Capital-Markets Surveillance Model Validation Director

The mandate

The operator has introduced a learned ranking layer over existing market-surveillance scenarios, but validation cannot reproduce several claimed improvements because labels include later investigation outcomes and alert populations changed during testing. The defined problem is to determine whether the model improves risk-based prioritisation without hiding missed typologies or distorting investigator attention.

The deliverable is an Independent Surveillance AI Validation Opinion containing model and data lineage, conceptual-soundness challenge, temporal partition tests, typology and market-regime coverage, leakage analysis, calibration, stability, investigator-use evidence, alert-capacity effects, limitations, monitoring triggers and a signed residual-risk assessment. It does not decide whether an observed pattern constitutes misconduct.

Work begins 11 January 2027. Milestone one on 22 January is the scope, data-lineage map and critical issue notice; milestone two on 12 February provides independent reruns, temporal and capacity tests and investigator findings; milestone three on 26 February delivers the signed opinion, committee paper, monitoring specification and validation evidence archive.

Acceptance requires model risk to reproduce independent results from retained code, data cut-offs, labels and environments, and surveillance leadership to trace sampled rank changes to investigator action without hindsight contamination. Material typologies must meet signed detection and prioritisation tolerances across relevant regimes; aggregate precision improvement cannot offset an unexplained severe miss.

The client will provide model artefacts, feature code, order and trade histories, scenario alerts, investigation labels, deployment and override logs, investigator access and secure compute. Legal and surveillance officers retain interpretation of conduct and escalation; the consultant will not investigate participants, change live thresholds, submit reports or remediate production code.

Why this is external work

Model developers own the performance claim, while investigators benefit from lower queues and internal validators lack recent temporal graph and ranking depth. The oversight committee needs a view independent of both productivity pressure and legacy scenario ownership. External validation supplies bounded technical evidence without replacing surveillance judgement or regulatory accountability.

What you will own

• Reconstruct model lineage across market data, entity links, features, labels, scenario alerts, investigation outcomes, overrides and release versions.
• Test temporal partitions and label availability to expose future information, post-alert investigation and changing participant identity leakage.
• Evaluate ranking and calibration across manipulation typology, instrument, liquidity, volatility, market event and participant structure.
• Quantify alert-capacity consequences through investigator workload, displacement of severe cases, time to review and reasons for override.
• Challenge stability under regime shifts, missing data, adversarial adaptation, entity-link error and changes to upstream scenarios.
• Define monitoring triggers for feature drift, rank concentration, typology coverage, calibration, override and investigator queue effects.
• Deliver an independent opinion with supported limits, required controls, unresolved uncertainty and conditions for continued use.

Candidate qualifications

• Independently validated machine-learning models used for market surveillance, financial crime, fraud or similarly consequential alert prioritisation.
• Detected temporal or label leakage arising from investigation outcomes, future events or changing entity resolution.
• Evaluated ranked alert systems under constrained investigator capacity and measured displacement, override and severe-case consequence.
• Challenged conceptual soundness across graph, behavioural or sequence features used to infer coordinated market activity.
• Presented a no-use, conditional-use or limited-use conclusion to an independent model or risk committee.
• Produced reproducible validation evidence while preserving confidential market, participant and investigation information.

Non-negotiables

• The named director must attend both Hong Kong investigator workshops and the regional market-data control review within seven weeks.
• Has no undisclosed employment, investment or paid relationship with model vendors, trading participants or investigators material to validation.
• Will not treat a model score as a misconduct conclusion or allow aggregate precision to conceal a severe typology gap.
• All order, trade, participant, alert and investigation data must remain inside the operator’s approved environment.

1. 49 words maximum. Describe one surveillance or financial-crime model whose result changed after you removed temporal leakage.
2. 49 words maximum. How would you test whether better ranking displaces severe alerts under a fixed investigator queue?
3. 49 words maximum. Confirm seven-week capacity, Hong Kong attendance and all relevant market, vendor or participant relationships.