Executive Summary

• Portfolio: Consisting of c. 3000 facilities across 3 segments comprosing HK Mortgage, HK CRE and HK Manufacturing sectors. Credit Quality of facilities is distributed between IG / NIG with a 10% non-investment grade (NIG) allocation.

• Expected Credit Loss: The ECL is computed scenario weighted using baseline (50%), upside (20%) and downside (30%). Computation of ECL differentiates between 12‑month and lifetime ECL as driven by facility staging through detection of any Significant Increase in Credit Risk (SICR) for facilities.

• Scenario generator: Scenarios are generated for baseline, upside, and downside economic conditions. Scenarios are devised sector specific consistent with historic credit cycle observed for HK sectors. The dependency structure of macro variables is captured through econometric VAR and ADL models. The CRE ECL engine captures non-linearity at the PD layer via regime switch and addresses supervisory critique’s on smoothing skikes in the macro variables. It does not fully capture non-linearity at the cycle-index and LGD layers though.

• Macro variables: Economic drivers are projected over the lifetime of loans, with macro variables used in VAR and Adl models comprising GDP growth, CPI Inflation, Unemployment, Hibor Interest rate, and Property price indices specific for HK Retail and Commercial Real Estate sectors.

• Post Model Adjustments: PMAs are generated with view of impact on ECL, simulated within the engine and broken down by PMA components: Model ECL, Staging, PD, LGD. Applies PMA overlays (PD multiplier, staging override, direct ECL add‑on) from a separate inventory table. The smoothing not adressed in PIT PD and sector cycle layer is remediated through a set of temporary PMAs (PMA-1, PMA-3, PMA-4a, PMA-4b), each of which is disclosed in the PMA inventory.

ECL engine components:

• One Scenario generator with single shock and vector shock transmitted through VAR and ADL model dynamics for projection of baseline, upside, and downside scenarios.

• Three PIT PD models used as satellite models for the portfolio segments HK Mortgage, HK CRE and HK Manufacturing.

• Three PIT LGD models used as satellite models for portfolio segments HK Mortgage, HK CRE, and HK Manufacturing (Downturn LGD as a function of indexed LTV, optionally with a lag toggle).

• SICR identification and Staging engine based on lifetime PD with relative and absolute SICR triggers used for stage 2 lifting at facility level (no qualitative backstops).

• PMA module: Note implemented yet a table of active overlays (type, quantum, expiry) that directly modify at PD/LGD/ECL or scenario weight level.

Workflow and Module Dependencies

Post Model Adjustments (PMA)

• Justification & documentation: Every PMA must be clearly justified documenting (i) the specific model documentation being addressed, (ii) the methodology underpinning the quantification of the PMA, (iii) the evidence supportive of the case for PMA (approach and size).

• Review & governance: PMAs should be subject to independent review before implementation. A PMA should not be used to correct a fundamentally flawed model. If a model is no longer fit for purpose, it must be re‑developed or replaced. PMAs must be governed by a robust framework.

• Monitoring: PMAs should be monitored and back‑tested against actual outcomes. However, permanent overlays that mask weak models and postpone model remediation / replacement are unacceptable.

Graphical Illustration of PMAs:

PMA Inventory

ECL Contribution by PMA

Scenario Components

Scenario Narrative

Each scenario has a qualitative story describes the forecast of macro variables. A baseline narrative would be consistent with the consensus forecast and forward curves. A downside scenario would project a severe but plausible stress.

• Baseline: GDP growth projected is starting at +1.26% (2025 Q1) and rising to +3.65% (2029 Q4). A probability weight of 50% has been attached to the scenario. The economy continues on its current trajectory with monetary policy following the expected path implied by forward curves, with no major shocks occurring and credit conditions remaining stable.
• Upside: GDP expansion is driven by a +1.5 SD structural shock; accompanied by a lower unemployment and stronger property prices. A probability weight of 20% has been attached to this scenario.
• Downside: these scenarios are devised to be sector specific: mortgage is HIBOR-driven, manufacturing and CRE are GDP-driven. Precisely, for the mortgage sector HIBOR spikes to 7.43% and GDP contracts moderately. For the Manufacturing sector GDP contracts to -3.83% and HIBOR declines gradually. For the CRE sector the path chosen is similar to manufacturing.

