Practical guide

How to create a financial forecast with AI

An AI-assisted financial forecast is still a financial model: its usefulness depends on a defined decision, controlled source data, explicit assumptions, reproducible calculations, and visible uncertainty. AI can help classify inputs, draft assumption notes, generate challenge questions, and explain movements. It should not silently invent drivers, replace governed calculations, or turn a.

By Two Prune

An AI-assisted financial forecast is still a financial model: its usefulness depends on a defined decision, controlled source data, explicit assumptions, reproducible calculations, and visible uncertainty. AI can help classify inputs, draft assumption notes, generate challenge questions, and explain movements. It should not silently invent drivers, replace governed calculations, or turn a scenario into a commitment.

The forecast and claims boundary

The UK government's AQuA Book says analytical quality assurance should be proportionate to intended use and risk, continue through the analytical lifecycle, distinguish verification from validation, document inputs and assumptions, 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 a disciplined forecast process; they do not make a forecast accurate or approve the use of AI in a particular finance environment. AQuA Book Government Actuary's Department: Model quality assurance

Sources: [1] [2]

Use the TRACE forecast method

### T — Target one decision

Write the decision before opening a model. Name the decision owner, decision date, forecast horizon, reporting currency, level of detail, and materiality threshold. A liquidity decision needs receipts and payment timing; a capacity decision may need volume, utilization, and unit economics. Do not use a general revenue forecast as a substitute for a decision-specific model.

Record a base case and the alternatives the decision maker can actually choose. Preserve a no-change or committed-plan comparator. If there is no decision, produce a planning estimate rather than presenting a decision forecast.

### R — Reconcile controlled inputs

Create an input register with source, owner, extraction date, period, unit, transformation, and reconciliation result. Separate recorded actuals, contracted commitments, operational drivers, management assumptions, and AI-generated suggestions. AI suggestions are hypotheses until an accountable person accepts and records them.

Reconcile opening balances and historical totals to the governed ledger or approved source. Test missing periods, duplicate records, sign conventions, currency conversion, late adjustments, and changes in definitions. Freeze the input snapshot used for an issued forecast.

### A — Author assumptions and scenarios

Each material assumption needs an ID, rationale, owner, effective period, evidence, range, dependency, expiry trigger, and approval status. Link formulas to assumption IDs rather than burying constants in cells or prompts.

Use scenarios to express coherent states of the world, not arbitrary percentage changes. For example, a downside case may combine delayed renewals, lower conversion, and slower collections because those drivers share a plausible mechanism. Keep sensitivities separate: a sensitivity changes one driver to show influence while a scenario changes a coherent set.

AI can extract candidate drivers from approved documents or draft scenario narratives. A human reviewer should compare every extracted value with its source and reject any driver that cannot be traced.

### C — Calculate through a visible model map

Map the path from sources to assumptions, calculations, outputs, and decision measures. Keep calculations deterministic and inspectable. Use AI-generated code or formulas only after review, tests, and version control.

Add checks for balance, units, signs, period alignment, roll-forward logic, duplicate inclusion, and impossible values. Recalculate material outputs independently or with a challenger model. Verification asks whether the implementation follows its specification; validation asks whether the method is fit for the intended decision.

### E — Expose uncertainty and issue controls

Show ranges and the drivers that matter, not just a precise point estimate. Explain which uncertainties are quantified, which remain descriptive, and where data is too weak for a useful range. Compare the current forecast with the prior issued version and actual outcomes, separating data changes, assumption changes, model changes, and timing effects.

Before release, record preparer, reviewer, approver, version, source snapshot, model hash, open exceptions, permitted uses, and next refresh date. Trigger reapproval after a material change to data, method, AI system, decision, or forecast horizon.

Forecast evidence pack

| Artifact | Minimum contents | Verification | | --- | --- | --- | | decision card | decision, owner, date, horizon, alternatives, materiality | owner confirms the forecast answers the stated decision | | input register | source, owner, extraction, unit, transformation, reconciliation | totals reproduce from frozen source extracts | | assumption register | ID, value or range, evidence, owner, status, expiry | every material constant maps to an approved ID | | model map | inputs, transformations, calculations, outputs | reviewer traces sample outputs back to sources | | test record | balance, boundary, sensitivity, regression, challenger tests | failures have dispositions and retest evidence | | movement bridge | prior forecast, actuals, data, assumption and model changes | bridge reconciles to the newly issued forecast | | release record | version, approvals, limitations, permitted use, refresh trigger | issued files and narrative carry the same version |

Worked example: synthetic 13-week cash forecast

A fictional services company must decide whether to delay a discretionary software rollout. Finance has bank actuals through Friday, an accounts-receivable aging file, approved payroll, supplier commitments, and a sales pipeline. The first AI draft describes likely collections but incorrectly treats every open invoice as collectible on its due date and invents a 90% pipeline conversion assumption.

The controller applies TRACE. The target is minimum weekly cash under the committed plan, delay option, and downside case. The input register reconciles the opening bank balance and removes three duplicate supplier invoices. Receivables are grouped by observed payment behavior, while the unapproved pipeline assumption is removed. Assumption A-07 records a 12-day collection delay for disputed invoices, with the credit manager as owner and a range of 8–20 days.

The model map rolls opening cash through receipts and payments by week. Boundary tests catch payroll posted one week early; an independent calculation reproduces the corrected minimum cash. The base case reaches a minimum of £410,000. The downside reaches £75,000, driven mostly by two disputed invoices, while delaying the rollout preserves £190,000. These are synthetic planning outputs, not probabilities.

The movement bridge explains the difference from last week's forecast: £120,000 from new actuals, negative £80,000 from collection timing, positive £35,000 from a corrected duplicate, and negative £25,000 from payroll timing. The release note recommends a conditional delay review if either disputed invoice misses its evidence-backed collection checkpoint. It does not claim that AI improved forecast accuracy.

Failure conditions

Return the forecast for repair when the opening position does not reconcile, material values lack sources, formulas depend on hidden prompt outputs, scenarios mix incompatible assumptions, or a reviewer cannot reproduce the decision measure. Hold the dependent decision when uncertainty could reverse the choice and no bounded contingency exists. Do not describe a model as validated merely because formulas run or historical values fit.

Two Prune boundary

Two Prune can assess whether a participant frames a forecasting decision, examines supplied evidence, challenges assumptions, verifies calculations, and communicates a bounded recommendation. It reports observed performance in that assessment. It does not build or approve an organization's financial 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. 1.AQuA Book · UK Government
  2. 2.Government Actuary's Department: Model quality assurance · UK Government

Explore this topic