Import the loan book. Measure every facility month by month. Show the evidence behind each assumption, prove the figure can be re-derived from its own inputs, and produce the disclosure note — all in one place, and all from the same numbers.
The demonstration book is synthetic and carries no customer data. It exists so the platform can be exercised at the scale it will meet in use, and so that any claim made about speed or capacity can be checked rather than taken.
CSV or Excel, read in the browser. Columns are matched by name, so most extracts load without preparation. Fourteen rules run before anything is measured, and three of them stop the import outright.
Lead sheets tie the book to the ledger, split it by stage, and reference the schedule supporting every line — the first thing an auditor asks for.
Every input is editable and sits beside the output it drives. Change a threshold and the book is re-measured; each change makes a new assumption version.
Fit the economic relationship with significance, collinearity, lag and stability tests. Develop recoveries to ultimate on a run-off triangle. Test whether a grouping separates risk at all.
Each facility carries its own schedule of default probability, exposure, loss rate and discount factor, under every scenario. The reported figure is the sum of those schedules.
Sixteen checks, each written as a plain question and answered with the numbers that decided it. Then the whole book is re-measured from its own stored inputs and compared to the last digit.
Three different people, three roles, and a proven reproduction. A signature captures the figure and the book fingerprint as they stood, so it cannot later be said to have signed something else.
The note, on a white sheet, generated from the measured book. Where a requirement cannot be met from the data held, it says so rather than dropping the paragraph.
The methodology lives in Python. Measurement, staging, the statistical calibrations and the working-paper generator are written in layers that keep the arithmetic separable from everything around it: immutable value objects, a measurement engine that is a pure function of its arguments and can reach no database or clock, an analytics layer for estimation, and append-only snapshots with versioned assumption sets. That separation is what makes a figure re-derivable — the engine, given the same inputs, has nothing else it could consult.
The result is published as local HTML. This workspace is that publication: a static page with no server, no build step and no network call once loaded. It runs from a network share, an internal web server or a laptop with the network unplugged, and the book it reads never leaves the machine.
It integrates with SAP. Export and load; or let a scheduled job drop the extract for the Python core to measure and publish; or have the core read loan, customer, collateral and recovery data directly and write the measured allowance back by facility. The route changes how the data arrives, never what is done with it.
Every estimate arrives with the diagnostics that say whether it should be used: significance, variance inflation, serial correlation, split-sample stability, and whether the fitted sign agrees with economic sense. An estimate without a diagnostic is an assertion.
The book is measured again from the stored inputs alone and compared facility by facility to the last digit. A figure that cannot be re-derived from its own inputs is not allowed to become the reported figure.
Where the data cannot support a requirement, the note says so instead of omitting the paragraph. An omitted disclosure reads as not applicable; a stated limitation reads as what it is.
Every sensitivity is the whole book measured again under that change. The measurement is not linear — exposure caps, collateral caps and stage-dependent horizons all bind — so scaling the reported figure would be wrong exactly where it matters.
The month-by-month schedule behind any facility is regenerated on demand from the facility and the assumptions. Because the run is exactly reproducible, storing a million schedules would cost a great deal to preserve what can be rebuilt at will.
The validation conclusion is computed from the checks that actually ran. A run in which nothing was checked reports not validated — never a pass.
The screens follow the order the work is actually done in, and the platform always shows a single next instruction — no one should have to work out which screen comes after this one.
The schedules a credit team circulates are there as reports, not as exports of an internal structure: an ageing by band with customer counts, a customer listing ordered by allowance, exposure or delinquency, and the full facility detail behind both. Everything opens in Excel.
Any facility can be opened to its month-by-month schedule — default probability, survival, exposure, loss rate, discount factor and the resulting allowance — so a question about one customer is answered in one click rather than by rebuilding the calculation.