Expected Loss Back-Testing Design Director

The mandate

This project will design and prove a back-testing framework for expected loss estimates that separates discrimination, calibration, timing, staging and overlay effects. The Director must turn realised outcomes into actionable model and accounting evidence while acknowledging censoring and long emergence periods. The commission does not rebuild models, calculate reported provisions or issue independent validation.

Named deliverables are an outcome and cohort definition paper, observation-window protocol, reconciliation design, metric library, decomposition method, overlay-performance review, significance and action thresholds, reporting template, pilot test and owner handbook. The framework must prevent apparent offsets among materially different populations from creating false comfort.

Vintage and cohort comparisons must preserve the information available at each original estimation date. Later corrections may inform learning, but cannot be backfilled into history in a way that flatters prior predictive performance.

Five milestones apply: scope and outcome definitions by 6 November 2026; data and censoring assessment by 18 December; metric and decomposition design by 12 February 2027; pilot and defect retest by 26 March; unseen internal run and acceptance by 30 April. A Model Performance Committee accepts milestones after accounting, modelling and validation representatives document challenge.

Final acceptance requires internal analysts to run the method on an unseen cohort, identify seeded cure, recovery and censoring errors, reconcile predicted to realised loss by key driver and recommend the correct governance route. Critical definition, lineage or aggregation defects block sign-off.

Management provides approved exposure histories, prediction snapshots, status events, recoveries, overlay records, existing monitoring and timely decisions. Consulting owns framework, pilot and transfer. Management owns models, estimates, provisions and remediation decisions. Redevelopment, production calculations, platform build, independent validation and recurring operation are excluded.

What you will own

• Define outcomes, cohorts, observation windows and snapshot rules appropriate to expected loss horizons.
• Address right censoring, cure, recovery timing, write-off, exposure change and data revision explicitly.
• Separate rank ordering, probability calibration, loss severity, exposure, staging and scenario effects.
• Attribute overlay performance to the stated deficiency and test whether exit conditions were met.
• Prevent aggregate offset by segmenting material populations and reconciling them to total reported movement.
• Set thresholds linking evidence to monitoring, recalibration, redevelopment, overlay or use restriction.
• Pilot the method, seed event-history faults and verify correction through an unseen internal run.
• Apply change control to redevelopment, validation, technology and recurring-production requests.

Candidate qualifications

• Demonstrate expected-loss back-testing design across long outcomes, censoring and changing exposures.
• Describe aggregate performance that concealed offsetting calibration weakness in distinct populations.
• Show how cure, recovery or write-off timing altered an apparent forecast error.
• Evidence decomposition that separated model, staging, scenario and management-overlay effects.
• Provide an action framework connecting statistical evidence to specific governance decisions.
• Explain how internal analysts proved reproducibility on an unseen cohort.
• Show project discipline when redevelopment or validation requests emerged.

Working terms and boundaries

• The project fee covers seven months, four days weekly and five milestone packages.
• The Model Performance Committee accepts after unseen-cohort execution and closure of critical defects.
• Management owns models, estimates, entries and decisions; consulting owns named back-testing artifacts.
• Data and decision delay enter formal dependency management before dates change.
• Redevelopment, production provision, platform build, independent validation and recurring operation are excluded.

Application

Applications for this mandate are received in one way only: through the India Board Terminal's application process. It is automated end to end. Your Executive Passport travels to the mandate holder in its confidential form, your answers to the three questions below are read before anything else in your file, and every stage that follows is recorded on your applications page.

There is no address to write to and no intermediary to call. The mandate holder reads what the Terminal delivers and nothing else, which is what keeps the process the same for every applicant and keeps your name out of it until you release it. Applications close on 8 October 2026. Mandate reference QNT-CON-2026-LIS-28.