E-19 — Risk Measurement

A point-in-time risk result — a VaR figure, an exposure, a sensitivity, a stress-test loss — and the record of how it was produced. The risk analogue of Valuation (E-07): a measured number, stored, with its method and confidence.

Purpose

Risk management produces numbers — value-at-risk, expected shortfall, a factor sensitivity, a counterparty exposure, the loss under a stress scenario. The Risk Measurement entity is the record of one such number, at one date, for one subject.

It is a deliberate modelling choice that risk results are an entity, not only metrics in the semantic layer. OpenIM treats metric definitions — how VaR is defined, how exposure is computed — as semantic-layer concerns (SD-13.8). But a risk result — this VaR figure, for this portfolio, on this date, from this model — is a stored artefact, for the same reason a Valuation is: model-risk governance and regulatory audit require risk numbers to have provenance. “What was our VaR on that date, which model produced it, and how confident were we” must be answerable from a record, not from a recomputation that may no longer reproduce the original. A risk number consumed by a governance decision is a fact that happened; the model stores it.

Like Valuation, Risk Measurement is append-only — a re-run or a restatement is a new row, never an overwrite.

Attribute schema

Notes

• risk_type and measure_type are two orthogonal axes. risk_type is the risk domain — which kind of risk is being measured (market, credit, liquidity, climate …), and which capability is the authoritative source for it. measure_type is the kind of number — how the risk is expressed (a VaR, an expected shortfall, a sensitivity, an exposure figure, a stress loss). The two cross rather than mirror: a single measure_type can appear under more than one risk_type (an exposure figure is produced both as a counterparty-risk measure and within a concentration measure), and a risk_type carries several measure_type values (market risk is expressed as VaR, expected shortfall and sensitivities). risk_type is therefore the discriminator that assigns a measurement to its single producing source; measure_type describes the figure once the domain is known. Neither subsumes the other, so both are kept. The one token that appears on both axes — concentration — is not an ambiguity: risk_type = concentration is the risk domain (the SD-07.4 partition), while measure_type = concentration is the concentration-ratio figure (a top-N or single-issuer share) produced under that domain — the same one-domain-one-figure relationship as risk_type = market with measure_type = var, simply sharing a word.
• The grain — risk_type is in the identity. measurement_id is the surrogate primary key: one stored measurement is one row, identified by measurement_id. risk_type is part of the measurement’s logical identity — it is the key the record set is partitioned on (a market-risk measure and a credit-risk measure are different instances by domain, produced by different authoritative sources), the same shape Holding / Position (E-04) uses with book. So risk_type is a required column that is part of the identity for ownership and partition purposes, while measurement_id remains the single-column surrogate row-handle; every measurement carries exactly one risk_type, and that value names the one source authoritative for it.
• Append-only. The set of Risk Measurements for a subject and measure is its risk trajectory over time. A re-run for a prior date — a corrected or restated number — is a new row; restatement is detected by comparing the new measurement to the prior one for the same as_of_date.
• The distinction from the semantic layer matters: the definition of VaR is a metric definition (SD-13.8); a VaR result is a Risk Measurement. Conflating them loses the provenance that model risk and audit depend on — the same separation OpenIM keeps between a performance metric’s definition and the stored Valuation it is computed from.
• model_id is the hook into Model Governance & AI Governance (SD-14.4): every stored risk number names the model behind it, so a model found to be flawed can be traced to every measurement it produced.
• A Risk Measurement is what a Risk Limit (E-16) is evaluated against; a crossing produces a Limit Breach (E-18).

Out of scope

• The definition of a risk measure — how VaR is defined, how exposure is computed — that is a semantic-layer metric definition (SD-13.8); E-19 is a stored risk result, not a metric definition.
• The configured constraint a measurement is evaluated against — that is E-16 Risk Limit; a crossing produces an E-18 Limit Breach, not a Risk Measurement.
• The scenario a stress result was produced under — that is E-17 Scenario; E-19 references it through scenario_id, but the definition is E-17.
• The value of a holding — that is E-07 Valuation; E-19 is the risk analogue, a measured risk number, not a position value.

Owned and consumed by

• Owned by: key-partitioned by risk_type. SD-07.1 Market Risk Management produces risk_type = market; SD-07.2 Credit & Counterparty Risk Management — credit / counterparty; SD-07.3 Liquidity Risk Management — liquidity; SD-07.4 Concentration & Exposure Risk — concentration; SD-07.6 Scenario Analysis & Stress Testing — scenario / stress; SD-07.8 Climate Risk Analytics — climate. Each measurement Service Domain is the sole authoritative source for instances of its risk type. SD-07.7 Investment Risk Reporting & Limits Governance is the consolidating consumer (limits-breach detection and risk-committee reporting); it is not a producing source. The full pattern is documented in ownership-map.md.
• Consumed by: SD-07.7 Investment Risk Reporting & Limits Governance (limit evaluation, risk reporting), SD-14.4 Model Governance & AI Governance, SD-16.1 Corporate & Fund Governance; SD-05.6 Liquidity-Aware Portfolio Management and SD-11.2 Liquidity Management consume the risk_type = liquidity partition for the per-holding liquidity classification.

Open extensions

• Restatement modelling — how a re-run risk number for a prior date relates to the original.
• The decomposition of an aggregate measure (a total-fund VaR) into its contributions (by portfolio, by factor) — whether each contribution is its own Risk Measurement or a sub-structure.
• The relationship to the R.E.S.T.-style risk-metrics catalogue at the evaluation layer.