Blog · Forecast and pipeline · Finance and FP&A teams
Separate persistent forecast bias, irregular variation, timing shifts, structural changes, and data mistakes. Use a forecast snapshot and actuals export to find the pattern before choosing a correction.
The main types of forecasting errors are persistent bias, irregular variation, timing shifts, structural changes, and data mistakes. Diagnose them by keeping the direction of each miss and examining its sequence by segment. An accuracy percentage alone cannot tell you which correction to make.
Start with the forecast version available when the decision was made. Retain its publication date, forecasted period, segment, horizon, amount, and currency. Join it to actuals at the same period and segment level, and check that each combination appears once.
For a US services company, a revenue forecast must meet recognized revenue actuals on the same basis. Bookings, invoices, and cash receipts answer different questions. A timing difference between invoicing and revenue recognition is not automatically a forecasting failure.
Use one sign convention throughout:
Signed error = Actual − Forecast
Absolute error = Absolute value of the signed error
Negative means the forecast was high; positive means it was low. Forecasting: Principles and Practice defines forecast errors as observed values minus forecasts, and distinguishes genuine forecast errors from residuals calculated on data used to fit statistical methods.
Keep the detailed measurement steps with the forecast accuracy and bias walkthrough. The purpose here is to identify what the pattern suggests and what evidence would confirm it.
This illustrative company froze a weekly revenue forecast for three regions. All figures are USD, all weeks use the same accounting calendar, and the actuals have closed. Columns A through F follow the table order below.
| Region | Week | Forecast (USD) | Actual (USD) | Signed error (USD) | Absolute error (USD) |
|---|---|---|---|---|---|
| North | 1 | 100,000 | 90,000 | −10,000 | 10,000 |
| North | 2 | 100,000 | 85,000 | −15,000 | 15,000 |
| North | 3 | 100,000 | 95,000 | −5,000 | 5,000 |
| North | 4 | 100,000 | 90,000 | −10,000 | 10,000 |
| South | 1 | 100,000 | 90,000 | −10,000 | 10,000 |
| South | 2 | 100,000 | 115,000 | 15,000 | 15,000 |
| South | 3 | 100,000 | 95,000 | −5,000 | 5,000 |
| South | 4 | 100,000 | 100,000 | 0 | 0 |
| West | 1 | 80,000 | 120,000 | 40,000 | 40,000 |
| West | 2 | 120,000 | 80,000 | −40,000 | 40,000 |
| West | 3 | 80,000 | 120,000 | 40,000 | 40,000 |
| West | 4 | 120,000 | 80,000 | −40,000 | 40,000 |
| Total | 1,200,000 | 1,160,000 | −40,000 | 230,000 |
In E2 and F2, fill these formulas through row 13:
=D2-C2
=ABS(E2)
The total signed error is −$40,000. The absolute errors total $230,000 because opposite directions do not cancel. Check =SUM(E2:E13)-(SUM(D2:D13)-SUM(C2:C13)): it must return zero.
Put the region names in H2:H4. These formulas calculate signed error, absolute error, and mean absolute error for the region in H2:
=SUMIF($A$2:$A$13,H2,$E$2:$E$13)
=SUMIF($A$2:$A$13,H2,$F$2:$F$13)
=SUMIF($A$2:$A$13,H2,$F$2:$F$13)/COUNTIF($A$2:$A$13,H2)
| Region | Forecast total (USD) | Actual total (USD) | Signed error (USD) | Absolute error total (USD) | Mean absolute error per week (USD) |
|---|---|---|---|---|---|
| North | 400,000 | 360,000 | −40,000 | 40,000 | 10,000 |
| South | 400,000 | 400,000 | 0 | 30,000 | 7,500 |
| West | 400,000 | 400,000 | 0 | 160,000 | 40,000 |
North's mean absolute error is 40,000 / 4 = $10,000. South's is 30,000 / 4 = $7,500, and West's is 160,000 / 4 = $40,000. These figures are exact; no rounding is required.
The three regional signed errors sum to −$40,000, and their absolute totals sum to $230,000. Both must match the detailed table. Reconcile the actual revenue total to the company's closed revenue report before interpreting any pattern.
| Type | What to look for | Evidence and response |
|---|---|---|
| Persistent bias | Repeated misses in one direction | Review assumptions and overrides across comparable horizons |
| Irregular variation | Mixed signs without an established repeatable pattern | Check more history and uncertainty before making a permanent adjustment |
| Timing error | Opposite misses in nearby periods | Match transaction IDs and expected versus actual dates |
| Structural change | Errors begin after a lasting business change | Separate the old operating conditions from the new ones |
| Data or definition mistake | Duplicates, missing rows, changing scope, or inconsistent units | Repair the comparison before assessing forecast quality |
North suggests overforecasting. All four errors are negative. Review unsupported renewals, optimistic utilization, and repeated manual uplifts. Four weeks justify investigation, not a permanent percentage reduction. The owner of each assumption should explain the evidence; forecast bias by rep covers the corresponding sales review.
South is a candidate for irregular variation. Its signed errors cancel, and its absolute misses are smaller than West's. That does not prove randomness. Check a longer sequence for recurring customer, calendar, or seasonal effects before treating the misses as unavoidable noise.
West suggests a timing issue. Each adjacent pair has equal and opposite errors. Confirm whether the same projects or transactions moved between weeks. If so, improve the date assumptions while retaining the published forecast history and the correctly recorded actuals.
Sequence matters. Penn State's time series teaching material explains that residuals should be unrelated across time; visible dependence can signal information a forecasting approach has missed. A short business review is a screening step, not a statistical proof.
A structural change needs different evidence: a closed location, a lost contract, a lasting price change, or an altered service offering. Mark the effective date and compare forecast errors before and after it. A single unexpected invoice does not establish a new demand level.
Set any acceptance threshold around the decision, volatility, and horizon; what good forecast accuracy means explains those conditions.
Bring the frozen forecast snapshots and the matching actuals export, including segment, period, currency, and version dates. Covirage maps the columns, and its tools compute and check signed errors, absolute errors, and reconciliations; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring a sample of your own data.
See forecast analysis for the supporting business use.
A forecast error is the difference between an actual result and its forecast for one observation. Bias is a persistent tendency to miss in the same direction across observations. A large individual miss does not establish bias, and offsetting misses can leave the average near zero.
No. Positive and negative errors can cancel even when every period has a substantial miss. Keep an absolute error measure beside the signed measure and inspect errors by period, customer, and segment before concluding that a forecast supports the business decision.
Look for an overforecast followed by an underforecast, then match the affected transactions to their expected and actual dates. A stable total across adjacent periods supports the hypothesis, but it is not proof. Confirm the movement with order, delivery, or accounting evidence.
No. Future outcomes contain uncertainty, and new events can change business conditions. Teams can correct data mistakes, improve assumptions, and reduce persistent bias. They should also measure the remaining variation and plan capacity or liquidity for plausible outcomes beyond the point estimate.