AI Demand Forecasting: How to Forecast Business Trends

AI demand forecasting is the use of machine learning on sales history and other signals to estimate how much customers will ask for in a future period. To forecast business trends this way, work through six steps in order. This guide gives the six steps, the decision to make at each one, a way to test a model on past periods and a way to review the result. The lead example is demand for one product line, and a sales target follows as a second example. The guide sits under the Signisys use case on trend forecasting.

AI Demand Forecasting: What You Need Before You Start

Three terms recur below:

• Demand: the quantity of a product or service that customers ask for in a period.
• History: the demand recorded for past periods.
• Forecast horizon: how far ahead the forecast looks.

IBM’s step-by-step demand tutorial describes demand forecasting as a supply chain planning process built on past sales, market trends and outside economic signals. It lists qualitative methods, such as the Delphi method and market research, beside quantitative ones, such as moving averages, exponential smoothing and regression.

Before step 1, have three things in hand:

• the demand history for the item you plan to forecast;
• the dates of past promotions, closures and stock-outs;
• a person who will use the forecast to make a decision.

If one of the three is missing, find it first. Without a person who uses the result, for instance, the later steps produce a figure that nobody acts on.

AI Demand Forecasting in Six Steps

The six steps below follow the order of IBM’s tutorial, which also has six steps: define the forecast period, collect and clean the data, create a baseline forecast, create adjusted forecasts, communicate and integrate, and monitor performance. The sequence is IBM’s, not ours. In Signisys’ working method, three specifics are added: a baseline set by a plain rule, a test on past periods that sets the model’s gap beside the baseline’s, and a published low, expected and high range. Each step has its own section after the list.

Step 1: Name the Decision the Forecast Serves

Start with the person who acts on the forecast and the date they act. The decision might be a purchase order, a staff rota or a production plan. IBM’s tutorial advises matching the forecast horizon to that decision cycle, and it separates short-term from long-term forecasts and macro from micro levels.

Then choose one level to forecast: the whole firm, a product line or a single item. Begin with one level, and add others once the first one works.

IBM also separates a passive forecast, built from historical sales data alone, from an active one that also allows for planned growth, campaigns and expected moves by competitors. Choose the one that fits the decision. For instance, a buyer placing a monthly order for one product line may start passive and add promotions later.

Output of this step: one written line with the decision, the horizon and the level. As an example: “Monthly order quantity for the kitchenware line, three months ahead, owned by the purchasing lead.”

Step 2: Gather the History and Mark the One-Off Events

Collect demand for the same item and level for each past period the firm holds. Note the period length, week or month, and keep it the same throughout. This guide sets no number of periods as enough; that depends on the item and the decision.

IBM’s tutorial describes this stage as collecting, cleaning and analysing the data, and it gives removing a one-time bulk order as an example of cleaning. In practice, marking comes before removing:

• Mark each past period that held a large single order, a closure, a promotion or a stock-out.
• For each mark, decide whether to keep the period as it is, adjust it or set it aside.
• Write that decision and the reason beside the period.

Keep the original figures as well. Anyone can then revisit a decision to adjust. Fixing wrong or duplicate values is wider work than this step. For that, see the use case on improving data quality.

Output of this step: a history sheet with each one-off event marked and a decision beside each.

Step 3: Set a Simple Baseline First

Before any model, forecast the next period with a plain rule the team can explain. Two plain rules are the same period last year and the average of the last few periods. IBM names moving averages and exponential smoothing among quantitative methods. Write the baseline forecast down and keep it.

The baseline matters for one reason: the team judges any later model against it. If a model does not beat the baseline on past periods, the extra work has not earned its place, and the baseline stays in use. That is reasoning, not a measured result.

Run the same rule for each past period you plan to test in step 4. Then both forecasts cover the same periods, and the comparison is fair.

Output of this step: a baseline forecast for each past period you will test.

Step 4: Add a Model and Test It on Past Periods

Here a model joins the work. Choose it together with the team that will maintain it, and ask three questions. Which inputs will it use beyond history, such as promotions or the calendar? Who can explain a figure it produces? Who will retrain it? IBM’s explanation of AI forecasting lists model development, evaluation and performance measurement, deployment and integration, and monitoring and continuous improvement as stages of the cycle.

Then run the test. Hide the last few periods, forecast them with the model and with the baseline, and compare both with the recorded figures. IBM describes backtesting, which tests models on earlier time periods, and checking whether a model leans too high or too low. For each test period, write the gap and whether the forecast was over or under.

Do not move to step 5 until each test period has the model’s gap and the baseline’s gap in writing. If the model does not beat the baseline, keep the baseline and record that choice.

Output of this step: a test table with the period, the baseline, the model and the recorded figure.

Step 5: Publish the Forecast with a Range

Publish three figures: a low, an expected and a high. Beside the low and the high, say what would move demand to that end, such as a promotion or a closed delivery route. A single number hides the spread that the test table showed. In Signisys’ working method, the width of the range comes from the size of the gaps in step 4.

Next, decide who receives the forecast and in what form. A one-page sheet next to the monthly report keeps the range in view of the person who acts on it. Add the date of the next forecast to that sheet as well. If those reports come together on a schedule, see the use case on automating management reporting.

Output of this step: a published range, its owner and the date of the next forecast.

Step 6: Review the Forecast Against What Happened

When the period closes, write the recorded figure beside the forecast, then the gap and its direction. Add a note on what happened, such as a one-off event or a change that nobody had fed into the inputs. Keep all of this in one log, with a named owner and a review date beside it.

Each review ends with one of three decisions: keep the forecast as it is, adjust the inputs, or return to step 2 and look at the history again. IBM’s cycle ends with monitoring and continuous improvement, where models are updated and retrained when accuracy slips or the data changes.

Compare each forecast with the figure recorded for its own period, not with a later forecast of that period. A later forecast has drawn on more data, and comparing against it can hide the original miss.

Output of this step: the log, and the next action with its date.

The Same Six Steps Applied to a Sales Target

For a second example, take an invented services firm whose sales team forecasts next quarter’s new orders to set targets. The firm exists only to show the method; it is not a Signisys client. Here, AI sales forecasting follows the same six steps:

1. The decision is the quarterly target, which the sales head sets.
2. The history is orders by month, with a single large contract in one month marked as a one-off.
3. The baseline is the same quarter last year.
4. The test hides the last two quarters from the model.
5. The range goes to sales leadership together with the target.
6. The sales head reviews the log at quarter end, and the next target starts from it.

Only the decision and the history differ from the demand example. A team can use the same steps to forecast business trends in sales or in demand.

AI Demand Forecasting: Decisions the Six Steps Cannot Make for You

Some choices sit outside the six steps. They do not pick a tool or platform, and they do not say how much history a given item calls for. Whether the firm’s data is ready and why forecasts miss in general are separate subjects, as are the choice between analytical approaches and rules for cleaning data. This guide is a method, not an audit, and it promises no result.

Ask someone outside the team to look when nobody can say who acts on the forecast. The same applies when the test table shows no gain over the baseline and the team still wants a model.

Signisys provides data and AI services through its business intelligence and analytics solution and its data and AI service page. This box describes a Signisys service, not neutral advice, and it makes no promise.

Frequently Asked Questions