Modellability: When Market Prices Become Regulatory Evidence

A risk factor can be economically essential and still lack sufficient observable evidence for a particular internal-model capital treatment. That does not make the risk disappear. It changes how the institution must represent, capitalise and control it. The challenge is to maintain an accurate account of the portfolio while applying evidence requirements consistently.

REFERENCE · NO CALL

This report consolidates the risk-factor eligibility and non-modellable-factor material in a September 2026 teaching series. The thresholds and examples describe the historical framework used there. They are methodological reference material, not a current jurisdiction-specific regulatory determination. All observation counts and capital figures below are illustrative.

Separate economic relevance from regulatory eligibility

Three questions should be recorded for each factor. Does it materially affect valuation? Is it represented in the risk engine? Does its observable-price evidence meet the relevant test for modellability?

These questions can produce different answers. A volatility-skew factor may be material and included in risk-theoretical P&L, but still fail the evidence test for inclusion in the modellable Expected Shortfall set. Omitting it from the risk representation to avoid the evidence problem would instead create an unexplained-P&L problem.

Economic relevance, representation and evidence require separate responses — Economic state: Material factor

Economic relevance, representation and evidence require separate responses — Economic state: Material factor

Economic relevance, representation and evidence require separate responses — Economic state: Material factor

Economic relevance, representation and evidence require separate responses — Economic state: Immaterial candidate

The inventory should make these distinctions explicit. A single yes/no field labelled modelled is not enough to support them.

A price is an evidence object

An observation is more than a numeric value and date. The historical framework differentiates actual transactions, verifiable arm’s-length transactions and committed quotations from indicative marks or model-generated values. A smooth daily surface can contain hundreds of prices without providing hundreds of qualifying market events.

An evidence record should identify the underlying event type, observation date, relevant instrument terms, factor mapping and the means by which the event can be substantiated. Data supplied through an intermediary still needs the evidence required by the applicable framework; a vendor label alone does not prove eligibility.

The public research conclusion is about the process. It does not require disclosing counterparties, private transaction identifiers or a reverse map from anonymised evidence to named institutions. Those fields belong in controlled internal records.

De-duplicate observations before counting

The same market event may arrive through several channels. It may be present in a transaction feed, a reconciliation file and a third-party package. Counting each appearance independently inflates the evidence without adding information.

The teaching framework permits no more than one qualifying observation per day for the relevant factor. A source example assembles 112 controlled observation days after qualification and de-duplication. Its significance is that eligibility is determined from the controlled result, not the raw number of vendor rows.

The process should also distinguish a corrected observation from a new one. When an original record is replaced, both the history and the current valid state should remain traceable. A retrospective data repair can change a historical count, but the change needs an explanation and an effective date.

Count and continuity are different tests

The historical material describes two alternative routes: at least 24 qualifying observations with a minimum of four in every relevant 90-day period, or at least 100 qualifying observations over the relevant twelve months. The exact period definitions and application belong to the governing rule version.

The following retained examples show why the alternatives cannot be collapsed into one annual count:

Controlled observation counts and continuity outcomes — Controlled evidence: 31 observations, one 90-day period with three

Controlled observation counts and continuity outcomes — Controlled evidence: 112 qualifying observation days

Controlled observation counts and continuity outcomes — Controlled evidence: Many raw records on a few dates

Monthly monitoring can warn of an approaching failure even when the formal classification is quarterly. A factor may lose eligibility because older observations roll out of the window, not because the latest market feed failed. The control should therefore forecast the expiry of evidence as well as track new arrivals.

Bucketing changes both evidence and fidelity

A volatility surface divided into a five-by-five grid has 25 buckets. A two-by-two grid has four. With the same qualifying observations, the coarser representation can concentrate evidence into fewer tests and make more buckets pass.

That improvement has a cost. The risk model now distinguishes fewer sources of movement. If the portfolio depends on strike-specific skew or tenor-specific variation, coarsening can increase the gap between full valuation and risk-theoretical P&L.

The economic choice is therefore not to maximise the number of passing buckets. It is to select a representation that is accurate enough for the portfolio, supported by evidence and assessed under the corresponding capital treatment. Eligibility and P&L attribution should be examined together on the same factor design.

A change in bucketing should have a documented counterfactual: what happens to unexplained P&L, capital, hedging interpretation and observation coverage under the old and new grids? A net reduction in one reported charge does not establish that the revised model is better.

Non-modellable treatment is a capital process

In the historical framework discussed by the source, non-modellable factors receive stress-based treatment with a specified confidence standard and a liquidity-horizon floor. The source describes a minimum 20-day horizon for that treatment, even where the corresponding modelled factor had a ten-day assigned horizon.

The stress calibration needs its own evidence. A lack of ordinary price observations makes this harder, not optional. Plausible proxies, conservative assumptions and the treatment prescribed for insufficient scenarios should be documented without relabelling model marks as observed transactions.

The factor inventory should connect the economic exposure to its stress charge. Otherwise a capital total can remain stable while the underlying factor coverage has become stale. Reconciliation is needed between valuation factors, the eligibility inventory, the modellable set, non-modellable charges and the P&L-attribution calculation.

Aggregation limits diversification

The source’s non-idiosyncratic aggregation example uses four separate charges of ₹1 crore and a correlation parameter of 0.6. A simple equal-correlation construction is

$$C=\sqrt{\sum_i c_i^2+2\rho\sum_{i<j}c_ic_j},$$

where $c_i$ are the individual charges and $\rho$ the prescribed illustrative correlation. With four equal charges,

$$C=\sqrt{4+2(0.6)(6)}=\sqrt{11.2}\approx₹3.35\text{ crore}.$$

At zero correlation the same arithmetic gives ₹2 crore; at perfect correlation it gives ₹4 crore. The example shows how a restricted diversification assumption changes the aggregate. It does not authorise using the simple formula for every class of non-modellable factor, since idiosyncratic groups and detailed rules can differ.

The main desk example separately uses ₹1.08 crore as a non-modellable charge. That number is an assumed teaching input in the capital exercise, not a charge reconstructed here from a complete evidence inventory.

An evidence lifecycle needs named responsibilities, not public names

The operational roles are straightforward even when the organisation is anonymised.

Responsibilities across the observation-evidence lifecycle

Each role needs a controlled handoff. A data team should not change bucket definitions solely to improve counts without model approval. A modeller should not treat an indicative mark as a qualifying quote to solve a calibration problem. A capital team should not assume that a missing factor has no charge because it is absent from the submitted inventory.

Exceptions should retain both the observed state and any approved override. A temporary judgement needs an expiry or review trigger. Otherwise a one-off treatment can become a permanent undocumented rule.

Monitor deterioration before the formal test

A practical dashboard should show distinct qualifying days, observations approaching expiry, continuity minima, evidence failures, mapping changes and economically material factors without adequate support. Concentration by evidence channel also matters: a factor depending entirely on one fragile source can pass today and fail after an operational interruption.

The dashboard should distinguish leading indicators from formal classifications. A projected shortfall is a prompt to investigate, not a declaration that the rule has already been breached. Conversely, a passing annual count should not obscure a failed continuity condition.

The decision is about complete representation

The strongest outcome is a factor design whose economic purpose, observation evidence and capital treatment can all be reproduced. Some factors will remain non-modellable, and their continued presence may be the correct result for a faithful model.

The objective is not to make the inventory uniformly favourable. It is to ensure that every material source of value change is represented, that evidence is counted honestly and that the cost of uncertainty is visible in the relevant control and capital process.