Forecast accuracy

Measure amount and timing errors against forecasts that were preserved before the events.

View Markdown
Browse documentation
On this page

A forecast is useful to evaluate only if you know what it said before the actual arrived. Balsa measures accuracy against preserved expectations and reviewed actual allocations, not reconstructed guesses about what the model might have predicted.

Build a measurable history

  1. Capture expectations in /reconciliation before their expected events occur.
  2. Preserve a plan or outlook review in /goals before the event and its actual payments.
  3. Post and explicitly allocate actual payments as they arrive.
  4. Complete the obligation, including an explicit final settlement when appropriate.
  5. In /reviews, choose Forecast accuracy and select that baseline.

A baseline with no captured expectations has no observations to score. Saving a snapshot after payment does not create a valid pre-event forecast.

Read the result

Measure Definition
Amount error Final allocated actual minus preserved expected amount
Timing error Final payment date minus preserved expected date, in days
Weighted absolute error Sum of absolute amount errors divided by sum of preserved expected amounts
Cash optimism How much the forecast overestimated cash: overstated receipts or understated spending

Positive timing error means late completion. Positive cash optimism means the forecast was too optimistic about cash.

For a synthetic $100,000 expected receipt that is explicitly settled in full for $90,000 five days late, amount error is negative $10,000, timing error is five days, and cash optimism is positive $10,000.

Measured, pending and excluded

Only completed, compatible obligations with valid allocations and a pre-event baseline are scored.

  • Partial and unmatched obligations remain pending; they are not treated as zero actuals.
  • Cancelled or missing obligations are excluded rather than treated as amount-error observations.
  • Reversed, unavailable or incompatible allocations invalidate the observation.
  • Baselines preserved on or after the expected event date, or after actual payment activity, are excluded.

Inspect these counts alongside the headline error. A small measured sample can omit much of the business. Zero completed observations means accuracy is unavailable, not perfect.

Use the learning carefully

Driver summaries show completed sample count, late observations and cash optimism. Fewer than three observations are explicitly insufficient for a repeated-pattern conclusion; three is not a statistical guarantee either.

The report does not automatically recalibrate amounts or timing. Review the source facts and decide whether to replace a recurring assumption or evaluate a scenario.

Saved accuracy reports preserve formulas, inputs, original source references and allocated actual identities. Follow Reviews and sign-off to record who reviewed the result.