when a loan migrates from Stage 1 to Stage 2, because the horizon stretches from 12-month to lifetime.
  • ECL is a forward-looking estimate. A loan that never actually defaults can still carry meaningful ECL on the bank's books for years.

Meet Rashid Trading Co. — a mid-sized importer of industrial equipment. In January 2024, they borrowed $1,000,000 from your bank to finance new warehouse machinery. Five-year term. 9% interest rate. Equal-installment repayment.

Below, we walk through what happens to this loan — and to its ECL number on the bank's books — over the next five years. Each Act is a real moment in the loan's life. The numbers update at each step.

I
Origination
The loan is created. Books look clean.
January 2024 · Day 1

Rashid Trading has a solid balance sheet, three years of profitable operation, and good cash flow coverage. The credit committee approves the loan at a 1.2% annual PD. The loan is classified as Stage 1 — performing, no credit deterioration. ECL uses the 12-month horizon.

The bank takes warehouse machinery and inventory as collateral, giving an effective LGD of 35%. Because EAD at origination equals full principal, the calculation is straightforward.

ECL Posted to Books
ECL = 1.2% × 35% × $1,000,000
$4,200
Starting point
II
Year 1 — Performing
A normal year. No surprises.
January 2025 · One year in

Rashid Trading paid all twelve installments on time. They submitted updated financials — slightly weaker margins due to import cost inflation, but DSCR still healthy at 1.8×. PD remains 1.2%. Stage 1 confirmed.

Through amortization, the outstanding balance has dropped from $1,000,000 to $830,000. EAD shrinks naturally. LGD is unchanged. ECL drops as a consequence — purely because there's less exposure to lose.

Updated ECL
ECL = 1.2% × 35% × $830,000
$3,486
$714 lower than Year 0
III
Year 2 — SICR Trigger
First warning signs. Credit migrates.
January 2026 · Two years in

The local currency depreciates 18%. Rashid's import costs spike. Their latest financials show 30-day overdue receivables, DSCR dropped to 1.1×, and an external credit rating downgrade from BB to B−. The bank's significant increase in credit risk (SICR) trigger fires.

The loan moves to Stage 2. PD jumps to 5.5% annually. Critically — and this is where ECL changes dramatically — the horizon stretches from 12 months to the remaining lifetime of the loan, which is 3 years. Cumulative lifetime PD is roughly 15.6%.

Behind the Number
How does 5.5% annual become 15.6% cumulative?

The 5.5% PD is an annual probability — chance of defaulting this year. The loan has 3 years remaining, and default could happen in any of those years. We need to add up the chances across all three.

But we can't just add 5.5% + 5.5% + 5.5% = 16.5%. That would double-count borrowers who default in year 1 and then "default again" in year 2 — which is impossible. We need to be smarter.

The trick: count survivors, not defaulters.

Instead of asking "what's the chance of defaulting", flip it. Ask: "what's the chance of surviving all 3 years without defaulting?" If the survival probability each year is independent, just multiply them.

Year Annual PD
prob of defaulting THIS year		 Annual Survival
1 − PD		 Cumulative Survival
product so far		 Cumulative PD
1 − survival
Year 1 5.50% 94.50% 0.9450 5.50%
Year 2 5.50% 94.50% 0.9450 × 0.9450 = 0.8930 10.70%
Year 3 5.50% 94.50% 0.8930 × 0.9450 = 0.8439 15.61%
In one line
Cumulative PD = 1 − (1 − annual PD)years
= 1 − (1 − 0.055)3
= 1 − (0.945)3
= 1 − 0.8439
= 0.1561 → 15.61%

Why not just 16.5% (= 3 × 5.5%)? Because every defaulter removes themselves from the pool for the next year. By year 3, only 89.3% of the original cohort is still around to potentially default. So the year-3 default rate, measured against the original 100 borrowers, is only 5.5% × 0.893 = 4.91%, not 5.5%. The naive sum overstates; the survival method correctly compounds.

Picture it: 100 borrowers, watching them thin out year by year
Start
100
surviving
After Year 1
94.5
5.5 defaulted
After Year 2
89.3
10.7 defaulted
After Year 3
84.4
15.6 defaulted
Updated ECL — Stage 2 (Lifetime)
ECL = 15.6% × 35% × ~$640,000
(avg EAD over remaining 3 years)
$34,944
10× increase from Year 1
IV
Year 3 — Default
The borrower stops paying.
June 2027 · Two-and-a-half years in

Rashid Trading misses three consecutive payments. They are over 90 days past due, triggering default classification. The loan moves to Stage 3 — credit-impaired.

ECL is no longer about probability of default — default has occurred. ECL now equals the bank's best estimate of actual loss: EAD at default minus expected recoveries. The collateral (machinery and inventory) is appraised at $420,000; legal and workout costs are estimated at $35,000. The current outstanding balance is $580,000.

ECL — Stage 3 (Impaired)
ECL = $580,000 − ($420,000 − $35,000)
= EAD − Net Recovery
$195,000
~6× higher than Stage 2 estimate
V
Year 4 — Workout & Resolution
Real recovery comes in. The estimate is reconciled.
December 2028 · Workout completed

After 18 months of workout proceedings, the bank successfully liquidates the warehouse collateral for $385,000 (below the original appraisal — collateral markets are illiquid in distress). Workout costs came in higher than estimated at $48,000. Total net recovery: $337,000.

Actual loss: $243,000. The bank had been carrying $195,000 of ECL — so a top-up of $48,000 hits the P&L in this final quarter. The loan exits the books. The ECL journey ends.

Realized Loss · ECL Closed Out
Actual Loss = EAD − Actual Recovery
= $580,000 − $337,000
$243,000
$48,000 P&L top-up vs estimate
The ECL Journey · Visualized
STAGE 1 STAGE 2 STAGE 3 · CLOSED $250k $150k $50k $0 $4.2k $3.5k $35k $195k $243k ACT I ACT II ACT III ACT IV ACT V Origination Year 1 SICR Default Resolution
Real Money Impact
What this $243,000 loss meant for the bank
Total ECL Posted: $243,000
P&L Impact
−$243,000
Net interest income on this loan over 3 years was roughly $135,000. The loan was a net loss for the bank — even though it generated revenue.
CET1 Capital
−$243,000
Provisions reduce retained earnings, which reduces CET1 capital one-for-one. For a bank with $100M CET1, this is a 24-basis-point hit to its capital ratio.
Dividend Capacity
−$243,000
Every dollar of provisions is a dollar less the bank can distribute to shareholders. Scaled across thousands of loans, provisions are why bank dividend policies are so tied to credit conditions.
For perspective: if a bank has 10,000 loans like this and even 1% of them follow Rashid's path, that's $24.3 million in losses per year — enough to reduce annual profits by 5–15% for a mid-sized bank. This is why ECL exists.
?
Knowledge Check · 3 Questions
Q1. Why did ECL jump 10× from Year 1 to Year 2 in Rashid's story?
Q2. In Act V, the bank had to add $48,000 to the P&L on top of the existing $195,000 provision. Why?
Q3. In Act II (Year 1), the loan was performing perfectly. Why did ECL still decrease?
Now You're Ready For The Framework
You've seen one loan's life. Now learn the formal IFRS 9 rules that govern this entire process.
A Brief Doctrine of Expected Credit Loss
From incurred-loss to expected-loss: how IFRS 9 rewrote the grammar of provisioning, and what every risk practitioner needs to internalize.
Reader note If you're new to credit risk, visit Start Here (Tab 0) first — it explains loans, default, and the ECL formula in plain English with no jargon. This tab assumes you already know what those words mean and goes into the formal IFRS 9 framework.
i
If you only remember three things from this tab
  • IFRS 9 replaced incurred-loss with expected-loss accounting — provisions now reflect forward-looking estimates, not just past defaults.
  • Every loan sits in Stage 1, 2, or 3, and the stage determines whether ECL is calculated over 12 months or lifetime.
  • The framework rests on three pillars: PD × LGD × EAD, all discounted at the loan's effective interest rate.
?
Start Here — What Is This About?
2-Minute Read
1
The Problem Banks Need to Solve When a bank lends money, some borrowers will not pay it back. The bank knows this in advance — it just doesn't know which borrowers. So how much money should the bank set aside today to cover future losses it expects but hasn't yet seen?
2
The Old Way (Before 2018) Banks waited until a borrower actually showed signs of trouble before setting aside money. This was called the incurred loss model. The problem: by the time you saw trouble, it was too late — the 2008 crisis exposed banks holding far too little against loans that were obviously going bad.
3
The New Way — IFRS 9 (Since 2018) Banks must now set aside money from day one of every loan, based on the losses they expect — even before any sign of trouble. This is the Expected Credit Loss (ECL) model. It's forward-looking, not backward-looking.
4
The Magic Formula Every dollar of provision comes from answering three simple questions: (1) What's the chance this borrower defaults? (2) If they default, how much do we lose? (3) How much money is at stake? Multiply these three together and you have your expected loss.
…car insurance. The insurer doesn't wait for you to crash before collecting premiums. They estimate (a) the chance you'll have an accident, (b) how expensive the repair would be, and (c) the value of your car. ECL works exactly the same way for loans.

