Sales Forecasting Method: data, MAPE, and KPIs
May 31, 2026
AI-based sales forecasting estimates demand for each product, store by store, by separating baseline demand from promotional demand. AI takes in more data points than a spreadsheet, but its value lies solely in the decisions it helps inform.
In retail, this decision is primarily about pricing or a promotion. A useful pricing forecast—one that estimates how much will be sold at each considered price—is evaluated based on its bias rather than its MAPE, and is validated against control stores.
A sales forecast can be accurate and still be useless. This is the case whenever it projects volume at a constant price, even though the question being asked in the meeting is: Should we lower this price, and by how much?
This guide outlines the definition, key data points, methods, and metrics. Above all, it advocates for one position: in retail, forecasting and pricing decisions must be derived from the same model.

Sales Forecasting Methods: What Are We Talking About?
AI-powered sales forecasting estimates how much of each product will sell at each retail location over the coming days or weeks. It analyzes historical data, prices, promotions, the calendar, and external factors such as the weather.
Its primary use in retail: to distinguish between two types of demand that do not follow the same rules—the baseline demand and the uplift generated by a promotion or event.

Baseline and uplift: two forecasts, not one
The baseline is what sells without any specific action. The uplift is the additional volume generated by a promotion, a flyer, or a special event.
Mixing them up throws off everything else. A forecast that interprets the peak of a promotion as baseline demand leads to overordering the following week, resulting in excess inventory.
What AI Is Really Changing
AI doesn't predict the future. It processes thousands of product-store combinations at once and takes into account factors that a spreadsheet overlooks: net price paid, long weekends, school breaks, and the weather.
The real change lies elsewhere. Human judgment no longer focuses on the calculation itself, but on the explanation behind the number and the decision that follows from it.
According to Gartner, large organizations will have adopted AI-based demand forecasting by 2030. The question is no longer whether to adopt it, but for what decisions.
(Gartner, press release dated September 16, 2025)
Our position: A retail forecast must answer the question, “What if I change the price?”
In a retail store, the first decision based on a forecast is almost never an order. It’s a price, a promotion depth, or a markdown date.
However, most forecasts are designed for the supply chain. They answer the question, “How much will I sell?” at a constant price. To set a price, that answer isn’t enough.
A useful pricing forecast is conditional: it indicates how much will be sold at each considered price level. This is where price elasticity is incorporated into the model, rather than being calculated separately in another file.
A product sells 100 units per week at €2.49, with a purchase cost of €1.60. The standard forecast predicts 100 units: it is accurate, but it says nothing about the decision that needs to be made.
A price-sensitive forecast projects approximately 118 units at €2.29 (elasticity of -2, assumption). Volume increases by 18%, but the margin drops from €89 to €81 per week.
Without this data, the decline would have been assessed based on volume. With it, we know that at least 129 units must be sold for the price reduction to pay off.
Practical implication: Price forecasting and pricing decisions must be based on the same model. Using two different tools means two different interpretations of the same demand, which ultimately end up contradicting each other at the worst possible moment.
The Data Behind a Reliable Forecast, and the Pitfall of Disruption
The foundation is a company-specific sales history, broken down by product and store, including the actual price paid, discounts, and coupons. The required depth varies by category: see what sales history is needed to make accurate forecasts.
The first pitfall is stockouts. A failed sale is not the same as a failed order: if the product was out of stock, the model learns to forecast lower quantities and sets the stage for the next stockout.
Next come the calendar (movable holidays, long weekends, school breaks) and external factors, with the weather being the top consideration for the garden, beverages, and apparel. Not all of these deserve to be included in the model: it depends on the category.
For a new product with no history, the model draws analogies with "similar" products that share the same attributes.
Statistics, machine learning, expert judgment: integrating them, not choosing between them
There are three distinct categories of methods. None of them is sufficient on its own for a catalog containing several thousand items.
Moving averages, exponential smoothing, seasonal decomposition. Reliable for stable sales. Lose their effectiveness as soon as an event disrupts the usual pattern.
