Cash forecasts fail when they hide the age of their data
By Equipo Quantum Developers
Summarize:
A forecast can reconcile mathematically and still be wrong for the decision at hand. The thesis is specific: an automated forecast is not trustworthy unless every figure states when it was observed, where it came from, and which decisions it is authorized to support. Yesterday’s bank balance, today’s receivable, and a week-old obligation do not become one coherent snapshot merely because they share a column.
Cash-flow automation projects often focus on adding sources or improving a forecasting model. A more basic risk remains hidden: records with different ages, closing rules, and authorities are blended without warning. A source-freshness ledger is a practical way to expose that risk.
Separate accounting classification from operating authority
IAS 7 classifies cash flows as operating, investing, or financing activities and requires reconciliation of cash and cash equivalents with the statement of financial position (IAS 7). That is an accounting foundation, not a daily treasury policy. Data suitable for a closed reporting period may not be fresh enough to authorize a transfer today.
Every use should therefore declare its horizon: observed position, forecast, scenario, or decision. Records should also distinguish confirmed cash, movement in transit, approved obligation, commercial estimate, and manual assumption. A system that blends those categories without labels produces a precise-looking number that cannot be defended.
The source-freshness ledger
The ledger may be a table or metadata object attached to every source. At minimum, include:
| Field | Purpose |
|---|---|
| source_id | identifies the system, account, or dataset |
| business_object | account, invoice, payment, payroll, or obligation |
| observed_at | when the value occurred or was valid |
| extracted_at | when it was read from the source |
| available_at | when the forecast could use it |
| expected_cadence | agreed update frequency |
| freshness_tolerance | maximum age for this specific use |
| reconciliation_state | confirmed, partial, pending, or disputed |
| provenance | process and version that produced it |
| owner | person accountable for meaning and correction |
| decision_right | inform, recommend, simulate, or commit |
These field names are an operating proposal, not a standard. The underlying discipline is well established. The Basel Framework summarizes BCBS 239 through expectations for accuracy, integrity, completeness, and timeliness, noting that timing depends on volatility and criticality (BIS, SRP 36). One freshness threshold should not be copied across every source; the decision determines the tolerance.
The W3C PROV model relates entities, activities, and agents to describe how information was produced and where responsibility sits (W3C PROV-DM). Applied here, do not retain only the forecast output. Preserve the source entity, transformation activity, and accountable process or person.
Decision rights are not technical permissions
Read access does not equal financial authority. Separate four levels:
- Inform: display a value and its state.
- Recommend: propose an action with visible assumptions.
- Simulate: calculate scenarios without changing source systems.
- Commit: create an instruction, transfer, or contractual change.
The final level requires an explicit policy, separate approval, and a freshness check immediately before action. A scenario approved in the morning should not enable execution later if critical inputs have changed. Authorization belongs to a specific scenario version and an identifiable set of inputs.
When a source exceeds tolerance, do not silently remove it. Mark the scenario as degraded, state which decisions are blocked, and request an update. Visible absence is safer and more informative than manufactured certainty.
Illustrative weekly treasury review
Consider a company preparing a weekly view from bank accounts, receivables, approved payments, and payroll. This is an illustrative design, not a customer result. Banks supply observed balances; the ERP supplies invoices and obligations; sales maintains collection assumptions; payroll contributes an approved calendar.
The ledger may allow a current balance to inform the position, permit a disputed invoice only in an optimistic scenario, and prohibit a sales probability from authorizing a transfer. If one account extract is stale, the scenario remains visible but is labeled incomplete. The agent may recommend obtaining the missing record or deferring a discretionary outflow. It may not invent the balance.
The meeting changes from “What is the number?” to “Which portion is observed, which is assumed, which source is late, and what decision may this version support?” That conversation takes more effort initially and becomes far easier to audit.
Bringing the ledger into a control plane
In Quantum Automation Center, each run can associate sources, status, timelines, artifacts, and logs. Financial analytics should expose freshness of the input set rather than only an aggregate value. Business objects let an invoice or payment remain linked to later events without losing identity; see the business-object documentation.
A useful view includes the scenario cut-off time, expired critical sources, reconciliation state without pretending it is a quality guarantee, manual assumptions, approver, applicable policy, and enabled decisions. An approved artifact should be immutable; a new data refresh should create a new version.
The counterargument: exploration can stay lightweight
A low-risk strategic conversation may need only a directional projection with explicit assumptions. Building field-level lineage before learning whether the view will be used can cost more than the insight. Start with sources that change decisions and fields that can block action.
Controls should be proportional. An exploratory worksheet may use manual labels. A recurring recommendation needs validation. A payment instruction needs evidence and separated authority.
When not to automate cash decisions
Do not automate commitments when accounts are unreconciled, sources have no owners, or liquidity policy depends on unrecorded conversations. Avoid it as well when integrating data costs more than the value of an infrequent decision; document a manual cut-off procedure instead.
The final discipline is simple. Every figure must answer what was true, when, according to whom, and for which decision. If the system cannot answer, it contains numbers—not a governed cash-flow capability.
