Facility Share Allocation¶

How the calculator attributes the undrawn commitment of a facility that several counterparties may draw against to one of them.

Regulatory reference: none for the allocation itself — see Regulatory status. The exposure value it produces cites CRR Art. 111 and Annex I, CRR Art. 166(8)(d); PRA PS1/26 Art. 111 Table A1 and Art. 166C. Design of record: Facility Share — Riskiest Member

What a Facility Share is¶

A Facility Share is one facility whose headroom more than one obligor may draw. The calculator emits a synthetic facility_undrawn exposure row for a facility's unused limit, and that row needs a counterparty: the risk weight, the approach, the model permission, the exposure class and any credit risk mitigation all follow from whose exposure it is. Where several obligors may draw the same line, "whose exposure it is" has no single answer in the input data, so the calculator has to elect one.

The policy it applies is capitalise the commitment against the riskiest member that could draw it, and "riskiest" is measured on each member's own applied approach — the internal-ratings formula for a member on a model, the standardised tables for a member without one — including that member's own credit risk mitigation.

Detection¶

There is no input flag for a share. A facility row carries exactly one counterparty_reference, its owner, and the facility reference is a unique key. Loans, contingents and sub-facilities attach to a facility through facility_mappings, and each child carries its own counterparty reference. That is the only route by which a second counterparty can enter.

The member set of a facility is therefore

• the facility's own counterparty_reference — the owner is always a member, because the owner is the legal borrower and can draw the whole line; unioned with
• the distinct counterparty_reference values on its descendant loans and contingents, each resolved up to its root facility.

A facility whose member set holds more than one member is a Facility Share.

Nothing is drawn-weighted. A member with a small drawn balance and a member with a large one are equal members, because either may draw the whole headroom.

Multiple Option Facility (MOF) parents are excluded. Their per-sub waterfall rows already carry each sub-facility's own counterparty, so the allocation question does not arise, and fanning them out would multiply the waterfall. The residual row that carries the parent's own risk type is not a candidate either.

Implementation: engine/hierarchy/facility_undrawn.py (_derive_facility_share_members, _apply_facility_share_fanout), described in Component Overview — Hierarchy Resolver.

Regulatory status¶

The allocation rule is firm policy grounded in conservatism, not regulation. Neither UK CRR nor PRA PS1/26 defines a facility share, and neither prescribes how to attribute a commitment that several obligors may draw. The conversion factor articles define the undrawn exposure value for a commitment with one obligor:

So the allocation carries no @cites of its own; the exposure value it selects cites the articles above. The values sit in the rulepack and are rendered, with their citations and for both regimes, in Regulatory Tables — this page does not restate them.

Details: the conversion-factor mechanics themselves live in Credit Conversion Factors.

Mechanism — compute, then choose¶

The comparison the policy asks for cannot be made where the facility rows are built. At the hierarchy stage the approach, the exposure class, the credit risk mitigation and the output-floor state are all still undecided; the only point at which every member's own-approach RWA and its standardised-equivalent RWA both exist is the exit of the calculators. The engine therefore computes first and chooses afterwards.

flowchart LR
    H["Hierarchy<br/>fan-out"] --> C["Classifier"]
    C --> M["CRM"]
    M --> K["SA / IRB / Slotting<br/>calculators"]
    K --> R["Aggregator head<br/>resolve_facility_shares"]
    R --> F["Output floor"]

1. Fan-out (hierarchy). A share's single undrawn row is replicated into one candidate row per member. Every candidate carries the full headroom — each is "as if this member drew the whole line", so nothing is pro-rated. A candidate is marked by:

2. Pricing (classifier, CRM, calculators). No pricing code knows about candidates. Each is classified, CRM-adjusted and priced as an ordinary row of its own member — its own exposure class, model permission, PD and LGD, its own collateral and guarantees — and, under Basel 3.1, its own standardised-equivalent RWA, because the standardised pipe runs on every row to build S-TREA.

3. Resolution (aggregator head). resolve_facility_shares keeps exactly one candidate per facility_share_group, drops the losers from the combined frame and from each of the three branch frames, and renames the winner back to <facility>_UNDRAWN. The aggregator's exit invariant is unchanged: one undrawn row per facility, so COREP, Pillar 3, reconciliation and the supervisory register see the shape they saw before.

There is no second implementation of any pricing logic anywhere in this design — that is the point of it. The mechanism it replaced was a standardised-equivalent risk-weight preview built at the hierarchy stage from entity type, credit quality step and country. That preview knew nothing about whether a member was on a model, about its PD or LGD, about its exposure class, about credit risk mitigation on the facility, or about the floor, and it was deleted on 2026-09-05.