Scenario Drivers

• GDP Growth (%) p.a. \(\rightarrow\) Corporate, CRE and Retail sectors
• Unemployment rate (%) p.a. \(\rightarrow\) Retail sector
• Hibor (%) p.a. \(\rightarrow\) Corporate sector (via debt‑service capacity)
• HPI Retail (%) p.a. \(\rightarrow\) Retail Mortgage sector
• HPI Office (%) p.a. \(\rightarrow\) CRE sector
• CPI (%) p.a. \(\rightarrow\) Corporate, CRE and Retail sectors
• FX rate (%) p.a. \(\rightarrow\) Corporate sector via foreign currency exposure(not implemented).

Scenario Generation and Expansion

• Time window: c. 5 years of quarterly projections (20 quarters).
• Model: Econometric VAR that forecasts core macro variables viz. GDP, Unemployment, CPI, Hibor.
• Model Satellites: Multiple Econometric ADL that forecast satellite macro variables consistently viz. HPI Retail, HPI Office, Sector Cycle Indices, PIT PD, PIT LGD.
• Baseline is the model’s forecast, following the joint historical dynamics of macro drivers captured in the core econometric model. The projection of satellite macro drivers is obtained from the scenario expansion through estimated fitted ADL models.
• Upside scenarios are generated by shocking the model residuals suitably to generate benign economic paths for key drivers in the core econometric model. Precisely, a Cholesky single shock, i.e. +1.5 SD structural shock to GDP is applied and the shock is transmitted to the other variables.
• Downside scenarios are generated specific to the portfolio sectors Mortgage, CRE, Manufacturing. The shock applied to these sectors is informed by sector-specific historical stress windows with an extreme structural shock being identified and scaled: a multiplier of 1.0 for mortgages, and 1.5 for manufacturing and CRE.
• Impulse-Response: The Impulse-Response Function IRF(h) - displayed for each macro variable in the right-lower corner of the pane - illustrates the dynamic transmission of a single shock, precisely the impact of a 1 standard-deviation structural GDP shock at \(h=0\), i.e. 2025 Q1, and its direct impact on the core variables unemployment, CPI and Hibor, over the quarters \(h=0,..,20\), with the contemporaneous impact of the shock captured at \(h=0\). For instance, a +1 SD structural shock to GDP produces a contemporaneous change in unemployment of -0.1730 units at \(h=0\), that is unemployment falls by 0.1730%. The peak response is at h = 3 at -0.2927% after which the effect gradually mean-reverts.Finally, the orthogonalized version of the IRF is used to allow for the same variable ordering as in the scenario generation.
  • Note: The impulse response function shows the model’s transmission mechanism: how a GDP shock propagates to other variables. The IRF is used to validate the model’s economic coherence, not the scenario itself. For the downside scenario, a negative GDP shock raises unemployment.

Cycle Index Signals and Economic Events

Point-In-Time PD conditional on Credit Cycle

Chart1: Historical PIT PD Paths by Sector. Displays the historical PIT PD paths when computed from sector specific credit cycles. The Through-The-Cycle (TTC) PD is displayed for each sector by a reference line. Shaded bands indicate recession periods.

Chart 2: PIT PD Fitted versus Actual. Actual PD is not directly observable but either a delinquency-based proxy for default rates or an approximation based on the credit cycle index. The Actual PD is hence implied by a Model. Include R², RMSE, and a 45-degree line. Consider using a scatter plot with the historical observations coloured by sector. The lagged cycle index has transformed the model from one that missed the spikes entirely to one that captures their timing, persistence, and most of their amplitude. This is the strongest evidence yet that the cycle index is the missing variable that addresses the smoothing criticism.

Chart 3: Projection of PIT PD for each sector by baseline, upside (benign), and downside (recession) scenarions.

