Name the event and management decision

Forecasting revenue, demand for goods, payments and closing transactions are different tasks. Choose a horizon and solution: procurement, team workload or cash flow plan. The unit must correspond to this decision: pieces, money, number of hits or probability of an event. The sales plan reflects the desired result, and the forecast reflects what is expected under the selected premises; It is useful to show them separately. Do not replace the forecast with the sum of managers' promises without describing the rules. Before choosing a model, record when the estimate is generated and what information is actually available to the company at that moment.

Check the meaning of historical values

The story may mix different currencies, units, returns, and event recognition dates. Changing the directory can divide one product into several codes, and merging divisions can create a jump in the series. Separate lack of sales from lack of information. Record changes to the process and accounting rules so that the model does not mistake the administrative transition for a market trend. It is not necessary to have perfect data for any experiment, but limitations should be visible. If the past cannot be reproduced by an agreed upon definition of a metric, a seemingly accurate forecast will be difficult to verify and explain to management. Event coordination is discussed in detail in the article aboutWhy are the sales amounts different in CRM and the report?.

Distinguish observed sales from demand

Zero sales during a stock-out period do not prove a lack of customers. Promotions, price changes, supply restrictions and store closures also change the meaning of the story. Keep the available signs of these conditions and note the unknown ones. A forecast of observed sales can be useful, but it is not automatically an estimate of total demand. For rare or new products, the history may be too short to draw a confident conclusion. In such cases, it is more useful to indicate uncertainty and the need for an expert solution than to fill any gap with a compelling number.

DEMONSTRATION EXAMPLE

Study situation: a product was not sold for two weeks due to out of stock. If we consider these weeks as normal zero demand, the system may suggest reducing the next purchase. Availability data helps explain the observation, although it does not by itself reconstruct the exact number of failed purchases.

Check in future periods regarding training

For a time forecast, it is important to reproduce the order of events: build an assessment based on the available past and compare it with the subsequent fact. Random shuffling may provide access to information that was not available at the date of the decision. Particularly dangerous are retroactively updated statuses and actual results included in the signs. Carry out several consecutive checks over different periods, taking into account the required horizon. Compare separately to a simple, agreed-upon reference—for example, the last known period or comparable season, if appropriate. A complex model must prove its benefits under the same conditions.

Questions to check your forecast
QuestionWhat to record
What date is the forecast for?Cut available at that time
What horizon?Period of management decision
What are we comparing with?Simple reproducible landmark
Where are we going wrong?Products, periods and direction of error

Evaluate a mistake by its consequences

The overall error across the network may hide the poor prediction of an important group, and the average error rate becomes inconvenient at values of zero. Select indicators for the task and show an analysis by segments. Underestimation and overestimation can lead to different consequences: shortages or excess inventory. Do not pass off the range of uncertainty as a promise: its suitability is also checked against past estimates and facts. Keep manual adjustments separate from the model, with the reason and author. Then the team will be able to understand whether the expert intervention improved the outcome or only hid an inconvenient prognosis.

Start with solution support

In the first step, show the forecast next to the fact, a simple guideline, and signs of problems with the data. Leave purchasing or changing the plan the responsibility of the employee. Agree on when the estimate is recalculated, who resolves variances, and what happens if the source is unavailable. After several test cycles, it will become clearer where the model is useful and what limitations remain. The purpose of the launch is to improve a specific solution, and not to replace any work plan with an automatic number without an explainable history and verification of the result. Preparing history and checking the forecast are related todata analytics for business solutions.

Apply to your task

Name a report or decision that is being hampered by discrepancies. Let's clarify the sources, calculation rules and verification of the result.

Discuss the analytics problem →
← Return to materials