The metric under CRR¶

Under CRR capital is additive, so the capital-maximising choice is made group by group and no portfolio state enters:

winner(g) = argmax over members i of g:  own-approach pre-floor RWA of i

The metric is RWA, not risk weight. The two orderings coincide whenever the exposure value is the same for every member, which is the common case; they diverge wherever the conversion factor or the credit risk mitigation depends on the obligor, and there the conservative quantity is capital rather than a gross weight. The own-approach RWA is taken after the SME and infrastructure supporting factors, so a member's relief is honoured rather than ranked away.

The metric under Basel 3.1 — the output floor¶

Under Basel 3.1 total risk exposure is a portfolio-level max, so "riskiest" is state-dependent: which member costs the most capital depends on whether the floor binds, and whether the floor binds depends in part on which member was chosen.

The engine's floor formula¶

The engine takes the max over the floor-eligible rows and adds the standardised book outside it:

TREA = SA_total + EQ + max( U_elig , x . S_elig + OF-ADJ )

where U_elig and S_elig are summed over floor-eligible approaches (internal ratings, slotting and counterparty-credit-risk-via-standardised), x is the Art. 92(5) phase-in percentage from the pack Schedule output_floor_pct (or output_floor_pct_full where the firm skips the transition), and OF-ADJ is the own-funds adjustment.

The floored-branch marginal¶

For a member i, let u_i be its own-approach pre-floor RWA and s_i its standardised-equivalent. Its marginal contribution to the floored branch is

b_i = x . s_i      if the member's applied approach is floor-eligible
      u_i          otherwise

The second limb is the one an implementation gets wrong. Floor-eligible and standardised do not partition the domain: counterparty-credit-risk-via- standardised and slotting are floor-eligible without being standardised, so a predicate written as "not internal-ratings gets u_i" sends both down the wrong limb and understates the floored assignment. A specialised-lending facility's undrawn candidate is a live slotting case.

The rule: evaluate two assignments, keep the larger¶

The un-floored branch is additive across shares, so it is maximised exactly by assignment A — every share to its argmax u_i. The floored branch is strictly increasing in S_elig, so it is maximised share by share by assignment B — every share to its argmax b_i. Since total risk exposure is the max of the two branches, the capital-maximising assignment is one of those two. The engine evaluates both end to end — recomputing the expected-loss summary, OF-ADJ and the floor on each assignment's surviving book, through the aggregator's own functions rather than a second copy of them — and keeps the assignment with the larger total. Ties go to A, for attribution stability.

When the floor state cannot change the answer¶

A useful boundary falls straight out of the two marginals. A standardised member's marginal is its full RWA under both assignments, while a modelled member's falls to x times its standardised-equivalent under assignment B. So:

No facility share whose standardised members all sit at or above the floor percentage — that is, whose own risk weight is at least x times the standardised-equivalent risk weight of every modelled member — can ever flip on the floor state.

A regulatory-retail member is the everyday case: its full risk weight already exceeds x times an unrated corporate's standardised-equivalent at every step of the output_floor_pct Schedule, so assignment A and assignment B agree and the election below changes nothing. Shares that can flip are the ones holding a low-PD modelled member with a much higher standardised-equivalent weight, against standardised members priced below the floor percentage.

The facility_share_metric election¶

CalculationConfig.facility_share_metric is a firm election, not a regime switch — the regime gate is the existing output_floor pack Feature.

Under CRR the election is accepted and inert: the output_floor Feature is off, so the un-floored rule applies either way. There is no OutputFloorSummary under CRR, so the only place the election is observable there is the metric_used column of the audit frame.

Candidates in obligor aggregates¶

Candidates are real rows from the classifier onwards, so the per-obligor windows see them. The design decision is no window special-casing anywhere: every candidate counts toward its own member's aggregates exactly as any exposure of that member would. Each member's lending-group and obligor totals therefore include the full undrawn, which is the conservative reading of "total amount owed" for a line that member may draw in full.

Direction, relative to the single-row behaviour it replaced:

• The eventual winner's siblings are unchanged. The single undrawn row already counted for them.
• A losing member's siblings gain the undrawn in their partition-local totals, so the retail and qualifying-revolving thresholds can only be crossed upward. Crossing upward moves an obligor to a less favourable treatment.
• The Art. 123A(1)(b)(ii) granularity denominator does not move at all. The retail-threshold carrier is built from the drawn amount only, so an undrawn row contributes zero to it; the denominator divides each obligor's aggregate by that obligor's own line count, so extra rows of the same obligor and standardised class leave its term algebraically unchanged. The invariance is pinned by a test, not assumed.