Before 2018, banks recognized credit losses only after a loss event had occurred — the so-called incurred loss model of IAS 39. This led to the criticism of "too little, too late": provisions arrived only after defaults, amplifying the procyclicality of bank earnings during the Global Financial Crisis.

IFRS 9 Financial Instruments, effective January 2018, replaced incurred loss with the Expected Credit Loss (ECL) model. Banks must now recognize expected losses from day one of every loan, with the time horizon and magnitude depending on the loan's credit risk migration since origination.

The ECL framework rests on three pillars: the probability of default (PD), the loss given default (LGD), and the exposure at default (EAD). These are not statistical curiosities — they are the operational levers that translate macroeconomic scenarios into balance-sheet provisions.

The Core Equation FUND·01
ECL = PD × LGD × EAD × D

Where D is the discount factor to present-value future losses at the effective interest rate. For lifetime ECL across multiple periods:

ECL = Σₜ PDₜ × LGDₜ × EADₜ / (1+EIR)ᵗ

The integral is taken over marginal (not cumulative) PDs, conditional on survival to time t.

Worked Example · The Simplest ECL
You lend $100,000 to a small business. Based on history, similar borrowers have a 3% chance of defaulting this year. If they do default, you typically recover 60% of the loan, meaning you lose 40%. ECL = PD × LGD × EAD
ECL = 3% × 40% × $100,000 = $1,200 So the bank sets aside $1,200 today against this loan, even though nothing has gone wrong yet. That's the entire philosophy in one calculation.
Three Stages · Three Horizons
The stages are like traffic lights for a loan. A new, healthy loan starts at green (Stage 1) — we expect it to perform, so we only worry about losses in the next 12 months. If we see warning signs the borrower is getting weaker, we move to yellow (Stage 2) — now we worry about losses across the entire remaining life of the loan, because the risk is real. If the borrower actually defaults or stops paying for 90+ days, we move to red (Stage 3) — the loan is "credit-impaired" and we treat losses as practically certain.
Stage 1 — Performing
12-moECL

Loans with no significant increase in credit risk (SICR) since origination. Banks provision for losses expected from default events within 12 months only.

Interest income is recognized on the gross carrying amount.

Stage 2 — Underperforming
LifetimeECL

SICR has been triggered (e.g., 30+ DPD, downgrade, watchlist) but the loan is not credit-impaired. Provisioning expands to losses over the full remaining life.

Interest income still recognized on gross carrying amount.

Stage 3 — Credit-Impaired
LifetimeECL

Default has occurred (typically 90+ DPD, bankruptcy, or other credit-impaired evidence). Lifetime ECL continues, but the "loss" is often close to certain.

Interest income recognized on net carrying amount (gross less ECL allowance).

What This Laboratory Will Teach You

Tab II — Components. Build intuition for how each component (PD, LGD, EAD) is constructed from data. Explore PD term structures, downturn LGD, and CCF-based EAD for off-balance commitments.

Tab III — Staging. Apply SICR criteria — quantitative (PD doubling), qualitative (watchlist), and backstop (30+ DPD) — to migrate exposures across stages.

Tab IV — Single Loan. Compute ECL for an individual loan under different stages, durations, and parameter choices. See the period-by-period decomposition.

Tab V — Portfolio. Aggregate a synthetic portfolio of 250 loans across rating buckets. Visualize stage distribution, coverage ratios, and the migration matrix.

Tab VI — Macro Overlay. Apply probability-weighted multi-scenario adjustments (base / upside / downside) using satellite models that link PD to GDP and unemployment.

Tab VII — Sensitivity. Run tornado charts and stress tests to identify which inputs move ECL the most.

→
How To Use This Laboratory
Suggested Path
A
Build the building blocks (Tabs II–III) Start with Components to see how PD, LGD, EAD are constructed. Then Staging shows how loans are classified into the three buckets. These two tabs teach the vocabulary you need.
B
Compute one loan, then a portfolio (Tabs IV–V) Single Loan walks through the calculation year-by-year for one borrower. Portfolio aggregates 250 loans so you can see how stage mix and rating distribution drive the total provision.
C
Add real-world complexity (Tabs VI–VII) Macro Overlay shows the IFRS 9 requirement to use forward-looking economic scenarios. Sensitivity reveals which assumptions matter most for your provision — invaluable for stress testing.
D
Reference the glossary (Tab VIII) Every acronym you'll meet in a credit committee, with plain-English definitions.
Methodological note. This simulator uses simplified analytical forms (logistic PD, Gaussian copula for correlation, constant-LGD assumption with downturn add-on). Real-world ECL engines may use through-the-cycle PD calibration, Markov chain migration matrices, and Monte Carlo for macroeconomic scenarios. The goal here is conceptual mastery, not regulatory submission.
Up next — Step II
Now that you have the framework, let's open the hood. The three components — PD, LGD, EAD — each have their own modelling traditions.
The Three Components
PD, LGD, EAD — each with its own data lineage, modelling tradition, and behavioural quirks.
i
If you only remember three things from this tab
  • PD answers how likely? · LGD answers how much if it happens? · EAD answers against how much?
  • The three multiply, so an error in any one is an error in ECL — but PD is usually the largest source of estimation uncertainty.
  • Each component requires its own data, methodology, validation cycle, and model risk controls. Banks routinely have separate teams owning each.
?
The Three Ingredients of Every ECL
Foundation
1
PD answers: "How likely is default?" Expressed as a percentage. A 2% PD means: out of 100 borrowers like this one, we expect roughly 2 to default in the time window we're measuring. Better-rated borrowers (AAA, AA) have tiny PDs; weaker borrowers (B, CCC) have large ones.
2
LGD answers: "If default happens, how much do we lose?" Also a percentage. If you lend $100 and recover $60 after default (through collateral sale, guarantees, etc.), your LGD is 40%. Secured loans (with real estate, vehicles) have lower LGD than unsecured loans.
3
EAD answers: "How much money is at stake?" Expressed in dollars. For a fully drawn term loan, EAD is roughly the outstanding balance. For a credit line where the borrower can still draw more (like an overdraft), EAD includes expected future drawings — captured by the CCF (Credit Conversion Factor).
…a weather forecast for a picnic. PD is the chance it rains. LGD is how badly the rain ruins your day if it does come (light drizzle vs. thunderstorm). EAD is how big and expensive the picnic is. Multiply the three and you get your expected misery.
Why these three? Because together they cover every dimension of loss. PD captures the likelihood of bad news. LGD captures the severity when bad news arrives. EAD captures the size of the bet. Miss any one and your provision is wrong.
Probability of Default PD
P(default) = 1 / (1 + exp(-(α + β·X)))

The likelihood a borrower fails to meet contractual obligations. Typically calibrated from internal default histories using logistic regression, survival models, or rating agency mappings.

Through-the-cycle (TTC) PD averages across business cycles; point-in-time (PIT) PD reflects current conditions. IFRS 9 prefers PIT, adjusted for forward-looking macro information.

Term structure: PD typically rises with horizon. A 1-year PD of 1.5% can imply a 5-year cumulative PD near 7%.

