How a domain expert builds a model of their own business's future
Seven steps from framing the problem to a decision, then a new iteration. Shown on a commercial loan book of a US regional bank; all figures are illustrative.
Framing
the system asks, the expert answers in their own words; the framing sheet fills in as they go
Chief Credit Officer “knows how the book behaves”
Framing interview · section 1 of 7
Who makes the decision, and how often?
The credit committee. Origination limits and pricing are set monthly for six borrower segments.
What is the objective?
Net interest income over the next 12 months: interest earned minus the cost of funding the loans.
Which levers, and what are the limits?
Monthly limit per segment and the spread over SOFR. Wholesale funding is capped at $500M a month; capital budget is $12M a month.
How fast does demand react when you raise the spread?
I don't know counts as an answer
“I don't know” does not stop the work: the question goes to data discovery and onto the list of open questions.
the system reads the warehouse: structure, a passport for every column, defects, answers to questions
daily loan panel 326,184 rows · 90 columns
loan_idtext · 5,866 loans missing 0 %
as_of_datedate · 61 days Apr 1 – May 31
principalnumber · USD 0 … 158M
past_duenumber · USD 3 similar columns
disbursednumber · USD non-zero 2 %
ratenumber · % SOFR + 1.75 … 5.0
naicscode · 2-digit missing 17.7 % of rows
allowancenumber · USD CECL, month-end only
15 more tables contracts, clients, rates, ratings
contracts70,645 × 11 33 duplicate numbers
clients500,000 × 5 truncated export
rate_history163,987 × 7 SOFR + spread by loan
risk_ratings22,784 × 7 Pass … Loss, by month
a passport for every column: type, meaning, missing values, examples; defects are named, not hidden behind “data is not ready”
No EIN or parent-company field: 734 guaranteed loans can't be grouped by borrowerwhat we did: top-10 concentration reported as a range, 20.8–30.4 %
47 borrowers re-papered in May under new loan IDs ($112M)what we did: phantom repayments and phantom originations netted out
What moved the book in May?
The book fell from $5.37B to $5.31B (−1.2 %). Only $29M came from 26 new borrowers; the rest is amortisation of existing loans. Two ways of measuring the change disagree: −$23M from flows vs −$64M from balances.
The question is asked in words, the answer comes back with numbers and is written into the project's knowledge: 15 entries after three sessions.
Model
a node is a metric, an edge means “enters the equation”, the colour is where the number came from
from datafrom a findingassumptionopen questionrandom drawdecisionobjective
Known exactly → formulaNII = interest income − funding cost
$186Mmedian net interest income over 12 months; 80 % of outcomes between $161M and $212M
The constraint is checked on every trajectory
Capital budget ≤ $12M a month
breached in 0 % of trajectories, shown in red
Check the model
where every number came from, a backtest against history, sensitivity to parameters
Starting book · $5.31B from data
Amortisation · 1.9 % / month from data
Legacy yield · SOFR + 3.6 pp from a finding
Core deposit spread · SOFR − 2.0 pp assumption
Demand sensitivity · 0.15 per pp open question
Risk weight, small borrowers · 1.25 open question
22 parameters: from data 4 · from findings 5 · assumptions 13, all 13 waiting for the expert
The answer depends most on the SOFR path and the amortisation rate; the demand sensitivity the expert didn't know is third. Those are the parameters to pin down first.
Choosing the decision
thousands of combinations; the system searches them under the constraints
The optimisation problem
max net interest income, 12 months
subject to
capital budget ≤ $12M a month
wholesale funding ≤ $500M a month
largest segment ≤ 15 % of the book
The effect of a decision is shown as a shift of the whole distribution: a higher median and a shorter left tail.
Report
the document is assembled from every step and updates together with the model
Commercial loan book, Oct 2026 – Sep 2027: origination plan and net interest income
Model report · 16 source tables · 22 parameters · 3 scenarios · 1,000 trajectories per scenario
1. SummaryMedian net interest income over 12 months is $186M, 80 % of outcomes between $161M and $212M. The capital budget is breached in 11 % of stress trajectories. The main sources of spread are the SOFR path and the amortisation rate.
2. Problem and decisionThe credit committee sets a monthly origination limit and a spread over SOFR for six borrower segments, within a $500M wholesale funding cap and a $12M capital budget.
3. Established from dataThe book fell from $5.37B to $5.31B in May; new borrowers brought $29M. 17.7 % of rows have no industry code. Top-10 concentration is 20.8–30.4 % against a 25 % watch level.
4. Expert knowledge15 entries: large borrowers negotiate their own rate, so the spread grid binds only for mid and small clients; new loans enter at Pass.
5. Model and simulation18 metrics, 22 parameters: from data 4, from findings 5, assumptions 13. Three scenarios. The stress scenario cuts the median to $149M.
6. ChecksBacktest on the held-out month: balances inside the 10–90 % band, new demand missed by 4× and sent back for calibration. Sensitivity: SOFR ±$11.9M, amortisation −$9.9M / +$8.2M.
7. RecommendationRaise limits for trade and food segments at SOFR + 3.25, hold small borrowers at the current limit; the median rises to $214M within the capital budget.
8. Open questionsDemand sensitivity to the spread, prepayment rates, risk weights: to confirm with the expert before the next iteration.
Sections follow the steps of the method
framing → problem and decision
data discovery → established from data
interview → expert knowledge
model and simulation → scenarios and constraints
check → backtest, sensitivity
choosing the decision → recommendation
The document rebuilds whenever the model, the data or the expert's answers change. The next iteration starts by reframing the problem.