Real-time demand forecasting: when faster updates help

Real time demand forecasting helps when new information improves the demand estimate available before a business decision closes. Measure that improvement with timestamped forecast versions, consistent actual demand and an explicit cutoff for each decision. A revision that arrives after a purchase commitment may be accurate and still have no value for that purchase.

Define what faster updates should improve

For a US distributor, the practical question is whether another update changes purchasing, labor or allocation in time to matter. Start with the demand forecasting definition of the target: which items, locations, units and future periods the forecast covers. Then name the decision it supports.

Write down the last time that decision can change without an extra charge or an approval. Keep a separate cutoff for each action. A buyer's order deadline and a warehouse manager's staffing deadline need not match, even when both use the same demand forecast.

Also distinguish the time an order happened, the time it entered the source file and the time the forecast was issued. An export generated later can contain earlier transactions that the planner could not have seen at the original cutoff. Reconstructing that earlier information set matters more than attaching a recent timestamp to a file.

Preserve the versions, then join the actuals

Keep one forecast record per item, site, target period and issue timestamp. Retain the forecast method or scenario name if it changes. The actual-demand file needs the same item and site IDs, the same target-period boundaries and the same unit of measure.

Record cancellations, returns and stock availability separately. Decide whether actual demand means requested units, accepted orders or shipments, and apply that choice to every comparison. Shipments during a stockout are a constrained observation, so flag those periods rather than quietly treating them as complete demand.

Forecasting: Principles and Practice explains why accuracy should be evaluated on observations that were not used to produce the forecast. Here, preserving the original forecast export is the business control that makes that evaluation possible.

Freeze actuals at a stated close date. If late cancellations change them later, label the restatement. Otherwise a forecast can appear to improve because the benchmark moved.

A worked comparison at a purchasing cutoff

Consider an illustrative Ohio distributor ordering one item for its Columbus warehouse. Its target is total requested units for each week beginning on the dates below. All forecast timestamps use Eastern Daylight Time.

For each target week, the distributor saves versions from the preceding week's Monday at 8:00 a.m., Tuesday at 4:00 p.m. and Friday at 4:00 p.m. Purchasing closes on that preceding Wednesday at noon. Monday and Tuesday are eligible; Friday is too late for this decision.

Use forecast minus actual as the signed-error convention: positive means overforecast. Take the absolute value before adding misses so opposite directions cannot cancel.

WAPE = Sum of absolute forecast errors / Sum of actual demand

MAE = Sum of absolute forecast errors / Number of matched observations

Monday WAPE is 600 / 4000 = 15.0%; Tuesday is 240 / 4000 = 6.0%. Friday is 35 / 4000 = 0.875%, displayed as 0.9% when rounding percentages to one decimal place. The corresponding MAEs are 150, 60 and 8.75 units, using the four observations in each column.

The usable Tuesday revision reduces absolute error by 600 - 240 = 360 units, or 360 / 600 = 60.0%. WAPE falls by 15.0 - 6.0 = 9.0 percentage points. Friday's lower error is useful evidence for a later decision, but cannot justify changing the earlier purchasing process.

Check the comparison before changing the cadence

The actual-demand denominator must remain 4000 units for every version. The compared row counts must also match. Excluding difficult weeks from only the newer forecast creates an apparent gain without a fair test.

Reconcile signed errors separately. Monday's errors are -100, -200, 200 and 100, summing to zero. Tuesday's -20, -100, 100 and 20 also sum to zero. Both total forecasts match actual demand, yet their absolute errors differ substantially. Keep forecast bias by SKU and site beside the accuracy measure.

This comparison tests the value of an extra update, including its shorter lead time. It does not isolate which forecasting method is better. To compare methods, hold their issue times and information sets constant. The textbook's time series cross-validation guidance also supports evaluating successive forecast origins and the horizons relevant to the decision.

Decide who acts on the result

The demand planner owns version definitions and matched comparisons. Purchasing owns its commitment cutoff; operations owns the decisions that can still change later. Together they should record whether a revision changed an action, then compare the resulting service and cost outcomes separately from forecast accuracy.

Repeat the evaluation across representative periods before setting a permanent cadence. Four illustrative weeks demonstrate the calculation, not a universal update policy. Use the demand forecasting process to assign review ownership, and choose refresh frequency by measure when different decisions need different schedules.

Where it goes wrong

• Using the final revision as the original plan. This imports later knowledge into the benchmark. Preserve every issue timestamp.
• Ignoring the decision deadline. A late forecast can improve reported accuracy without changing the action under review.
• Mixing actual-demand definitions. Requested units and fulfilled shipments answer different questions, especially during stockouts.
• Dropping unmatched rows silently. Report missing versions and compare a common set of targets before calculating improvement.
• Treating every change as useful information. Investigate reversals and incomplete orders before adding more review cycles.

Compare forecast versions on your own file in Covirage

Upload timestamped forecast snapshots, matched actual demand and your decision-cutoff file; Covirage maps the columns, and its tools compute and check the supplied comparisons. 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 Covirage for supply chain teams for the supporting business context.