Portfolio and model monitoring

Formula / definition

PSI = Σᵢ(current shareᵢ − reference shareᵢ) × ln(current shareᵢ / reference shareᵢ), with disclosed zero-share smoothing.

Interpretation

Compresses changes in a binned score or population distribution; inspect bin contributions and source counts.

Population

Comparable reference and current populations with the same bin definitions.

Horizon

A dated reference window and monitoring interval are explicit.

Assumptions

• Bins, smoothing, weights, and missing-value treatment are fixed or explained.

Useful segmentation

• Product, geography, channel, and score distribution.

Limitations

• Depends on binning and baseline choice; drift does not prove performance decay.

Formula / definition

Apply the PSI contribution formula to reference and current shares for each feature bin.

Interpretation

Highlights which feature distributions moved, so the cause of a population shift can be investigated.

Population

Comparable feature values from the model's reference and current populations.

Horizon

The same reference and monitoring windows used for the related model.

Assumptions

• Feature bins and category mappings are consistent.
• Missingness is a distinct, monitored state.

Useful segmentation

• Feature, product, time, and materially affected segment.

Limitations

• A single index hides direction and can be unstable for sparse categories.

Formula / definition

Compare declared distribution summaries (e.g. quantiles, category shares, missing rate) between reference and current windows.

Interpretation

Separates shifts in source values and pipeline behavior from changes in model performance.

Population

Point-in-time feature snapshots generated under the same definitions and transformations.

Horizon

Reference window, current window, and data arrival lag are stated.

Assumptions

• Feature units, joins, and transformations are stable or versioned.

Useful segmentation

• Feature, source system, product, and time.

Limitations

• Distribution change can be expected seasonality or a pipeline defect; drift alone does not establish impact.

Formula / definition

Missing rate = missing or invalid observations / eligible observations, with null, blank, and sentinel rules disclosed.

Interpretation

Tracks data availability and may reveal upstream outages or population changes.

Population

The eligible rows for each field, with denominator and sentinel mapping recorded.

Horizon

Monitor by processing date and align with source delivery schedules.

Assumptions

• Null, blank, sentinel, and not-applicable values are classified consistently.

Useful segmentation

• Field, source, product, segment, and time.

Limitations

• A stable rate may still hide changed missingness mechanisms or imputation.

Formula / definition

Bad rate = matured positive/default outcomes / eligible units, with exposure weighting stated if used.

Interpretation

Describes realized event frequency for a cohort; distinguish account-weighted from balance-weighted measures.

Population

Eligible originated or active accounts under a stated inclusion and censoring rule.

Horizon

A defined event horizon and maturity cutoff; exclude immature observations or label them censored.

Assumptions

• Default event and denominator grain are consistent.
• Exits and restructures follow a stated policy.

Useful segmentation

• Origination vintage, product, rating, geography, and risk band.

Limitations

• Changes with portfolio mix, macro conditions, policy, and collections.

Formula / definition

Mᵢⱼ = count or share moving from grade i at start to grade j at end, with start-grade row denominators shown.

Interpretation

Shows upgrades, downgrades, stability, and transitions to default or exit over a stated interval.

Population

Accounts rated at the beginning of the transition window with start and end states captured.

Horizon

A consistent transition interval and treatment of exits/defaults.

Assumptions

• Rating scale mappings and observation dates are consistent.
• Missing transitions are reported.

Useful segmentation

• Start grade, product, vintage, and transition period.

Limitations

• Transition rates may reflect policy changes and macro cycles; aggregates may hide segment shifts.

Formula / definition

Roll rate b→c = accounts or balance moving from delinquency bucket b to c / eligible accounts or balance in bucket b.

Interpretation

Tracks curing, remaining delinquent, or worsening between buckets during one transition interval.

Population

Accounts with an observable opening bucket and a defined subsequent snapshot.

Horizon

Specify snapshot interval and treatment of payments, charge-offs, and exits.

Assumptions

• Buckets, balances, and transition handling are fixed.

Useful segmentation

• Opening arrears bucket, product, servicing channel, and time.

Limitations

• A roll rate is affected by collection strategy and payment seasonality.

Formula / definition

Cumulative cohort rate by age t = cohort events observed by age t / eligible cohort units at risk under a stated censoring rule.

Interpretation

Compares equally aged origination cohorts and helps separate seasoning from calendar-time changes.

Population

Cohorts defined by origination or activation period and comparable eligibility rules.

Horizon

Compare like-for-like months-on-book and state right-censoring treatment.