Gathers prices, promotions, weather, and schedules from a wide range of catalogs. Shifts human judgment from calculation to explanation.
A competitor goes out of business, construction in front of the store, a change in product listings. These are things to manage, not eliminate.
The real question isn't which one to choose, but who decides when they differ.
Our belief: Machine learning doesn't replace the expert; it shifts the nature of their work. The expert no longer performs calculations; instead, they point out what the data cannot reveal.
Evaluating a Forecast: MAPE, Bias, and What They Don't Tell You
MAPE measures the average percentage error. It is easy to communicate, and that is its drawback: it gives the same weight to a product that sells 2 units per week as to a loss leader that sells 2,000.
Estimating 3 instead of 2 results in a 50% error for a single unit of deviation. Estimating 2,100 instead of 2,000 results in a 5% error for 100 units. Take the average, and the long tail determines your score.
We prefer to look at a volume- or margin-weighted error (WAPE), and especially the bias. A forecast that is wrong in both directions is less costly than one that is always wrong in the same direction.
A sustained positive bias fills the reserves. A negative bias empties the shelves, then censors history and perpetuates itself.
The ultimate judge is business: break-even points, markdowns, and margins. A backtest proves that a model can interpret past data, not that it improves decision-making.
The proof is in the results. For this project, every price change was compared to control stores on a weekly basis before any expansion took place.
Each price change is measured against a control group.
The new prices were first rolled out in a few pilot stores and compared to control stores. The results are reviewed weekly and supplemented by an in-store price perception survey.
- Issue
- Move away from a robust but manual method using Excel across multiple retail chains.
- What Has Changed
- Pilot program with a control group, benchmark indices, and capped price fluctuations.
customers surveyed in-store to gauge their perception of the pilot's price.
Pricing Department · Food Retail Group. Case study written by Booper based on an actual client project; the client has been anonymized.
Where Sales Forecasting Pays Off First in Retail
The uses are well known: determining the size of a promotion, reducing stockouts and excess inventory, managing end-of-season markdowns, adjusting prices, and allocating inventory among stores.
Our approach: start where decisions are frequent, reversible, and quantifiable. Promotions check all three boxes. Product assortment rarely does.
A promotion based on a realistic uplift prevents both sell-outs on the first day and the need to sell off remaining inventory at a discount afterward. However, it is essential to measure the incremental impact of a promotion, not just its volume.
Each use case has its own data and metric:
| Use Cases | Key Data | Key Performance Indicator | Business Profit |
|---|---|---|---|
| Promotion | Promotion history, schedule, net price | Uplift / MAPE | Campaign Profitability |
| Restock | Real-time inventory, supplier lead times | Service Rate | Zero stockouts on the shelves |
| Markdown | Local sales, idle inventory | Gross Margin | Optimal markdown of End-of-Season Inventory |
| Pricing | Price elasticity, competition | Revenue | Dynamic Pricing |
| Allowance | Sales by Region, Product Attributes | Inventory Turnover | Multi-Site Inventory Balancing |
Why Forecasting Projects Fail on the Organizational Side
The model is rarely the weak link. What causes a forecasting project to fail almost always has to do with how the process is organized.
- No one owns the figure. A forecast without a business owner is disregarded the first time it contradicts on-the-ground intuition.
- Plans are never compared with actual results. Without regular reviews, no one knows whether the forecast is improving or getting worse.
- A budget and a forecast are one and the same. A single figure cannot be both a commitment and an estimate: see sales forecast and budget.
- Each department has its own forecast. Sales, purchasing, and supply chain each base their decisions on their own interpretation of demand, and the category manager is left to deal with the resulting contradictions.
Booper's sales forecasting software projects demand over several weeks based on three scenarios—conservative, balanced, and aggressive—rather than a single figure presented as a certainty.
Each forecast is accompanied by an “AI explanation” that lists the factors contributing to the result, including weather conditions. The projected price is simulated before being applied: impact on sales, margin, and inventory.
Where to Start
Using a test category and a specific question, not a data project covering the entire catalog. A category with frequent promotions, a reasonable history, and a business metric selected in advance.
