Practical guide
AI financial forecast verification checklist
Use this checklist after a financial forecast has been calculated and before it is issued for a decision. It tests whether the forecast answers the stated question, reconciles to controlled evidence, exposes material assumptions, behaves correctly under challenge, and carries an appropriate release record. A polished narrative, plausible trend, or successful recalculation does not compensate for.
Use this checklist after a financial forecast has been calculated and before it is issued for a decision. It tests whether the forecast answers the stated question, reconciles to controlled evidence, exposes material assumptions, behaves correctly under challenge, and carries an appropriate release record. A polished narrative, plausible trend, or successful recalculation does not compensate for an unreconciled opening position, hidden driver, broken formula, or uncertainty that could reverse the decision.
What this checklist can verify
The UK government's AQuA Book says assurance should be proportionate to intended use and risk, continue across the analytical lifecycle, distinguish verification from validation, retain traceable records, and communicate uncertainty. The Government Actuary's Department describes financial-model assurance as checking calculations, data, assumptions, methods, documentation, scrutiny, limitations, and uncertainty. These sources support disciplined review. They do not determine whether a particular forecast is correct or suitable for an organization's decision. AQuA Book Government Actuary's Department: Model quality assurance
Sources: [1] [2]
How to use the AUDIT checklist
Mark every item pass, fail, not applicable, or not evidenced. A pass needs an evidence ID or a reproducible check. Assign each failure an owner, due date, and consequence. Do not average critical failures into an overall score.
### A: Align the decision and version
Stop condition: hold issue when the decision or model version is ambiguous, or when a narrative describes a different result from the controlled model.
### U: Underpin inputs and opening positions
Stop condition: do not issue a cash, revenue, cost, or capacity conclusion when its opening position or material input population cannot be reconciled.
### D: Document drivers, assumptions, and scenarios
Stop condition: return the forecast when a material conclusion depends on an undocumented assumption or on an AI-generated driver that no accountable owner accepted.
### I: Inspect calculations and model behavior
Stop condition: hold any conclusion that depends on an irreproducible calculation, circular logic, broken test, or unexplained override.
### T: Treat uncertainty and release controls
Stop condition: do not present a precise point estimate when unresolved uncertainty could change the action and no bounded contingency exists.
- [ ] The decision, accountable decision maker, decision date, alternatives, horizon, currency, reporting unit, and materiality threshold are explicit.
- [ ] The issued model, narrative, dashboard, source snapshot, and approval record carry the same version.
- [ ] Actuals, commitments, estimates, scenarios, and management targets are visibly separated.
- [ ] Permitted uses and users are stated, including any purpose for which the forecast is not fit.
- [ ] Changes since the previous issued version are classified as data, assumption, method, correction, or timing changes.
- [ ] Opening balances and historical totals reconcile to the governed ledger or approved source.
- [ ] Every material input has a source, owner, extraction date, period, unit, transformation, and reconciliation result.
- [ ] Duplicate records, missing periods, sign conventions, currencies, late adjustments, and definition changes have been tested.
- [ ] Contracted commitments, observed actuals, operational drivers, and unapproved suggestions are not mixed.
- [ ] AI-extracted values have been checked against the original approved source.
- [ ] Each material assumption has an ID, rationale, evidence, owner, effective period, range, dependency, status, and expiry trigger.
- [ ] Hard-coded values in formulas, scripts, prompts, and presentation files map to approved assumptions.
- [ ] Scenarios represent coherent states rather than unrelated percentage changes.
- [ ] Sensitivities change one driver at a time and identify the outputs most affected.
- [ ] Base, upside, downside, and stress labels do not imply probabilities unless probabilities have a defensible method.
- [ ] Management targets are not presented as forecast outcomes.
- [ ] A model map traces sources through transformations and calculations to decision measures.
- [ ] Balance, unit, sign, period, roll-forward, boundary, duplicate, and impossible-value tests pass.
- [ ] A reviewer can reproduce sampled outputs independently.
- [ ] Formula, code, or prompt changes are version controlled and regression tested.
- [ ] Challenger calculations or reasonableness checks cover the material outputs.
- [ ] Verification confirms the implementation follows its design; validation separately asks whether the method is fit for the decision.
- [ ] The forecast shows a useful range or scenario set and identifies the drivers that matter.
- [ ] Quantified and unquantified uncertainties are distinguished.
- [ ] Actual-versus-forecast checks show whether prior errors reveal bias, missing drivers, or process failures.
- [ ] The decision remains clear under plausible alternatives, or a contingency trigger covers reversal risk.
- [ ] Open exceptions, limitations, residual uncertainty, preparer, reviewer, approver, release date, and refresh trigger are recorded.
- [ ] Material changes to data, method, AI system, horizon, or decision trigger reapproval.
Verification record
| Area | Result | Evidence IDs | Failure or uncertainty | Owner | Due date | Decision consequence | | --- | --- | --- | --- | --- | --- | --- | | decision and version | | | | | | | | inputs and reconciliation | | | | | | | | drivers and assumptions | | | | | | | | calculations and behavior | | | | | | | | uncertainty and release | | | | | | |
Overall disposition: verified for the stated decision, verified with visible limitations, return for repair, or hold the dependent decision. The disposition applies only to the reviewed version and evidence packet.
Worked verification: synthetic subscription forecast
A fictional software company prepares a six-month revenue forecast for a hiring-capacity decision. The deck shows £4.8 million revenue and a narrow confidence range. The model uses actual recurring revenue, a renewal schedule, sales pipeline, and AI-drafted account notes.
AUDIT finds that the model and deck share a version, but the input check does not pass. Twelve expired opportunities remain in the pipeline, three renewals appear twice after a CRM merge, and the model treats every AI-classified account note as an approved probability change. Removing duplicates and expired opportunities lowers the base case by £310,000. The commercial owner accepts only four of 27 suggested probability changes and records the supporting account evidence.
Calculation tests identify a formula that shifts churn one month late. The corrected model produces £4.32 million in the base case, £3.88 million in the downside, and £4.55 million in the upside. A sensitivity check shows that renewal timing, not new-logo conversion, drives the capacity threshold. The prior deck's narrow range is therefore not supported.
The disposition is return for repair. The owner corrects the churn timing, freezes the source snapshot, links each probability override to an approval, and adds a contingency: defer two planned starts if two named renewals remain unsigned at the review date. The checklist does not declare the forecast accurate or approve the hiring decision.
Two Prune boundary
Two Prune can assess whether a participant verifies a supplied forecast, reconciles evidence, challenges assumptions, checks calculations, and communicates a bounded disposition. It reports observed performance in that assessment. It does not audit or approve an organization's forecast, provide financial advice, certify model controls, predict workplace performance, or make employment decisions.
Sources
UK Government: The AQuA Book
- Government Actuary's Department: Model quality assurance
Sources: [1] [2]
Sources
- 1.AQuA Book · UK Government
- 2.Government Actuary's Department: Model quality assurance · UK Government