Move the rating from AAA to CCC. Notice how the PD scale changes dramatically — a CCC borrower can have 100× the default probability of an AAA. Now extend the horizon from 1 to 10 years and watch cumulative PD rise.
Loss Given Default LGD
LGD = 1 − Recovery / EAD

The fraction of exposure lost if default occurs, after recoveries from collateral, guarantees, and legal proceedings. Net of recovery costs.

Downturn LGD reflects recoveries during economic stress (typically 10–20pp higher than long-run averages). IFRS 9 requires unbiased, probability-weighted LGD.

Drivers: collateral type & LTV, seniority, jurisdiction (creditor rights), workout efficiency.

Computed LGD
35.0%
Base LGD 25.0% · Downturn +10pp
Switch from Unsecured to Cash / Deposit. LGD drops dramatically — because cash collateral is virtually risk-free. Now try Real Estate at LTV 30% vs 100%. Higher LTV means less equity buffer, so more expected loss.
Exposure at Default EAD
EAD = Drawn + CCF × Undrawn

The amount outstanding at the moment of default. For amortizing loans, this is the future principal balance. For credit lines, it includes future draws via the Credit Conversion Factor (CCF).

CCF typically ranges from 20% (uncommitted retail) to 100% (committed corporate revolver). Behavioural studies show drawdowns accelerate as borrowers approach distress.

For revolvers, IFRS 9 uses the contractual cancellation period when applicable.

Computed EAD
$1,375,000
= Drawn + CCF × Undrawn
Set Drawn = $1M, Undrawn = $500K, CCF = 75%. EAD = $1M + 75% × $500K = $1.375M. Why CCF? Because distressed borrowers tend to draw down their unused lines right before they default — so we should assume some of that $500K will be drawn.
✓ Quick Check
A loan has PD = 5%, LGD = 60%, EAD = $200,000. What's the ECL?
Answer: 5% × 60% × $200,000 = $6,000. The bank holds $6,000 in reserve against this loan. Three numbers. One multiplication. That's the whole job — everything else is just better estimates of those three numbers.
From Numbers to Reality
What if this $6,000 ECL were on every loan?
Per-Loan ECL: $6,000
Across 10,000 Loans
$60M
Total provisions sitting on the balance sheet — a real liability-side reserve the bank cannot lend out or distribute as dividends.
Annual P&L Impact
−$60M/yr
If portfolio composition stays stable, $60M flows through the income statement as impairment expense each year — reducing net profit by that exact amount.
Capital Implications
−60 bps CET1
For a bank with $10B of capital, $60M in provisions reduces CET1 by 60 basis points. Enough to materially affect dividend capacity and stress test outcomes.
Why this matters: these numbers feel small per loan but enormous in aggregate. PD, LGD, and EAD are not abstract inputs — every basis point of estimation error scales to millions on a real bank's books. Estimation precision is a competitive advantage.
?
Knowledge Check · 3 Questions
Q1. Two identical loans have same PD and EAD, but Loan A has LGD = 30% while Loan B has LGD = 60%. What's the ECL relationship?
Q2. Which component is typically the largest source of estimation uncertainty in ECL?
Q3. A credit card has $5,000 drawn and $15,000 undrawn limit. What's the relevant EAD concept?
Up next — Step III
You can now measure each piece. But IFRS 9 doesn't apply them uniformly — it depends on which stage a loan sits in. Let's see how staging works.
The Staging Decision
Where ECL is allocated — and how exposures travel between stages. The fulcrum of every IFRS 9 implementation.
i
If you only remember three things from this tab
  • The single biggest driver of ECL is whether a loan is in Stage 1 (12-month) or Stage 2 (lifetime).
  • What triggers the move isn't the level of credit risk — it's the significant increase (SICR) from origination.
  • Migration is asymmetric: easy to deteriorate (Stage 1 → 2), hard to cure (Stage 2 → 1). Designed to avoid flattering coverage ratios.
?
Why Staging Matters So Much
The Big Idea
1
The horizon flips when a loan moves from Stage 1 to Stage 2 In Stage 1, you only count losses you expect in the next 12 months. The moment a loan moves to Stage 2, you must count losses across its entire remaining life. For a 5-year loan, that's roughly a 5× expansion of the provision overnight.
2
The trigger is "SICR" — Significant Increase in Credit Risk IFRS 9 doesn't tell banks exactly when to call SICR. Each bank designs its own rules. But they must combine three types of evidence: a numbers test, a judgement test, and a backstop nobody can override.
3
The three SICR tests (a) Numbers — has the borrower's PD doubled since the loan was originated? (b) Judgement — is the borrower on the watchlist, in forbearance, or breaching covenants? (c) Backstop — has the borrower been 30+ days late on payment? Any one of these can trigger Stage 2.
4
Stage 3 is the "default" zone Once a borrower is 90+ days past due, or shows clear evidence of credit impairment (bankruptcy filing, restructuring), they're in Stage 3. Lifetime ECL continues, but now interest is recognized on the net balance (after deducting the ECL provision).
…a hospital triage system. Stage 1 = walking patients (routine monitoring). Stage 2 = patients with worrying symptoms (full workup needed). Stage 3 = patients in critical condition (full intensive care). The system reacts to severity, and the resources committed grow with it.
Worked Example · A Loan Migrating Through Stages
A bank originates a 5-year, $1M loan to a BBB-rated company. PD at origination is 0.4% (12-month). At each reporting date: Day 1 (Stage 1) · PD unchanged → ECL = 0.4% × 40% × $1M = $1,600 (12-mo only)
Year 1 (Stage 2) · Industry downgrade, lifetime PD now 4% → ECL = lifetime ≈ $60,000
Year 2 (Stage 3) · Borrower 95 days past due → ECL = severe lifetime ≈ $280,000 Watch how the same loan can require provisions ranging from $1,600 to $280,000 depending on its stage. Stage migration is the single largest driver of P&L volatility in IFRS 9 reporting.

Significant Increase in Credit Risk (SICR) is the gateway between Stage 1 and Stage 2. It is not defined precisely by the standard; banks must develop their own framework, typically combining:

Quantitative trigger: a relative or absolute deterioration in lifetime PD since origination. A common rule is "PD has doubled and exceeded a minimum threshold."

Qualitative trigger: watchlist inclusion, covenant breach, forbearance, sector downgrade, or any internal flag of elevated risk.

Backstop (mandatory): 30+ days past due (rebuttable). 90+ DPD triggers Stage 3 unless rebutted with strong evidence.

SICR Trigger Calculator
Start with origination PD 1.5% and current PD 1.6%. The loan stays in Stage 1. Now slowly raise current PD — at what point does Stage 2 trigger? (Answer: when current ÷ original ≥ 2.0×.) Then set DPD to 35 — Stage 2 triggers immediately regardless of PD ratio. That's the backstop in action.
Migration Matrix · One-Year Transitions
How to read this table. Each row says "if a loan starts the year in this stage, here are the probabilities of where it'll be at year-end." For example, the top row says: a Stage 1 loan has a 92% chance of staying in Stage 1, a 6% chance of slipping to Stage 2, and a 2% chance of jumping straight to Stage 3 (a severe shock). These probabilities should add up to 100% across each row.
Stage-to-Stage Transition Probabilities (Illustrative)

A migration matrix captures the probability of exposures moving between stages over one year. Diagonal terms represent staying put; off-diagonal terms capture deterioration (upper-right) or cure (lower-left).

Cure considerations. Stage 2 exposures can return to Stage 1 only after a probation period (typically 3 months minimum) of consistent performance, evidencing that the SICR has reversed. For Stage 3, cure requires meeting tighter conditions over a longer period (often 12 months). Reverse migration discipline is essential to avoid artificially flattering coverage ratios.
Visualizing PD Over Time — 100 Borrowers, Year by Year

To build intuition for what "2% PD" really means, imagine 100 identical borrowers, each with a 2% annual probability of default. Adjust the slider below to watch what happens each year. Surviving borrowers stay green; defaulted ones turn red and shrink.

Notice that even with a low annual PD, cumulative defaults add up over a multi-year loan. This is exactly why Stage 2's lifetime horizon matters so much.

