A missing revenue observation is not a zero-sales day

An empty daily revenue field can mean no transactions occurred, the feed failed, the business was closed or the reporting population did not yet exist. Replacing all those cases with zero gives the forecasting model a stronger statement than the source supports.

Hyndman and Athanasopoulos emphasize that missing observations need interpretation in their forecasting context. Missingness can itself be informative, and different methods handle it differently. The first task is therefore to identify what happened, rather than choose the fastest spreadsheet replacement.

Use distinct statuses for observed zero, observed nonzero, missing and not applicable where the data process permits. Preserve the date and expected reporting coverage. A blank row that disappears during an export can otherwise make both the activity and the calendar look different from reality.

Consider an invented five-day revenue record containing $20,000, $22,000, a missing observation, $21,000 and $23,000. The four observed values total $86,000. Treating the missing day as zero produces a five-day average of $17,200.

Averaging only the four observed days gives $21,500. If the original source later confirms that the missing day actually generated $24,000, the complete five-day average becomes $22,000. Neither the zero replacement nor the observed-only average recovered the true value automatically.

A deliberately simple ten-day forecast using those averages would be $172,000, $215,000 or $220,000 respectively. This is not a recommendation to forecast revenue from five daily observations. It isolates how a data-treatment decision can materially change a financial projection before any sophisticated modeling begins.

Review source-system status, reporting cutoffs and transaction counts. A failed transfer is different from a store closure or an unresolved accounting adjustment. Identify whether the amount is absent from the source itself or merely missing from the analytical extract.

Consider whether missingness is related to activity. If a feed tends to fail on exceptionally busy days, the observed-only average can understate revenue. If a manual report is skipped on quiet days, excluding missing observations can have the opposite effect. The direction cannot be determined from the blank alone.

Check whether transactions appear in a later period. A delayed batch may create an apparent zero followed by an inflated day. Recovering only the empty observation without removing duplicated amounts from the later day would overstate the total. Reconcile the date basis as well as the amount.

Where possible, reconstruct the observation from authoritative transaction or accounting records appropriate to the forecast measure. Match entity, currency, revenue definition and recognition or transaction date. A payment-processor deposit is not automatically a substitute for daily revenue because fees and settlement timing can differ.

Reconcile the repaired series to an independent period total. That helps detect omitted or duplicated transactions across the missing interval. Keep the original extract and the repair record so a reviewer can distinguish recovered facts from later analytical choices.

If recovery is not possible, describe the remaining gap and its materiality. A short isolated gap in a stable series presents a different problem from missing all peak-season observations. The forecast method and uncertainty should reflect how much relevant information is absent.

Interpolation, a comparable-day value or a model-based estimate can sometimes make a series usable. Each relies on assumptions about the missing period. A straight line between adjacent days may be inappropriate across weekends, promotions or a known operating interruption.

Store the estimated value separately or retain an explicit imputation flag. Do not allow a replacement to become an indistinguishable actual in the next extract. Otherwise the forecasting system can gradually train on its own earlier guesses while reporting an apparently complete history.

Evaluate sensitivity to plausible alternatives when the missing amount affects a commitment. The goal is not to generate many arbitrary versions. Focus on whether reasonable replacements would change staffing, purchasing or funding decisions, and explain that dependency where it remains unresolved.

A replacement estimated using observations after the forecast date can be useful for reconstructing historical data. It may be unavailable in a realistic backtest of what the team could have forecast at that time. Keep retrospective reconstruction separate from forecasting with contemporaneous information.

Apply the chosen missing-data procedure within each historical training window when evaluating the operating method. Fitting the replacement model on the full dataset before withholding test periods can introduce future information into the supposedly historical forecast.

Do not score a prediction against an imputed outcome as though the outcome were independently observed. If the true value remains missing, label the evaluation gap. A forecast agreeing with a replacement produced by a related model is not strong evidence of actual accuracy.

A confirmed closure day with zero revenue belongs in the record when the target is calendar-day revenue. If the target is revenue per open trading day, use the appropriate denominator and retain the closure calendar separately. The answer depends on the forecast's defined population.

Real zeros can carry useful information about intermittent activity, cancellations or operational disruption. Removing them merely because a selected model cannot accept zero values changes the business question. Choose a suitable method or a supported transformation instead of discarding inconvenient outcomes.

A newly opened location also has periods before it existed. Those are not necessarily zero-performing operating periods. Define the active population and opening date so growth from added locations is not confused with improved performance in the existing business.

Report material gaps, recovered observations and estimated replacements with the forecast version. Assign the unresolved source issue to someone who can correct the process. Repeated gaps deserve a data fix rather than an increasingly elaborate imputation routine.

The finished series should preserve what the business knows and what it estimated. That distinction lets finance explain a forecast change when the real observation arrives, and prevents an empty field from becoming an unsupported statement that nothing happened.