Then comes the expansion, once the gain has been measured against a control group. Gain is measured in terms of prevented breakages, reduced markdowns, and profit margin—not in terms of model scores.
An AI prediction is only as valuable as the decision it informs. In retail, that decision is, first and foremost, a price.
An accurate forecast is useless if it doesn't tell you what price to set. Let's see how to link your forecasts to your pricing and promotional decisions.
See also in this feature
- Comparing Sales Forecasting Solutions: 5 Questions to Ask Before the POC
- External Events: How to Incorporate Them into a Sales Forecast
- Who Should Lead Sales Forecasting? Sales, Purchasing, and Supply Chain Governance
Frequently Asked Questions About the Sales Forecasting Method
It is a forecast of future demand for each product at each retail location, generated by machine learning models. These models analyze sales history, prices, promotions, the calendar, and external factors such as the weather.
It does not provide certainty but rather a probability, ideally in the form of scenarios. Its value in retail lies in distinguishing between baseline demand and promotional demand, and in showing what would happen if the price were changed.
Statistical methods (smoothing, seasonal decomposition) remain reliable when sales are stable. Machine learning takes over as soon as the number of signals increases: promotions, competitors' prices, weather, and a broad product catalog.
The two are not mutually exclusive. A good system combines them and leaves room for expert judgment regarding what the data cannot reveal.
The baseline is a product's natural sales volume, without any marketing efforts. The uplift is the additional volume generated by a promotion or event.
Separating them allows us to base the next promotion on a realistic uplift and prevents a promotional spike from being misinterpreted as a sustained increase in demand, which can lead to excess inventory.
A sales history specific to the product-store combination, the actual price paid, past promotions, inventory levels to identify stockouts, and the schedule.
Weather, traffic, and competitors' prices are then added, category by category. Imperfect data doesn't prevent you from getting started, as long as you know where it's inaccurate.
A week with a sales gap shows low sales even though demand existed. The model should reconstruct this demand—for example, based on similar products or data from previous weeks—rather than interpreting the dip as a loss of interest.
Without this correction, the forecast decreases, orders follow suit, and the shortage occurs again. The mechanism is explained in detail here.
MAPE (Mean Absolute Percentage Error) is simple but overweights small SKUs. WAPE weights the error by volume and better reflects the actual impact on inventory and margin.
Bias indicates whether the forecast is consistently wrong in the same direction. It is often the most useful indicator: a persistent bias is more costly than a random error.
There is no universal figure: the return depends on the decision that the forecast informs. The initial gains are generally seen in promotions and markdowns, because these decisions are frequent and measurable.
The best approach is to measure the gain against control stores in a test category. In terms of pricing, teams working with Booper have seen margin increases of +0.5 to +3 points in just a few months (Hugues Lafitte, Informations Entreprise, September 2026).
Yes, whenever the forecast is used to determine a price or a promotion. A constant-price forecast tells us how much we’ll sell if nothing changes—not what will happen if we lower the price by 20 centimes.
Incorporating price means incorporating price elasticity into the forecasting model. Pricing decisions and forecasts are then based on the same interpretation of demand.
Well-defined business rules are often sufficient for a narrow, stable, and loosely interdependent product catalog. AI becomes cost-effective when decisions number in the thousands, products influence one another, and price sensitivity varies.
But you still need a usable sales history. In most cases, the right answer is a hybrid approach: the rules set the framework, and the model optimizes within it.
AI handles the volume at every stage of the data journey: collection, product matching, verification, forecasting, and simulation. The expert selects competitors, resolves ambiguous cases, provides context, formulates the strategy, and makes decisions.
At Booper, Booper Mind makes suggestions, MPS monitors, and the user decides. Knowing how to interpret elasticity and question a forecast is more important than knowing how to code.
Dynamic pricing is permitted in France. Personalized pricing based on automated decisions is allowed, but customers must be notified of it. Algorithm-based price-fixing remains prohibited, even in the absence of contact between competitors.
The strongest safeguard for a retailer: being able to say, for each price, which rule generated it, what data it was based on, and who approved it.
.avif)