Annual PD (Marginal)
2.0%
Probability of default this year, given the loan is still performing at year-start.
Cumulative PD
9.6%
Probability of default at some point within the years elapsed.
Lifetime PD (Total Term)
18.3%
Cumulative probability over the entire remaining loan life — this is what Stage 2 ECL uses.
Cohort of 100 Borrowers After 5 Years
90 still performing
10 defaulted

Why this visualization matters for ECL: a Stage 1 loan uses the marginal annual PD for next 12 months. A Stage 2 loan uses the cumulative PD over remaining life. The bigger the gap between the two, the bigger the Stage 1 → Stage 2 jump.

For a borrower with 2% annual PD and 10 years remaining, the lifetime PD is ~18.3% — that's 9× higher than the 12-month figure. This is the math behind the dramatic ECL jump on stage migration.

The Stage 1 → Stage 2 Cliff — Two Identical Loans, One Migrates

Both loans below are identical: $500,000 principal, 2% annual PD, 40% LGD, 5-year term. The only difference is that Loan B has just experienced a Significant Increase in Credit Risk (SICR) and migrated to Stage 2. Watch what happens to their respective ECL provisions.

Loan A · Stage 1
12-Month Horizon
Expected Credit Loss
$4,000
= 2% × 40% × $500,000
A healthy performing loan. ECL counts only the loss expected if default happens within the next year.
SICR →
4.8×
Higher
Loan B · Stage 2
Lifetime (5-Year) Horizon
Expected Credit Loss
$19,229
= 9.6% × 40% × ~$500,000 avg EAD
where 9.6% = 1 − (1−0.02)⁵
Same loan, same PD, same LGD. But the horizon is now lifetime. Cumulative PD over 5 years compounds the annual 2% into 9.6% — and ECL with it.
The provisioning cliff in real bank terms
If a bank has 10,000 loans like Loan A, total Stage 1 provisions = $40 million. If a recession migrates 15% of those loans to Stage 2, the lifetime ECL on those 1,500 loans alone = $28.8 million — a $22.8 million provision charge in a single quarter. This is exactly why banks' quarterly earnings swing so much with credit cycles — the stage migration effect, not actual losses, drives the volatility.
?
Knowledge Check · 3 Questions
Q1. What triggers the move from Stage 1 to Stage 2?
Q2. Why does Stage 2 ECL increase so dramatically vs Stage 1, even when nothing about the loan's actual risk has changed yet?
Q3. A Stage 2 borrower starts making timely payments again. Can the loan immediately move back to Stage 1?
Up next — Step IV
You've classified the loan. Now let's actually compute its ECL — cashflow by cashflow, discounted to present value.
Single Loan Computation
Build ECL one cashflow at a time. See exactly where the provisioning comes from.
i
If you only remember three things from this tab
  • ECL is built by computing year-by-year expected losses and summing — Stage 1 sums one year, Stage 2 sums all remaining years.
  • Each year's contribution is discounted to present value using the loan's effective interest rate.
  • Most of ECL typically comes from the middle years of a loan's life — early years have low default probability; later years have low remaining exposure.
?
How ECL Is Built — Year by Year
The Mechanics
1
For Stage 1: just one year of losses ECL = PD × LGD × EAD. That's it. You compute the expected loss for the next 12 months only. Simple.
2
For Stage 2 and Stage 3: every year of the loan's life Now you compute the expected loss for each future year, then add them up. Year 1 has its expected loss, Year 2 has its expected loss, and so on. The total across all years is your lifetime ECL.
3
But here's the twist: survival matters To default in Year 3, the borrower must have survived Years 1 and 2. So the probability of "defaulting in Year 3" is actually: (chance of surviving Years 1-2) × (chance of defaulting in Year 3 given survival). This makes later-year contributions smaller than they look.
4
And one more twist: discounting A $1,000 loss next year is worth less today than a $1,000 loss this year (because of the time value of money). We divide each year's expected loss by (1 + interest rate)year to bring it back to today's value. This is called present-valuing the loss.
…calculating expected revenue from a subscription business. You estimate Year 1 revenue, Year 2 revenue (accounting for churn = "survival"), Year 3 revenue (accounting for two years of churn), and so on — then discount each to today. ECL is the same math, just for expected losses.
Worked Example · Lifetime ECL for a 3-Year Loan
$1M, 3-year loan, annual PD = 2%, LGD = 40%, EIR = 8%. Y1: Survival=100%, Marg.PD=2.0%, EAD=$1M → Loss=$8,000, PV=$8,000÷1.08 = $7,407
Y2: Survival=98%, Marg.PD=1.96%, EAD~$700K → Loss=$5,488, PV÷1.08² = $4,703
Y3: Survival=96%, Marg.PD=1.92%, EAD~$350K → Loss=$2,688, PV÷1.08³ = $2,134
Total Lifetime ECL ≈ $14,244 Notice how (a) survival rates compound, making later years contribute less, (b) EAD shrinks as the loan amortizes, and (c) discount factors reduce later-year losses further. Three forces all pulling later contributions down.
Loan Parameters
Total ECL
$0
— of principal
Coverage Ratio
0%
Stage 1
Period Decomposition — Marginal ECL by Year
Marginal ECL
EAD Profile
Period-by-Period Computation Table
Reading the table. For each period t: PDt is the marginal probability of defaulting in year t conditional on surviving prior years. EADt reflects the amortization schedule. The discount factor 1/(1+EIR)t brings future losses to present value. Stage 1 truncates summation at t=1; Stage 2/3 sums over the full lifetime.
Start with default settings and look at the total ECL number. Now switch the stage from 1 to 2 — watch the number explode (5–10× larger, depending on tenor). That single change is the most important lever in IFRS 9. Now toggle amortization from Equal Installment to Bullet — bullet loans keep the full principal exposed until maturity, so ECL is higher. Try increasing the term from 3 to 10 years to see compounding effects.
✓ Quick Check
Why does a Stage 2 ECL grow with the loan's remaining tenor, but a Stage 1 ECL doesn't?
Answer: Stage 1 ECL only counts losses in the next 12 months — it doesn't care if the loan has 2 years left or 20 years left. Stage 2 ECL sums losses across the entire remaining life — so a 10-year loan has roughly 10× as many years of losses to count as a 1-year loan.
Putting Numbers In Context
What does one loan's ECL mean for the bank?
See Per-Loan ECL Above
Balance Sheet
Liability
The ECL becomes a loan-loss reserve on the liability side. It reduces the carrying value of the loan asset by the same amount — the loan now appears "smaller" on the books.
Income Statement
Hits P&L
Provisioning hits the impairment expense line. Every dollar of new ECL is a dollar less of pre-tax profit. Reversing provisions (when credit improves) flows back through as negative impairment — a P&L boost.
Capital & Ratios
Reduces CET1
P&L flows to retained earnings, which is part of CET1 capital. Provisions therefore reduce the regulatory capital ratio one-for-one — affecting dividend capacity and stress-test results.
