Supervisors ask how a figure was produced. So do auditors, validators and your own management body. This is how Halden answers, for data matching, validation, scenario design, model development and governance.
Download the methodology paper (PDF)Every value in the template comes from a mapping rule: a source range, an optional transform and a target area. You write rules by pointing and clicking, or ask the optional AI assistant to draft them. Either way, a rule is explained, checked and approved before it fills anything.
You pick a source range and a template area. Only that evidence is used, and the app shows it before anything is sent.
Rules are written by hand or drafted by the AI assistant against the template's row and column identifiers.
Each draft carries its reasoning and passes the same technical checks as a manual rule: target exists, no calculated cells, no overlaps.
A checker approves the rule set; approved rules are deterministic and versioned. AI plays no part in the fill itself.
| Check | What it catches | When | Status |
|---|---|---|---|
| Rule checks | Targets outside the template, writes to calculated or formula cells, two rules writing the same cell | When a rule is saved | Available |
| Template data validations | Values outside a cell's dropdown list, with a suggested correction | Before export | Available |
| Template-aware tab checks | Sign conventions, percentages entered as whole numbers, zeros where blank selects a different methodology, empty selectors | Before export | Available |
| Calculated totals | Totals recomputed from the template's own formulas, so you see what the supervisor will see | In review | Available |
| Four-eyes approval | A second person approves the exact rule-set version; any change needs new approval | Before export | Available |
| EBA validation rules | Breaches of the published EBA validation rule set | Before export | Planned |
| Plausibility & reconciliation | Movements against the prior cycle; differences against FINREP, COREP and the ledger | Before review | Planned |
| Stage | Method in Scenario Creator 0.1.0 |
|---|---|
| Data | Series from the ECB Data Portal, Eurostat and CBS StatLine, downloaded from your computer, or your own CSV. Stored with source address, retrieval time and SHA-256. Annual paths; incomplete years are left out. |
| Baseline | An AR(1) per variable, x(t) = μ + ρ (x(t−1) − μ), estimated by least squares on the history window, with ρ limited to 0–0.95. Alternatives: convergence to the mean, flat, or an entered path. |
| Shocks | Deviations from the baseline that rise to a peak year and then decay by 0.35, 0.7 or 1.0 a year. Templates size them in each series’ historical σ; severities scale them by 0.5, 0.75 or 1.0. |
| Propagation | A VAR(1) on standardised annual data with ridge shrinkage (λ = 0.1 × T). Shocked variables are imposed; the others take their conditional expectation. Linear, and can be switched off per scenario. |
| Climate | NGFS Phase 5 GDP impacts relative to Current Policies, converted into a growth deviation for the matching country. The carbon price is a fixed path. |
| Checks | Blocking: missing values, impossible values, sample data at submission. Review: stale data, short history, an adverse path milder than the baseline, GDP down without unemployment rising, no narrative. |
| Approval | A content hash covers data, methods, shocks, paths and narratives. With a team server, a checker who is not the maker approves that exact hash. |
The checks support a reviewer; they do not certify that a scenario is severe but plausible. The methods are documented approximations and do not replace expert judgement or supervisory guidance. Scenario Creator →
In Model Studio, agents carry out the work between gates. Each gate is a decision by a named person, recorded with the evidence it was based on.
Availability, depth and quality
Gate · methodology approval
Selection, segmentation, calibration
Performance, stability, back-tests
Gate · owner sign-off
Gate · independent review
Ongoing performance tracking · Planned
The EBA Filler and Scenario Creator separate maker and checker. Model Studio adds a model owner and an independent validator, who open its gates. No one approves their own work.
Every rule set is versioned and hashed; an approval is bound to that exact version. Re-run any approved version to reproduce its export.
From any reported value back to the rule, the approver and the source row.
Scorecards and rating models, through-the-cycle and point-in-time.
Model details →Workout and cure-based LGD, including downturn adjustments.
Model details →Credit conversion factors for revolving and off-balance exposures.
Model details →Lifetime PD term structures, staging and macro scenarios.
Model details →Macro-to-risk links for stress testing and climate scenarios.
Model details →Prepayment and non-maturing deposits for IRRBB and NII.
Model details →How rules are built, checked and approved, what the template checks cover, how lineage and the audit trail work, and what is not yet covered. No form, no gate.
We'll walk through your use case and show the product on data structured like yours.