Chart 4: PD Term Structure and SICR. and scenario Important for IFRS 9, but “SICR impact” is not yet defined. SICR is a facility-level staging trigger, not a sector-level aggregate. At the segment level, you can show the proportion of facilities that would migrate to Stage 2 under each scenario, but this requires a staging model. Replace with a PD term structure at the scenario horizon — showing how the PIT PD evolves over the 5-year horizon under base, up, and down. This is the natural link to staging: the divergence between scenarios at year 1 vs year 3 drives the SICR assessment.

• What PIT PD is — point-in-time probability of default, conditional on the current credit cycle position.

• How it is derived — static Vasicek formula using sector-specific credit cycle indices (Z_corporate, Z_cre, Z_mortgage) as inputs, with TTC PD and asset correlation ρ as parameters.

• The sector-specific cycle indices — brief description of how each index is constructed (GDP, HIBOR, HPI-Office, HPI-Retail, unemployment, FX rate combinations).

The validation approach — comparison against delinquency-based default proxies, with R² and RMSE metrics.

• The scenario paths — how base/up/down PIT PD paths are generated from the cycle index scenario paths.

• Key observations — which sector is most cyclical, which is most stable, and what this means for ECL.

• SICR Trigger Zones: two threshold lines define three zones:

  • Below the absolute threshold no SICR (stage 1).
  • Between the lower and upper bound threshold judgement required (stage 2).
  • Abover the uppper bound threshold SICR recognized (stage 3).

For the validation pane in your flexdashboard, you can display these four validations concisely:

Panel 1: Implied Cycle vs. History

A single chart overlaying the VAR downside-implied cycle on the historical cycle, with shaded bands for the worst historical stress episodes.

Panel 2: Severity Metrics Table

A small table comparing the VAR scenario severity to historical benchmarks, with green/amber/red traffic lights for each metric.

Panel 3: Turning Point Accuracy

A simple hit/miss table: for each historical turning point, did the VAR forecast the correct direction of change within two quarters?

Panel 4: PD Sensitivity Check

A scatter plot of VAR-implied PD vs. cycle-implied PD, with a 45-degree line and the correlation coefficient.

IRF sign and shape Do VAR shocks produce cycle responses consistent with history? Ensures the VAR dynamics are economically sensible Severity benchmarking Is the VAR downside as severe as historical stress? Regulators expect scenarios comparable to history Turning point detection Does the VAR capture the timing of recessions and recoveries? Credit losses are driven by transitions, not just levels PD sensitivity Is the VAR → PD pipeline consistent with a direct cycle → PD model? Detects smoothing in the transmission channel

• Macro Scenarios based on fitted VAR You gather 15 to 20 years of quarterly macro data and fit the VAR model. Once the model understands how these variables historically interact, you use it to project the future. Unconditional Forecasts: The model runs freely to generate the Base Scenario. Conditional Forecasts: You inject specific shocks (e.g., forcing a 4% drop in GDP for the first 4 quarters) to generate the Downside Scenario. The VAR mathematically determines how unemployment and interest rates would realistically react to that specific GDP crash.

• Economic Response Models to produce PIT PD and LGD The Satellite Credit Model (The Bridge). Now that you have your distinct multi-year macro paths (Base, Upside, Downside), you pass them into a separate “Satellite” model. This model bridge connects the macro economy to your bank’s internal credit data. It calculates how those specific projected macro pathways will drive up or down the Point-in-Time Probability of Default (PiT PD) and risk stage migrations for your portfolio.

• ERM for linking Macro to PIT PD The PD satellite model is the most heavily scrutinized component of an IFRS 9 framework. Because raw historical default rates are bounded between 0% and 100%, a standard linear regression cannot be used. To map macro factors to a portfolio’s Point-in-Time (PIT) PD, banks apply a logit transformation to the historical default rate. This creates an unbounded index \(Y_{t}\) that can be modeled via logit:

\[Y_t=\frac{DR_t}{(1-DR_t)}=\beta_0+\beta_1 \cdot GDP_{t-1}+...+\epsilon_t\]