The hidden link: a single number on the ECL screen is connected to three different financial statements, regulatory capital, and shareholder distributions. This is why ECL governance is treated as a board-level matter at every bank.
?
Knowledge Check · 3 Questions
Q1. A 5-year loan in Stage 2: which year typically contributes the most to the total ECL?
Q2. Why must each year's expected loss be discounted to present value in the ECL calculation?
Q3. A bullet loan has higher lifetime ECL than an equal-installment loan with the same principal, term, PD and LGD. Why?
Up next — Step V
One loan is just one loan. Real banks have thousands. Let's scale up to a full portfolio and see how coverage ratios emerge.
Portfolio Aggregation
A synthetic book of 250 loans across rating buckets. Watch coverage ratios respond to your assumptions.
i
If you only remember three things from this tab
  • Portfolio ECL is just the sum of individual loan ECLs — but the composition across stages drives the headline coverage ratio.
  • The coverage ratio (ECL ÷ total exposure) is the metric regulators and analysts watch most closely.
  • A small shift in Stage 2 share (e.g. 5% → 12%) can move coverage ratios more than any single PD or LGD assumption change.
?
From One Loan to a Bank's Balance Sheet
Scaling Up
1
Total ECL = sum of every loan's individual ECL Each loan is computed independently using its own PD, LGD, EAD, and stage. Then the bank's total provision is just the sum. No magic — but the composition of the portfolio dominates the total.
2
Three things shape the portfolio's total ECL (a) Rating mix — what fraction is AAA vs CCC. (b) Stage mix — what fraction is Stage 1 vs Stage 2 vs Stage 3. (c) Sector concentration — if too much is in one industry, correlated defaults can dominate.
3
"Coverage ratio" is the headline metric Coverage = Total ECL ÷ Total EAD, usually expressed in basis points. A retail bank might have 100–200 bps coverage; a stressed corporate bank might have 400–600 bps. Comparing coverage across banks and over time is how analysts judge provisioning adequacy.
4
"Stage 2+3 share" is the leading indicator If a growing slice of the portfolio is in Stage 2 or 3, total ECL is about to rise sharply — because lifetime ECL is so much bigger than 12-month ECL. Watching the migration into Stage 2 is more informative than watching ECL itself.
…a hospital tracking its patient census. Total provisions = total patients × average cost per patient. But mix matters enormously: a hospital with mostly ICU patients (Stage 3) has wildly different costs than one with mostly outpatients (Stage 1), even at the same headcount.
How to use this tab. The portfolio is randomly generated each time you change a slider. Rating Skew tilts the book toward investment-grade or high-yield. Industry Concentration packs more loans into fewer sectors. Macro Stress multiplies everyone's PD (simulating a recession). Watch how the four KPIs at the top respond — these are exactly the metrics a risk committee tracks each quarter.
Total EAD
$0M
250 loans
Total ECL
$0K
— bps coverage
Weighted PD
0%
Exposure-weighted
Stage 2+3 Share
0%
— loans
Portfolio Generator Controls
Stage Distribution by EAD
S1
S2
S3
ECL Contribution by Rating Bucket
Portfolio Sample — Top 15 Exposures by ECL
Experiment 1: Set Rating Skew to High Yield. Watch coverage jump from ~50 bps to ~400 bps. Experiment 2: With balanced skew, slide Macro Stress from 1.0× to 3.0×. The Stage 2+3 share explodes — this is how a recession shows up in IFRS 9 numbers. Experiment 3: Hit "Regenerate Portfolio" a few times with the same settings. Notice the variation — small portfolios are statistically noisy, which is why real banks need thousands of loans before estimates become reliable.
✓ Quick Check
If a bank's Stage 2+3 share jumps from 8% to 15% in one quarter, what happens to total ECL?
Answer: It rises sharply — typically 2–4× the percentage move — because the loans newly classified as Stage 2 now require lifetime ECL instead of just 12-month ECL. This is the mechanic that made Q1 2020 (COVID-19) and Q3 2022 (inflation shock) such painful provisioning quarters for global banks.
Up next — Step VI
A portfolio number is only as good as its assumptions about the future. Let's add macroeconomic scenarios — base, upside, downside.
Macroeconomic Overlay
IFRS 9 demands forward-looking, probability-weighted ECL. Three scenarios — base, upside, downside — combined into a single, defensible provision.
i
If you only remember three things from this tab
  • ECL is probability-weighted across scenarios, not a single point estimate — typical practice uses 3 scenarios (base, upside, downside) with assigned probabilities.
  • Because the loss function is convex, the probability-weighted ECL is higher than the ECL of the base-case scenario alone.
  • Scenario probabilities are themselves judgmental — auditors and regulators scrutinize whether they reflect current macro conditions or just a fixed historical average.
?
Why Banks Don't Just Use One Forecast
Forward-Looking Information
1
PDs aren't static — they change with the economy Default rates rise sharply during recessions and fall during expansions. A 2% PD calibrated on calm-times data will badly understate losses heading into a downturn. IFRS 9 requires banks to look forward and adjust.
2
The future is uncertain — so banks consider multiple scenarios Banks build at least three economic scenarios: a base case (what they think is most likely), an upside (things go better), and a downside (recession). Each scenario produces its own PD path and its own ECL.
3
Then they weight the scenarios by probability If base case = 50%, upside = 25%, downside = 25%, then: Final ECL = 50% × ECL(base) + 25% × ECL(up) + 25% × ECL(down). This is the "probability-weighted, unbiased" requirement of IFRS 9.
4
The mechanism: a "satellite model" A statistical model links macro variables (GDP, unemployment, house prices) to PDs. When GDP falls, the model says "PDs should rise by X%." Each macro scenario plugs into this model to generate a stressed PD multiplier, which is then applied to baseline PDs.
…weather forecasting for a wedding. You don't just plan for the most likely weather — you plan for the chance of rain too. A 70% sunny / 30% rainy forecast means you book a tent. ECL works the same way: even if base case is most likely, the downside is bad enough that ignoring it would underprovide.
The convexity trick. If you just computed ECL on the "average" economic forecast, you'd get the wrong answer. Why? Because PD rises faster during bad times than it falls during good times — the relationship is curved (convex), not straight. The probability-weighted ECL across three scenarios is always larger than the ECL at the average scenario. This convexity uplift is real money — typically 5–15% of total provisions.