One residual, disclosed rather than fixed. A member demoted out of the qualifying revolving retail class by its enlarged obligor aggregate is priced on the ordinary retail correlation, which decays with PD, instead of the fixed qualifying-revolving correlation. Above a crossover PD the decaying correlation falls below the fixed one, so for such an obligor the demotion lowers RWA. The effect is partition-local — it touches only that member — and the crossover sits well above any ordinary revolving portfolio's PD; the correlations are in engine/irb/formulas.py. No fixture reaches it, so it is recorded as owed coverage: a share with a revolving retail member near the qualifying limit and a PD above the crossover.

The rejected alternative was an "own-inclusive" rule in which non-candidate rows see no candidate. It is exact for losers but drops the undrawn from the winner's siblings, which is RWA-reducing relative to the behaviour it would replace.

Tie-breaks and the fallback¶

Ordering within a share, for the own-approach ranking and for assignment A:

1. own-approach RWA, descending;
2. risk_weight, descending;
3. pd_floored, then cqs, descending — the worse credit;
4. counterparty_reference, ascending.

Assignment B is the same chain with b_i in the first rung. Nulls sort last throughout. Rungs 3 reach only within an approach — pd_floored is populated on modelled rows and cqs on externally rated ones — so on a mixed-approach share the final rung is what guarantees a total order. It always does.

Where every candidate of a share carries a non-finite or absent own-approach RWA, the ranking metric decides nothing. The group is never dropped — dropping every candidate would delete the facility's undrawn commitment from the submission outright. The engine falls back to risk_weight descending, then counterparty_reference ascending, records metric_used = "fallback_deterministic" with a null rank on every candidate, and emits an AGG003 warning naming the group and the candidate count. A fallback nobody is told about is indistinguishable from a ranked outcome, which is why it is a warning and not a log line.

What is exposed for audit¶

AggregatedResultBundle.facility_share_resolution — one row per priced candidate, in both regimes:

Because every member is priced, the frame is a per-facility allocation sensitivity table: the RWA the commitment would have carried under each member, on each metric. That is directly useful in reconciliation against a legacy system whose own share rule is unknown.

OutputFloorSummary, on the Basel 3.1 runs, gains two fields:

• facility_share_metric_used — which assignment decided the book.
• facility_share_trea_alternative — the total the other assignment came to.

Attribution flipping with the floor state is a designed consequence of the floor-aware default, but it may never be silent. These two fields are what the flip is read against; it should never be inferred from a moved COREP row.

Worked example¶

One shared facility with three members: the owner, a standardised corporate; a second standardised corporate at a better credit quality step; and a foundation-internal-ratings corporate with a low internal PD and a high standardised-equivalent weight. Two variants of the same portfolio differ in one input, the PD of an unrelated anchor loan, which is what makes the floor bind in one and not the other.

Reading the four rows:

• Under CRR the owner's full RWA beats the modelled member's, which is much lower than its standardised-equivalent. Capital is additive, so that is the end of it.
• Under a binding floor the modelled member's marginal becomes x times its standardised-equivalent, which here exceeds the owner's full RWA, so assignment B wins and attribution moves to the modelled member. Note that the undrawn row's own rwa_final is lower while the portfolio total is higher — the member's own-approach RWA is what sits on the row, and its contribution to capital arrives through the floor.
• Under the election, the same binding book is pinned to assignment A. Total risk exposure falls by the difference between the two floored branches. This is the direction the election warning above describes.
• Where the floor does not bind, both assignments agree and the election is inert.

Known findings¶

Recorded here because they bound what this page claims, and filed for separate decisions. None is fixed by this design.

1. The engine floors the modelled subset, not the firm. The floor max is taken over floor-eligible rows with the standardised book added outside it, where Art. 92(2A) takes it over whole-firm totals. The engine's form is conservative; correcting it would be RWA-reducing and needs its own evidence and sign-off.
2. The Art. 110A due-diligence input never reaches the standardised branch. The two input columns are accepted by the loan and contingent schemas but are not declared on the hierarchy edge contract, so they are dropped silently before classification and the SA004 warning fires on every Basel 3.1 run regardless of input.
3. The binding-floor add-on has no reported home in the internal-ratings templates. The pro-rata floor add-on lives inside rwa_final and the reporting projection mirrors it into the after-adjustments column, which then exceeds the sum of its stated components by exactly the floor shortfall. The fix is a reporting-basis decision, not a calculation one.

Related pages¶

• Facility Share — Riskiest Member — the design of record, with the five decisions and the rejected alternatives
• Component Overview — where the fan-out and the resolution sit in the pipeline
• Output Floor — the floor mechanics the two-assignment rule depends on
• Credit Conversion Factors — the undrawn exposure value
• Regulatory Tables — every pack entry named on this page, rendered with its citation for both regimes