• ERM for linking Macro to PIT LGD For secured portfolios (e.g., residential mortgages or commercial real estate), the macro-to-LGD link is usually modeled by directly shocking the underlying asset values using your BVAR projections. For unsecured debts (e.g., credit cards), recovery rates are heavily depressed when unemployment rises because collection agencies struggle to recoup funds. Because LGD is strictly bounded between 0 and 1 with heavy clustering at the edges (0% loss or 100% loss), institutions fit Tobit or Beta regressions. The LGD model is built to explicitly include the portfolio’s contemporaneous default rate (or the systematic risk factor \(Z_{t}\)) as an explanatory variable. This approach captures the PD-LGD correlation that is observed in Downturn LGD.

\[LGD_t=\beta_0+\beta_1 \cdot \text{Collateral}_t+\beta_2 \cdot PD_t\]

Point-in-Time LGD

Chart1: Point-in-time LGD conditional on the credit cycle. Displays historical PIT LGD paths computed from sector specific credit cycles. The Through-The-Cycle (TTC) LGD is displayed by a reference line. The computation of sector LGD is based on a comoving assumption, which for loss and default rate distributions is a useful approximation to capture a rise in losses when default rates rise.

• Mortgage: Low TTC LGD due to strong collateral, low LTVs, limited recourse mortgage law in HK. Loss correlation chosen to match cyclicality.
• CRE: Moderate TTC LGD backed by collateral but subject to a fragile office market thatis prone to cycles. Cyclicality reflected in higher loss correlation.
• Manufacturing: High TTC LGD due to exposure to unsecured lending, LGD affected by trade and finance channels.

Shaded bands in the LGD plot indicate recession periods:

Chart 2: Dependency PIT LGD versus PIT PD. LGD vs contemporaneous PD for each sector. This shows the PD-LGD correlation that drives downturn LGD.

Chart 3: LGD scenario paths for sectors. Each sector’s plot is displayed separately.

What PIT LGD is — point-in-time loss given default, conditional on the credit cycle.

The two channels — collateral channel (property prices drive recovery values) and default correlation channel (when defaults rise, recoveries fall).

• Model specification — Tobit regression for bounded LGD, with sector specific macro variables e.g. HPI-Retail / HPI-Office growth and contemporaneous sector default rate as drivers.

• Validation — comparison against historical LGD data or proxies (recovery rates from HKMA data or industry benchmarks).

• Key observations — which sectors have the highest downturn LGD, and the cyclicality of LGD vs PD.

Optimizing IFRS 9 Frameworks based on Regulatory Heatmap

• Post Model Adjustments (PMA) Supervisors like the HKMA and the PRA expect PMAs to be temporary, evidence-based, and granular — any overuse, persistence without remediation and materiality in top-down portfolio overlays will attract challenge.

• Significant Increase in Credit Risk (SICR) The 30-Days-Past-Due (DPD) backstop used in SICR engines should not be used in isolation, but in conjunction with forward-looking indicators. Moreover, refinancing risk; arising from borrower cliff effect as fixed-rate debt expires and refinances at higher rates; should be modelled explicitly.

• Scrutinized Sectors and Portfolios Portfolios in vulnerable sectors — particularly Commercial Real Estate and SME lending — are subject to targeted review, with granular monitoring of debt-service-coverage ratios and covenant adherence.

• Capture of Macro Spikes in ECL Transmission Models that smooth out macroeconomic spikes mask the non-linear effects of inflation, geopolitical volatility, and rapid interest-rate cycles on defaults. Supervisors increasingly use back-testing and transition-matrix analysis to detect smoothing: ECL that fails to move under a severe scenario is itself a red flag.

• Supervisory Stress Testing Stress‑Testing the ECL is used as a tool in regulatory testing to examine the what if impact on capital. The standard IFRS-9 calculation already incorporates a range of scenarios with weights. However, the framework must be ready to respond to scenarios beyond the standard probability-weighted range.

• Validating the ECL framework Model credibility under stress is demonstrated through technically sound in-house validation. Where stressed ECL diverges materially from base ECL, banks must be able to explain the divergence — this is the reconciliation test that supervisors now apply.

• Transparent Data Lineage Data lineage in IFRS9 computations is expected to be traceable through to the ECL engine to verify timeliness and integrity of data.