The satellite model links macroeconomic variables (GDP growth, unemployment, interest rates, house prices) to default rates. A common functional form regresses log-odds of default on lagged macro variables:

log(PDₜ / (1 − PDₜ)) = α + β₁·ΔGDPₜ + β₂·URₜ + β₃·ΔHPIₜ + ε

Each scenario produces its own PD path. ECL is computed under each, then probability-weighted to satisfy the unbiased requirement. The non-linearity of PD-to-macro means the weighted-PD ECL ≠ average-scenario ECL — convexity matters.

Scenario Definition
Base Case
Weight · 50%
Upside
Weight · 25%
Downside
Weight · 25%
Base Case
PD Multiplier: 1.00×
Upside
PD Multiplier: 1.00×
Downside
PD Multiplier: 1.00×
Probability-Weighted Result
PD Path Under Each Scenario
Up
Base
Down
PW
Single-Scen ECL (Base)
$0K
Baseline reference
Prob-Weighted ECL
$0K
— vs base
ECL Contribution by Scenario
Convexity uplift. Because PD is a nonlinear (convex) function of macro stress, the probability-weighted ECL exceeds the ECL computed at the weighted-average macroeconomic state. This is why simply running ECL on a "base" scenario understates required provisions — the downside scenario's high-loss outcomes are not offset symmetrically by the upside.
Experiment 1: Set all weights to Base=100%. Note the PW ECL equals the base ECL. Experiment 2: Now shift to Base=50%, Up=25%, Down=25%. The PW ECL rises above base, even though the weights are "symmetric." This is the convexity uplift. Experiment 3: Make the downside really severe (GDP -5%, UR 12%). Watch the PD multiplier explode and the PW ECL jump — this is what happened to bank provisions in March 2020.
✓ Quick Check
Why does a bank put weight on a downside scenario even when it thinks the base case is most likely?
Answer: Because IFRS 9 demands an unbiased, probability-weighted estimate. Even a 10% chance of a severe downside contributes meaningfully to expected loss — and ignoring it would mean systematically under-provisioning. Banks would much rather hold a bit extra than face a regulatory finding that their ECL was biased toward optimism.
Up next — Step VII
You have a full provisioning model. But which assumptions move the number the most? That's what sensitivity analysis answers.
Sensitivity & Stress
Which dial moves provisions most? A tornado decomposition for the curious risk manager.
i
If you only remember three things from this tab
  • ECL is most sensitive to PD shifts at the lower rating end and LGD shifts at the higher exposure end.
  • The bigger the Stage 2 share, the more sensitive total ECL becomes to every assumption — because lifetime horizons compound errors.
  • Stress tests should flex the worst-case scenario probability, not just the worst-case parameter values — this is what regulators look for.
?
Knowing Which Lever Moves the World
Stress Testing 101
1
Not all inputs are equally important ECL depends on PD, LGD, EAD, and tenor. But if you shock each by 20%, do they all move ECL by 20%? No — some inputs amplify, some dampen. Understanding which is which is the heart of stress testing.
2
A "tornado chart" ranks driver sensitivity For each input, we shock it up by X% (red bar to the right) and down by X% (green bar to the left). The longer the total bar, the more sensitive ECL is to that driver. The chart is sorted with the most powerful drivers at top — like a tornado funnel.
3
PD and LGD are typically near-linear; tenor is non-linear A 20% increase in PD usually means a 20% increase in ECL — linear. But a 20% increase in tenor (e.g., 4 years to 5 years) can mean a much larger ECL increase because each new year adds another full year of expected losses. Tenor is the most powerful single lever in many portfolios.
4
The 2D heatmap shows combined stress Real stresses don't move one variable at a time. Recessions push PD and LGD up together. The PD × LGD heatmap shows ECL at every combination — the corner where both are highest is the "severely adverse scenario" regulators love to see banks survive.
…an aircraft pre-flight check. Pilots don't just verify the engine works — they test every system to find which failure mode would be most catastrophic. Tornado charts do the same for ECL: which input failure would hurt provisions most?
Why this matters operationally. When you go to a Risk Committee with a provision number, the first question is always: "How sensitive is this to your assumptions?" The tornado answers that question directly — and tells you where you need the strongest model validation. If LGD is the dominant driver, your LGD methodology needs the most scrutiny. If tenor matters most, the front office's tenor assumptions need challenging.
Base Configuration
Tornado interpretation: Each bar shows the change in total ECL when the named driver is shocked up (+) and down (−) by the shock magnitude, holding all others constant. The longer the bar, the more sensitive ECL is to that input.
Base ECL
$0K
Most Sensitive To
PD
Range Width
$0K
Tornado — Sensitivity of ECL to Drivers
+Shock
−Shock
Two-Way Stress · PD × LGD Heatmap
ECL ($K) Across Joint Stress Scenarios
Experiment 1: Set shock to 20% and observe the tornado ranking. Now raise shock to 50%. Notice the ranking can change — non-linear drivers (like tenor) become more dominant at extreme shocks. Experiment 2: Push tenor from 2 to 10 years and rerun. ECL grows more than 5× even though "tenor only increased 5×" — the compounding effect of multi-year losses. Experiment 3: Find the cell in the heatmap where PD is 2× and LGD is 1.3× baseline. That cell represents a typical "moderate adverse" stress scenario.
✓ Quick Check
A risk manager says "our ECL is most sensitive to LGD." What does that imply for where the bank should spend its validation budget?
Answer: On LGD modelling. If LGD is the dominant driver, then collateral valuation, recovery assumptions, downturn adjustments, and workout cost estimates deserve the bulk of model validation effort. Spending 80% of validation resources on PD when LGD drives the provision is a misallocation — and a finding waiting to happen at the next audit.
Up next — Step VIII
You now understand every component, every dial, every sensitivity. Time to put it all together in the Workbench — pick your own methodology for each input.
The ECL Workbench
Build ECL from the ground up — choose your methodology for every input, see each formula live, and trace each calculation. The full atelier of credit risk modelling.
i
If you only remember three things from this tab
  • Real banks use different methods for different portfolios — logistic regression for retail, Vasicek/IRB for wholesale, migration matrices for rated exposures.
  • The choice of method is itself a model risk — independent validation must justify why one approach is more appropriate than another.
  • The Workbench mirrors the structure of a real ECL system: independent component models feeding a master engine. Change any input, the master recomputes.
⚒
What This Tab Does
Open The Black Box
1
Six modules — one per input Configure loan terms, then choose how PD is estimated (Vasicek / Logistic / Migration / Direct), how LGD is built (Workout / Market / Collateral / Direct), and how EAD is modelled (CCF / Behavioral / Amortizing). Each module exposes the formula, accepts your inputs, and shows the math.
2
Every methodology shows its formula and trace Pick Vasicek and see the asymptotic single-risk-factor formula with your asset correlation. Pick Logistic regression and supply your own coefficients. Every method shows what it computed and why, step by step.
3
Master ECL aggregates everything At the bottom, all three components combine into the final ECL — recomputed live as you change any input or method. You see the full chain from raw inputs to a single provision number.
…a workshop where each station does one job. Station 1 measures the loan. Stations 2-4 each estimate one ingredient (PD, LGD, EAD), and each station offers a choice of tools. The output of every station feeds into the final assembly. Change a tool at any station and watch the final product change.
1
Loan Terms
Define the instrument
Outputs → Tenor, Cashflows
Loan Amount (Principal) notional at origination
Term (Years) contractual maturity
Interest Rate (% p.a.) coupon / contractual rate
Effective Interest Rate (% p.a.) discount rate for ECL
Amortization how principal repays
Stage determines ECL horizon
Why these inputs?

The principal, term, and amortization schedule together define the EAD profile over time — how much will be outstanding when default could occur. The EIR is mandated by IFRS 9 as the discount rate for ECL (using a different rate, like the funding rate, would not faithfully represent the cash shortfall to the lender).

For amortizing loans, the EAD declines as principal repays — pushing late-year ECL contributions down. Bullet loans keep full exposure to maturity, dramatically inflating lifetime ECL.

Live Preview
Repayment Schedule & EAD Profile
How outstanding balance — and therefore Exposure at Default — evolves over the loan's life under each amortization method.
EAD over Time — Three Methods Compared
Equal Installment
Linear Principal
Bullet
If the borrower defaults in any given year, the height of the line at that year is what the bank is exposed to. Notice how Bullet loans stay at full exposure right until the end — much riskier from an ECL perspective.
Selected method: Equal Installment Same total payment every year. Early years are mostly interest; later years are mostly principal.
Year-by-Year Schedule (Selected Method) Showing payment composition + closing balance = next year's EAD
Year Opening Balance
(= EAD this year)		 Payment Interest Portion Principal Portion Closing Balance
Annual Payment
—
—
Total Interest Over Life
—
cost of borrowing
Average EAD
—
average outstanding over term
EAD at Year 1 vs Final Year
—
range of exposure
▸ Why this matters for ECL
—
2
PD — Probability of Default
Choose your estimation method
Output → PDt path
Method 1 · Direct Input

The simplest case: you already have a 12-month PD from your rating system (S&P, Moody's, internal scorecard, regulator-published averages). The lifetime PD path is built by holding the marginal rate roughly constant, with optional drift to mimic seasoning effects.

When to use: proxy portfolios, simple retail products, or when your rating system already provides reliable 12-month PDs.

12-Month PD (%) point-in-time
Annual PD Drift (%) period-on-period change
Marginal PD path PDₜ = PD₁ × (1 + drift)t-1
Method 2 · Logistic Regression Scorecard

The workhorse of consumer and SME PD modelling. A logistic regression maps borrower-level features (financial ratios, behavioral data, demographics) to a default probability bounded between 0 and 1.

The S-curve is the key insight: the function is nearly flat for very safe and very risky borrowers, and steepest in the middle — which matches reality, where small changes in financial health matter most for borderline obligors.

Logistic / Sigmoid function PD = 1 / (1 + exp(−z))
where z = β₀ + β₁·DSCR + β₂·Leverage + β₃·CurrentRatio
β₀ — Intercept baseline log-odds
DSCR debt service coverage
β₁ — DSCR Coefficient expect negative
Leverage (Debt/Equity) solvency ratio
β₂ — Leverage Coefficient expect positive
Current Ratio liquidity
β₃ — Current Ratio Coefficient expect negative
Method 3 · Vasicek Asymptotic Single Risk Factor (ASRF)

Oldrich Vasicek's 1987 model is the theoretical backbone of Basel II/III internal-ratings-based capital and underpins many regulatory stress-testing exercises. Each borrower's asset value is decomposed into a systematic factor (the economy) and an idiosyncratic factor (firm-specific). Default occurs when assets fall below a threshold determined by the borrower's PD.

The model produces a conditional PD — what the PD becomes when the economy is in a stressed state. This is exactly what IFRS 9 forward-looking adjustment needs.

Vasicek conditional PD (Basel IRB formula) PD(stress) = Φ((Φ⁻¹(PD) + √ρ · Φ⁻¹(conf)) / √(1 − ρ))
where Φ is the standard normal CDF, ρ = asset correlation, conf = stress confidence
Through-the-Cycle PD (%) long-run average
Asset Correlation ρ (%) comovement with economy
Stress Confidence (%) 99.9% = Basel; 50% = neutral
Exposure Type determines typical ρ
Interpretation

At conf = 99.9%, you get the Basel regulatory capital PD — what your PD would be in a one-in-a-thousand-year recession (typically 5–10× the TTC PD). At conf = 50%, you get the median-state PD, which is slightly below the TTC due to the convexity of the PD-to-factor relationship. To recover an IFRS 9 unconditional/PIT PD, use a confidence between 60–90% depending on your scenario weighting.