Assumptions

• Cohort membership is stable.
• Event capture and exposure period are aligned.

Useful segmentation

• Origination quarter, product, channel, score band, and age.

Limitations

• Mix, underwriting, servicing, and macro differences can confound comparisons.

Formula / definition

Example: delinquency rate = accounts past a defined days-past-due threshold / eligible accounts; state whether balance weighted.

Interpretation

Tracks payment stress before default and supports collections and portfolio early-warning review.

Population

Eligible active accounts with current contractual due dates and valid payment histories.

Horizon

Snapshot date and days-past-due definition are explicit.

Assumptions

• Grace periods, holidays, restructures, and partial payments are handled consistently.

Useful segmentation

• Days-past-due band, product, vintage, geography, and channel.

Limitations

• Operational timing and payment posting can create temporary spikes.

Formula / definition

Cure rate = defaulted accounts meeting a defined return-to-performing rule / eligible defaulted accounts; re-default uses a later stated window.

Interpretation

Separates temporary cures from durable recoveries and makes the cure policy visible.

Population

Default episodes with complete follow-up and distinct episode identifiers.

Horizon

Define probation period, cure date, and re-default window.

Assumptions

• Cure and default definitions match policy and dataset labels.
• Repeated episodes are not collapsed without a rule.

Useful segmentation

• Default reason, product, workout strategy, time, and collateral type.

Limitations

• Cure classification is policy-sensitive and affected by observation censoring.

Formula / definition

Period hazard = voluntary prepayments during period / accounts at risk at period start, with competing exits defined.

Interpretation

Describes early repayment or account closure that changes exposure and lifetime cash flows.

Population

Accounts eligible to prepay and observed at the start of each period.

Horizon

Specify period length, account age, and treatment of default or transfer as competing events.

Assumptions

• Exit reasons are distinguishable or limitations are disclosed.

Useful segmentation

• Product, rate environment, vintage, term, and customer segment.

Limitations

• Exit coding and contract features can bias observed hazards.

Formula / definition

Recovery rate = discounted eligible recoveries / a stated exposure denominator; also report time-to-recovery and collection cost.

Interpretation

Shows cash collected after default and the timing/cost profile used in workout analysis.

Population

Default episodes with reconciled recovery cash flows and exposure records.

Horizon

From default or charge-off date through a defined workout observation window.

Assumptions

• Recoveries, collateral proceeds, costs, and write-offs are consistently classified.

Useful segmentation

• Secured status, collateral, product, workout channel, and vintage.

Limitations

• Recent defaults may have immature recovery paths; legal and macro processes differ.

Formula / definition

Realized LGD = discounted net loss cash flows / exposure at default under the stated workout or market definition.

Interpretation

Compares realized loss severity with modeled LGD at the same grain and maturity.

Population

Mature default episodes with EAD and complete recovery/cost cash-flow information.

Horizon

State workout horizon, discount rate, cure treatment, and maturity rule.

Assumptions

• Modeled and realized definitions align.
• Recoveries and costs use consistent timing and discounting.

Useful segmentation

• Collateral, seniority, product, default vintage, and workout strategy.

Limitations

• Censored recoveries and small segment event counts create uncertainty.

Formula / definition

For a revolving facility, CCF = (EAD − drawn at reference date) / undrawn at reference date, subject to a declared sample rule.

Interpretation

Compares observed exposure growth before default with modeled CCF/EAD assumptions.

Population

Defaulted facilities with aligned reference balances, limits, and exposure at default.

Horizon

Specify reference date, default date, observation horizon, and facility eligibility.

Assumptions

• Limit changes and restructures are treated consistently.
• Zero-undrawn denominators are excluded or handled explicitly.

Useful segmentation

• Product, commitment type, rating, utilization, and time.

Limitations

• Draw behavior changes with policy, availability, and borrower distress.

Formula / definition

Compare realized cash shortfalls with expected ECL on a matched population and horizon; define any ratio as observed / expected.

Interpretation

Highlights forecast and measurement differences but requires matching scope, scenario, discount, and exposure conventions.

Population

Matched exposures with comparable contractual cash-flow and outcome data.

Horizon

State stage, horizon, observation date, and realization window.

Assumptions

• Observed and expected measures use aligned definitions and discounting.
• Portfolio changes and write-offs are reconciled.

Useful segmentation

• Stage, product, scenario, vintage, rating, and time.

Limitations

• Realized shortfall is noisy; ECL incorporates forward-looking estimates and is not simply a default count.