Choosing sales forecasting software: 

5 Questions to Ask Before the POC

Choosing Sales Forecasting Software: What Are We Talking About?

Choosing sales forecasting software means selecting a tool that will estimate future demand for each SKU at the appropriate level of detail and explain the reasoning behind its estimates.

The number alone isn't enough. A sales forecast is only valuable if it influences a decision: a price, a promotion, an order, or a markdown.

Our belief as practitioners is simple:a forecast is only as valuable as the decision it helps inform. A highly accurate model that no one uses is of no use.

This article is intended for pricing, sales, procurement, and category management teams that are comparing software vendors before conducting a trial. The calculation method itself is covered in our guide , “AI-Powered Sales Forecasting: Methods and Metrics.”

Here, we ask the five questions that truly set solutions apart—and what each answer reveals.

Why a demonstration does not settle the debate over forecasting methods

In the demonstration, all the curves match the historical data. That's to be expected: a model fitted to past data can always reproduce it.

The real question lies elsewhere. How does the solution handle your data gaps, your data breaks, your past promotions, and your new products? And how does it handle its own errors?

Not all of these losses are due to forecasting. But a forecast that is off in one direction fills the inventory, and in the other direction empties the shelves.

Hence our approach: to ask a publisher about the ways in which their tool makes mistakes, rather than about the beauty of its graphs.

Question 1: How does the tool handle demand masked by shortages?

An empty shelf means zero sales. Yet demand did exist. So your sales history doesn't measure demand—it measures what you actually managed to sell.

This is known as a censored request. A model that learns from this history without corrections underestimates the number of items that were out of stock, recommends ordering fewer items, and causes the next stockout.

The mechanism is explained in detail in our article on the sales history needed for accurate forecasting. Its impact on the sales floor is described in the section on the link between stockouts and poor forecasting.

When dealing with a publisher, the crucial question isn't "Do you know how to handle stockouts?" Everyone will say yes.

Instead, ask how the tool detects a data gap when your inventory data is incomplete, and what it uses to fill in the zero. A vague answer on this point suggests that the forecast will be biased downward.

Can forecasting software reduce both stockouts and excess inventory?

Yes, provided that the historical data is adjusted in both directions. Data gaps cause demand to be underestimated; past promotions and clearance sales cause it to be overestimated if the tool does not filter them out.

A solution that addresses only one of the two biases merely shifts the problem. It trades off stockouts for markdowns—or vice versa.

Question 2: What stitch size should I use, and why that one?

A national forecast by product category may be accurate on average but unusable everywhere. Decisions are made at a more granular level: by product line, retail location, week, or day.

The right allocation depends on the decision made. Prices and promotions are often determined by product line and by group of retail locations; orders are determined by product line, by location, and by day.

In B2B distribution, the relevant metric is typically the number of SKUs per customer or per agency. In retail pharmacies, it’s the number of SKUs per pharmacy, where low volumes make each sale more significant.

Is there a demand forecasting tool designed for multi-store operations?

This is one of the most frequently asked questions on the subject. A demand forecasting tool designed for a network performs calculations at the point-of-sale level and then consolidates the results, rather than distributing a national figure.

Ask the publisher at what level of detail it actually performs its calculations, and at what level it simply aggregates or distributes data. Then ask how it handles slow-moving items, where the relative error skyrockets.

A reliable solution recognizes that a mesh that's too fine produces noise. It knows to move up a level when the data is too sparse, and it lets you know.

Question 3: Can the solution explain a number?

A buyer who doesn't understand a forecast corrects it manually. After a few weeks, the tool is left idle, and the spreadsheet makes a comeback.

A useful forecast is backed by reasons: seasonality, a past promotion, a holiday, a price change, or the weather.

Signals that are external to the historical data are discussed in our article on exogenous events in sales forecasting.

At Booper, each forecast includes an explanatory section that lists the factors taken into account, including the weather. And it is presented as three scenarios—conservative, balanced, and aggressive—rather than a single number.

Why three scenarios? Because an order, a promotion, and a price reduction involve different levels of risk. The discrepancy between the scenarios also indicates when the forecast is uncertain.

The issue goes beyond forecasting: it is the same as for any calculated recommendation, as our article on the explainability of AI-generated price recommendations shows.

Question 4: Is forecasting linked to your pricing and promotional decisions?

This is the issue that specifications most often overlook. In many companies, forecasting is handled by the procurement department, while pricing decisions are made elsewhere, based on different figures.

However, demand depends on price. A forecast that fails to take into account next week's planned promotion—or next month's price increase—is bound to be wrong from the start.

Ask the vendor via this link. Can their solution simulate the impact of a price change on sales, margin, and inventory before a decision is made? Does it factor inthe price elasticity of each SKU?

When it comes to promotions, the forecast must distinguish between base sales and promotional sales, and measure the impact of the promotion on sales of other products. We go into detail on this in “Simulating a Promotion: Incremental Sales, Cannibalization, and Margin.”

When it comes to markdowns, the forecast for remaining inventory at the end of the season determines the schedule for discounts: in other words, estimate your remaining inventory before clearing out your stock.

What ROI can you expect from sales forecasting software?

Be wary of any percentage figure given before you've seen your data. The return on investment is measured in improved decisions: fewer out-of-stocks on key items, fewer unplanned markdowns, and better-targeted promotions.

It also depends on who is responsible for the forecast. When sales, purchasing, and the supply chain each keep their own figures, no tool can make the decision for them: that’s the topic of our article on sales forecast governance.

Also, don't confuse a forecast with a goal. The budget specifies what we want to sell, while the forecast indicates what we will likely sell, as explained in the section on the difference between sales forecasts and budgets.

Question 5: How can we accurately measure the accuracy of a forecast?

MAPE, or mean absolute percentage error, is the most commonly cited metric. It is also the easiest to make look better than it actually is.

When calculated on an aggregated basis by family, on a monthly basis, using stable benchmarks, it shows an excellent result. When calculated by benchmark, by retail location, and by week, it tells a different story.

Three requirements make all the difference. First, a test using historical data that the model has never seen—known as a backtest—rather than an adjustment based on data that is already known.

Next, measuring bias: a model that is always wrong in the same direction is more costly than one that makes random errors. Finally, a volume- or value-weighted analysis, so that the references that matter carry the weight they deserve.

Ask the publisher to compare its model to a simple benchmark, such as sales from the same week last year. If it doesn't perform better than that benchmark based on your data, the rest doesn't matter.

A technical detail does not necessarily prove a benefit.

Test a forecasting tool before committing to it

The five questions above are answered based on your data, not on a demo. A test is only as good as its data: our article on the data to prepare for a POC explains why results that seem too good to be true should raise a red flag.

But there’s no need to wait for an IT project. At Booper, you can get started right away using your existing exports—point-of-sale data, pricing, and catalog—without needing a connector: see how to integrate your data without a major IT project.

One last point: check whether the forecasting tool uses the same data source as your pricing, promotions, and markdowns, or if it adds yet another tool to the mix. We believe in consolidating pricing, promotions, and markdowns into a single tool.

Here's how our AI-powered sales forecasting softwareworks: forecasts by SKU and by store, complete with their underlying factors and scenarios, linked to price and promotion simulations.

FAQ: Choosing Sales Forecasting Software