Method 4 · Markov Migration Matrix

Build a multi-year PD curve from a one-year rating transition matrix. Each cell M[i,j] = probability of moving from rating i to rating j in one year. Defaults accumulate via repeated matrix multiplication: M² gives two-year transition probabilities, M³ gives three-year, etc.

This is the most theoretically clean way to build a lifetime PD term structure from observable annual data, and it naturally handles rating drift (where today's BBB obligor becomes tomorrow's BB obligor before defaulting).

Cumulative PD after t years Cum PD(t) = Mᵗ[i, default]
where i is current rating, M is the 1-year transition matrix
Current Rating starting state
Matrix Source calibration data
12-Month PD
2.00%
Lifetime Cumulative PD
0.00%
Method: Direct Input · Term: 5 years
3
LGD — Loss Given Default
Choose your estimation method
Output → LGD %
Method 1 · Direct Input

Specify LGD directly from your loss database, regulatory floor, or expert judgement. Useful as a baseline to compare against more granular methods.

Base LGD (%) long-run average
Downturn Add-On (pp) IFRS 9 / Basel adjustment
Method 2 · Collateral-Based (Haircut Approach)

Build LGD from the collateral side: estimate the recoverable value after applying market-value haircuts, then deduct legal and recovery costs. The unrecoverable portion of EAD is LGD.

This is how most secured corporate lending and commercial real estate portfolios are modelled. Haircuts reflect distressed-sale discounts; in a downturn, real estate sells for less than appraisal.

Collateral-based LGD Recovery = Collateral × (1 − Haircut) − Costs
LGD = max(0, 1 − Recovery / EAD)
Collateral Type determines haircut
Collateral Market Value ($) current appraisal
Custom Haircut Override (%) leave 0 to use default
Legal & Workout Costs (%) of EAD
Method 3 · Workout LGD (Discounted Cashflow Approach)

The gold standard for sophisticated banks. Model the post-default cashflow timeline: when does the borrower start partial payments, when is collateral seized, when is the asset sold, when are legal costs incurred? Discount each cashflow back to the default date at the EIR.

The result is the economic loss — what you actually lose in present-value terms, not just what you fail to nominal-recover.

Workout LGD PV(Recoveries) = Σ CFₜ / (1 + EIR)ᵗ − PV(Costs)
LGD = 1 − PV(Recoveries) / EAD
Recovery Year 1 ($) partial payments / cash
Recovery Year 2 ($) collateral sale, etc.
Recovery Year 3 ($) final settlements
Workout Costs ($) paid in Year 1
EAD at Default ($) denominator
Discount Rate (% p.a.) EIR
Method 4 · Market-Implied LGD

For traded debt (corporate bonds, syndicated loans), market prices around default time give an immediate, market-validated LGD estimate. The price of defaulted debt typically reflects expected recovery, discounted for uncertainty and the time/cost of workout.

Industry-standard rule: LGD ≈ 100% − price-at-default. A bond trading at 35 cents on the dollar 30 days after default implies a market LGD of about 65%.

Market-implied LGD LGD = 1 − (DefaultedPrice / 100) × (1 − LiquidityAdj)
Defaulted Debt Price (cents on $1) 30 days post-default
Liquidity Adjustment (%) illiquid debt
Seniority payment priority
Computed LGD
50.0%
Method: Direct Input
4
EAD — Exposure at Default
Choose your estimation method
Output → EADt profile
Method 1 · Amortization Schedule

For term loans, EAD at year t is simply the outstanding principal balance at year t, computed from the contractual amortization schedule. The EAD profile inherits the amortization type (equal installment, linear, or bullet) from Module 1.

Average EAD over the loan's life is the typical input used for lifetime ECL — though more refined approaches use the period-by-period EAD profile directly.

Equal-installment payment (mortgage-style) PMT = P × r / (1 − (1+r)⁻ⁿ)
EADₜ = EADₜ₋₁ × (1+r) − PMT
Method 2 · Credit Conversion Factor (CCF)

For credit cards, revolvers, and overdrafts, the borrower can draw additional funds before defaulting. CCF estimates the fraction of the undrawn commitment that will be drawn by the time of default.

Empirical finding: as borrowers approach default, they draw heavily on available credit (the "race to default" effect). CCF values: ~75% for committed revolvers, ~40-60% for credit cards, ~20% for uncommitted lines.

CCF-based EAD EAD = Drawn + CCF × Undrawn
Currently Drawn ($) outstanding balance
Undrawn Commitment ($) available limit
CCF (%) expected drawdown
Facility Type suggests CCF
Method 3 · Behavioral / Loan Equivalent (LEQ)

An advanced variant of CCF. Instead of applying CCF to current undrawn, behavioral models predict the limit at default (which itself may have grown or shrunk). LEQ = fraction of additional draws relative to the limit observed empirically at default.

For credit cards: the limit at default tends to be 5-15% higher than the limit a year prior (as good repayment behavior earlier in the year triggered limit increases that the borrower then used in their downward spiral).

Behavioral EAD with limit dynamics Limitdefault = CurrentLimit × (1 + LimitGrowth)
EAD = Drawn + LEQ × (Limitdefault − Drawn)
Currently Drawn ($) balance now
Current Limit ($) authorized line
Limit Growth to Default (%) empirical observation
LEQ Factor (%) % of available drawn
EAD at Default
$1,000,000
Avg EAD Over Life
$500,000
Method: Amortization Schedule
5
Forward-Looking Macro Adjustment
Apply scenario weighting
Output → PD multiplier
Why apply a macro overlay?

The PD from Module 2 reflects current conditions, but IFRS 9 demands an unbiased, probability-weighted estimate over multiple economic scenarios. We apply a multiplier to scale the PD up (if the weighted-average scenario is worse than today) or down (if it's better).

For a complete three-scenario implementation, use Tab VI. Here we use a simplified single-multiplier approach.

Apply Macro Overlay? on/off
PD Multiplier 1.0 = no change
Scenario Description label only
∑
Master ECL Computation
Period-by-period assembly
FINAL OUTPUT
Lifetime ECL — the master equation ECL = Σₜ PDₜ(marg) × LGD × EADₜ × DFₜ
where PDₜ(marg) = Survivalₜ₋₁ × PDₜ, DFₜ = 1 / (1 + EIR)ᵗ
Expected Credit Loss · Master Result
$0·USD
PD (Effective)
0.00%
LGD
0.0%
EAD
$0
Coverage
0bps
Compare methodologies side by side. Set PD = Direct (2.0%), LGD = Direct (40%), note the ECL. Switch PD to Vasicek with ρ = 15%, conf = 75%. Notice the conditional PD lifts to ~3-4%. Now switch LGD to Workout with the default cashflows — observe how present-value discounting reduces LGD versus undiscounted recovery. Each methodology answers the same question with a different lens; the variance across methods is itself a useful uncertainty estimate.
✓ Quick Check
Why does the Vasicek model produce a higher PD than the through-the-cycle input when confidence > 50%?
Answer: Because Vasicek converts an unconditional PD into a PD conditional on a stressed economic state. At 50% confidence, you're asking "what's PD when the economy is average?" — answer: the unconditional PD. At 99.9%, you're asking "what's PD when the economy is at its 1-in-1000 worst?" — answer: much higher. The Basel IRB capital formula uses 99.9%; IFRS 9 typically uses 50-90% depending on scenario weights.
Up next — Step IX
You've mastered the mechanics. Now let's see them in action — real moments when ECL accounting reshaped financial history.
Historical Case Studies
Real moments when expected-credit-loss accounting (or its absence) shaped financial history. Each case anchors an abstract rule in a story that actually happened.
i
Why these stories matter
  • The 2008 crisis was caused, in part, by accounting rules that delayed loss recognition.
  • IFRS 9 was designed to fix this — but its 2018 introduction caused its own shock to bank books.
  • COVID-19 in 2020 became the first real-world stress test of forward-looking ECL.
2007–2009
The Incurred-Loss Disaster
$1.3TGlobal Bank Losses

Before IFRS 9, banks worldwide used the incurred-loss model. A loan-loss provision could only be recognized when a default event had already occurred or was demonstrably imminent. Banks would happily originate billions of subprime mortgages and book zero impairment on them, because — technically — no losses had yet been "incurred."

