ARIMA time series analysis: when the forecast fits

ARIMA time series analysis uses earlier values and prediction errors, with differencing where needed, to describe a dated series and forecast later observations. For a business team, the practical test is whether supplied forecasts outperform a credible baseline on results that were unavailable when the forecasts were made. A convincing historical fit does not settle that question.

What ARIMA describes

ARIMA stands for autoregressive integrated moving average. Autoregressive terms relate a series to its earlier values. Differencing replaces levels with changes between observations to help address a changing level. Moving-average terms use earlier prediction errors; they are different from simply averaging recent sales.

Forecasting: Principles and Practice introduces ARIMA as an approach that describes autocorrelation in a series. That is dependence between observations at different time lags. Seasonal variants can incorporate relationships at seasonal intervals as well.

These ideas explain what a supplied method is trying to capture. They do not tell a finance manager which settings to select, or establish that every historical relationship will persist. Ask whoever prepared the forecast to document the measure, history cutoff, seasonal treatment and forecast horizon. Detailed selection and statistical diagnostics belong with that preparation work.

For ordered observations, trend and calendar patterns more broadly, start with time series analysis for business data. This page focuses on judging forecasts already supplied to the business.

Decide whether the business history is relevant

ARIMA can be a candidate when a consistently measured series has useful relationships with its own past. Examples include established monthly revenue or recurring service demand. It needs a history whose definitions and reporting boundaries make comparisons meaningful.

A new product, a major acquisition or a pricing policy change can weaken that connection. Historical revenue may also hide stockouts or capacity limits, so repeating the recorded pattern can repeat a constraint rather than underlying demand.

Calendar structure matters. If customers renew in a recurring month, compare the candidate with the same month last year as well as considering seasonal treatment. Seasonality in sales measures covers how calendar comparisons change the reading. Do not infer that a short history proves an annual pattern.

Ownership is shared. Finance owns the revenue definition and evaluation window. The forecast preparer documents the method and information available at each forecast date. Sales and operations explain changes in contracts, capacity or demand that historical values alone cannot capture.

Assemble the supplied forecasts and actuals

Keep one row per business segment, target period, forecast creation date and horizon. Include the forecast value and method label. The actuals export needs the same segment, target period, currency and measure definition.

Join using those business keys, retaining unmatched rows for review. Do not allow several forecast revisions to join to one actual and inflate the observation count. Choose a fixed revision rule, such as the approved forecast saved immediately before the target month starts.

The forecast accuracy guidance in Forecasting: Principles and Practice calls for evaluation on new observations rather than relying on fit to preparation data. A dated, preserved forecast file makes that principle checkable in an ordinary business review.

Use an explicit baseline. A naive forecast repeats the last available actual, while a seasonal naive forecast repeats the last observation from the corresponding season, as described in the textbook's simple forecasting methods. Select a baseline that reflects what the team could reasonably have used at the time.

Work the comparison on the same revenue rows

A US service company evaluates supplied monthly ARIMA forecasts for January-April 2026. Each forecast was saved before its target month began at a one-month horizon. December 2025 actual revenue was $96,000, which supplies January's naive baseline; each later baseline uses the preceding month's actual.

Use forecast minus actual as the signed error, so a positive result means overforecasting. Some references reverse the sign; record the convention beside the report. Absolute error ignores direction and measures miss size.

The supplied forecast's mean absolute error is 15,000 / 4 = $3,750. The baseline's is 34,000 / 4 = $8,500. The candidate has the lower average miss in these observations, although the baseline has the smaller miss in March and ties in January.

For a scale-relative view, divide total absolute error by total actual revenue. This gives 15,000 / 435,000 = 3.45% for the supplied forecast and 34,000 / 435,000 = 7.82% for the baseline. These are weighted absolute percentage errors, not averages of monthly percentage errors.

Net signed error divided by total actual revenue is 5,000 / 435,000 = 1.15% for the candidate and -24,000 / 435,000 = -5.52% for the baseline. Percentages are rounded to two decimal places; source dollar amounts remain exact.

Prove the totals before choosing a method

Two identities check the signed errors: 440,000 - 435,000 = 5,000, and 411,000 - 435,000 = -24,000. Both must match the sum of their monthly signed-error rows. Absolute errors must be calculated before summing, because opposite errors cancel in a signed total.

Reconcile actual revenue to the closed ledger, then verify that forecast and baseline each cover the same target periods. Keep the forecast timestamps: February's baseline legitimately uses January actuals, but only because January had closed before that forecast date.

These observations support further evaluation of the supplied forecast, not an automatic method choice. Extend the comparison across relevant business conditions and decision horizons. The guide to evaluating time series forecasts owns that broader evaluation process, and the forecast accuracy and bias workbook guide covers spreadsheet implementation.

Where it goes wrong

• Revised forecasts replace the originals. Later information makes an earlier decision look better than it was. Preserve each approved version and creation date.
• Horizons are mixed. A forecast made just before month-end has more information than one made before the month starts. Compare like horizons.
• A weak baseline creates an easy win. A seasonal business may need a seasonal reference; last month's result may be an unsuitable comparison.
• Aggregate bias hides offsetting misses. Inspect monthly and segment errors alongside the net signed total before interpreting accuracy.
• Business changes are ignored. Acquisitions, new contracts and capacity changes can break the usefulness of relationships learned from earlier periods.

Compare your supplied forecasts in Covirage

Bring the supplied forecast export, actual revenue rows and a clearly defined baseline, or a Google Sheet containing those inputs. Covirage maps the columns, and its tools compute and check the comparisons; the external AI model explains the result with every figure traced to its rows. Create a free Covirage account in about a minute, then set up your business and bring a sample of your own data.

See Covirage for finance teams for the supporting business context.