By mid-2007, the cracks were obvious. Subprime delinquencies were rising. Housing prices had peaked. But the accounting rules forbade banks from reflecting expected future losses on their books. Earnings looked healthy through 2007 even as the underlying portfolios were deteriorating.

When the dam broke in late 2008, banks recognized massive provisions all at once. Citigroup, Bank of America, and Wells Fargo each took write-downs exceeding $30 billion in a single quarter. The pro-cyclicality of the incurred-loss model — silence on the way up, panic on the way down — amplified the crisis.

Why this case matters
The G20 explicitly identified incurred-loss accounting as a structural cause of the crisis. The IASB began work on what would become IFRS 9 in response — a regime requiring banks to recognize expected losses before they happen, not after.
January 1, 2018
The IFRS 9 Day-One Shock
−€140BEU Bank Equity Hit

On January 1, 2018, IFRS 9 went live across most of the world (the U.S. uses a parallel rule, CECL, which followed in 2020). Every bank had to re-classify every loan into Stages 1, 2, or 3 and re-compute provisions on the new expected-loss basis.

The aggregate effect was substantial and immediate. European banks took an average 9% increase in loan-loss provisions overnight. UK banks averaged 13%. Some emerging-market banks saw provisions jump 20–40%. The hit went directly to retained earnings — and therefore directly to CET1 capital ratios.

To soften the blow, the Basel Committee permitted a five-year transitional add-back: banks could phase in the capital impact at 95%, 85%, 70%, 50%, and 25% reductions over 2018–2022. Without this relief, several mid-sized European banks would have breached their minimum capital requirements on day one.

Why this case matters
The day-one shock revealed how much "hidden" expected loss had been accumulating under the old incurred-loss regime. It also became the template for how regulators handle accounting transitions: recognize the new reality, but soften the path so banks aren't forced into fire sales.
March 2020
The COVID Overlay Debate
$130BQ1 2020 US Bank Provisions

March 2020 was the first real stress test of forward-looking ECL. Economic models suddenly forecasted Great-Depression-magnitude shocks. If banks mechanically applied their ECL models with these new macro inputs, the resulting provisions would have been catastrophic — wiping out years of bank earnings in a single quarter.

Three responses emerged. Regulators (the IASB, ECB, and PRA) issued guidance urging banks to apply judgment rather than mechanically pumping recessionary scenarios into ECL models. Banks applied large "post-model overlays" — top-side adjustments that smoothed the impact and reflected unprecedented government support measures. Auditors were caught between two truths: ECL is meant to be forward-looking, but its outputs in March 2020 felt detached from reality.

The result was striking. U.S. banks took $130B of provisions in Q1 2020 — but then released back $50B+ in 2021 as the world reopened faster than models predicted. Critics argued ECL had over-reacted; supporters argued it had performed exactly as intended.

Why this case matters
COVID showed that forward-looking accounting amplifies short-term volatility. Real risk managers learned that ECL is a tool requiring judgment, not a black box — and that "post-model overlays" became a permanent part of the IFRS 9 toolkit.
March 2023
SVB and the Held-to-Maturity Loophole
$200BBank Failure (Assets)

Silicon Valley Bank's collapse in March 2023 wasn't a credit story — it was an interest-rate story. SVB held large quantities of long-duration U.S. Treasuries that lost market value as rates rose. The losses existed but were not reflected on the balance sheet because the bonds were classified as "held to maturity" and not subject to mark-to-market under U.S. GAAP.

Crucially, these were government bonds — zero PD, zero LGD. ECL was zero. The accounting rules said "no credit loss expected, therefore no provision needed." But the bank still failed, because the economic loss had accumulated even if the accounting loss had not.

Why this case matters
ECL covers credit risk only. Market-value losses on rate-sensitive assets are a different beast. The SVB failure prompted serious discussion about whether ECL should be supplemented with broader fair-value disclosures — a debate that continues today.
The pattern across all four cases: accounting rules evolve reactively. Each major crisis exposes a gap in the existing rule, and the next iteration tries to close it. ECL is the current state of the art — but it will not be the last word. Understanding why the rules exist is more durable than memorizing the rules themselves.
Final step — X
One last stop: a working glossary. Every acronym and term you'll meet in a credit committee or audit query, with plain-English definitions.
A Working Glossary
The lexicon of expected credit loss — the terms that will appear in every credit committee memo, audit query, and Pillar 3 disclosure you will write.
?
How To Use This Glossary
Quick Reference
1
The five terms you must know cold ECL, PD, LGD, EAD, SICR. If you understand these five, you can follow 90% of any IFRS 9 discussion. The rest are refinements and operational details.
2
Pairs to keep straight 12m-ECL vs Lifetime ECL (horizon difference). PIT vs TTC (cyclical vs averaged). Stage 1 vs Stage 2 (the SICR boundary). Each pair is the same concept measured two different ways — knowing why both exist is half the battle.
3
Operational terms vs theoretical terms CCF, DPD, Coverage Ratio appear in everyday risk operations. FLI, Migration Matrix sit one level deeper in methodology documents. Bookmark this tab and return to it as you work through the other tabs.
Expected Credit LossECL
The probability-weighted estimate of credit losses (i.e., the present value of all cash shortfalls) over the expected life of a financial instrument. Replaces the incurred-loss model of IAS 39.
Probability of DefaultPD
The likelihood that a borrower defaults within a given time horizon. Can be point-in-time (PIT) or through-the-cycle (TTC); 12-month or lifetime; conditional or unconditional.
Loss Given DefaultLGD
The fraction of exposure not recovered following default. Equals 1 minus the recovery rate, net of recovery costs and time value of recoveries.
Exposure at DefaultEAD
The expected outstanding amount at the moment of default. For revolving facilities, includes future drawings through the Credit Conversion Factor (CCF).
Effective Interest RateEIR
The rate that exactly discounts estimated future cash payments through the expected life to the gross carrying amount. Used as the discount rate for ECL.
Significant Increase in Credit RiskSICR
The threshold for migrating from Stage 1 (12-month ECL) to Stage 2 (lifetime ECL). Assessed via quantitative, qualitative, and backstop criteria.
Credit Conversion FactorCCF
The percentage of an undrawn commitment expected to be drawn by the time of default. Critical for revolving credit facilities and undrawn loan commitments.
Lifetime ECLLECL
ECL resulting from all possible default events over the expected life of the financial instrument. Applied to Stages 2 and 3.
12-month ECL12m-ECL
The portion of lifetime ECL representing the expected credit losses from default events possible within 12 months from the reporting date. Applied to Stage 1.
Forward-Looking InformationFLI
Macroeconomic forecasts (GDP, unemployment, house prices, etc.) incorporated into ECL estimates to make them unbiased and probability-weighted. Implemented via scenario analysis.
Days Past DueDPD
The number of days that a contractual payment is overdue. 30+ DPD is the rebuttable presumption for SICR (Stage 2); 90+ DPD for default (Stage 3).
Coverage RatioCR
ECL allowance divided by gross carrying amount (or EAD). Headline metric for adequacy of provisions; compared across peers, sectors, and over time.
Through-the-Cycle PDTTC
A PD estimate averaged across economic cycles; relatively stable through time. Used as a benchmark and for Basel regulatory capital.
Point-in-Time PDPIT
A PD estimate reflecting current macroeconomic conditions; cyclical. Preferred for IFRS 9 ECL (with forward-looking adjustment).
Migration MatrixMM
A square matrix giving the probabilities of moving between credit states (ratings or stages) over a fixed horizon. Foundation for term-structure construction.
You've reached the end
From the simple ECL formula to portfolio aggregation, macro overlays, sensitivity analysis, and historical case studies — you now hold the full conceptual machinery of expected credit loss accounting.