Demand Forecasting in Retail and Supply Chain: The Lever Most Companies Underestimate

Here's something we notice at nearly every client engagement in retail and supply chain: when you trace almost any operational problem back to its root cause, you eventually arrive at the same culprit.

Not the ERP. Not the warehouse. Not the team.

The forecast.

Demand forecasting has been around for decades. It's not new, it's not exciting, and most companies feel like they've already "handled it." But over the last few years, across dozens of projects with retailers, food manufacturers, and consumer goods companies in Belgium and the Netherlands, Crunch Analytics keeps finding the same thing: better forecasting is still one of the highest-value, most underexploited levers in the supply chain.

Here is what that actually looks like in practice.

Signs Your Demand Forecast Is Costing You Money

Most companies think they're fine. They have a process, a planner, maybe even a tool. Forecasting is "handled."

But ask yourself honestly:

• Do you occasionally run out of stock on products you should have had? Not the lumpy, unpredictable ones. The regular movers?
• Do you sit on inventory at the end of a season and have to discount it?
• When you run a promotion, do you know whether you ordered too much or too little? Or do you find out weeks later, after the fact?
• Does your planning team spend more time managing exceptions and firefighting than actually planning?

If yes to any of these, your forecast is costing you money. The question is how much.

The inconvenient truth is that most companies' baseline forecast is some variation of "same as last year" or "same as last week." Sometimes adjusted manually by a planner with years of experience. That works, to a point. But it's a heuristic, not a model. And heuristics don't scale, don't improve with data, and don't tell you when they're wrong.

At a leading retailer, the existing forecasting model was built on an 8-week rolling sales average. It was systematic and consistent, and it was systematically creating overstock on slow movers and stockouts on fast ones at the same time. Replacing that with a model that actually learned from the data patterns in their assortment led to lower inventory tied up in stock, fewer empty shelves, and a meaningfully improved forecast accuracy across their product range.

That's not a unique story. It's a pattern.

The Business Case for Better Demand Forecasting: Stockouts, Overstock, and Working Capital

The value of better demand forecasting isn't abstract. It shows up in four very concrete places.

1. Fewer Stockouts and Less Hidden Lost Sales

Out-of-stock situations are expensive in ways most companies underestimate. The obvious cost is the lost sale. The less obvious cost is the demand signal you just lost: when a product isn't on the shelf, customers buy something else or go elsewhere. Your data records zero demand, not frustrated demand. Over time, that distorts your view of what customers actually want.

Better forecasting reduces stockouts. But smarter forecasting also helps you identify where historical lost sales have already corrupted your demand signal, and corrects for it.

2. Overstock and Working Capital

Every unit of excess inventory on your shelves or in your warehouse is working capital you're not using elsewhere. Safety stock (the buffer you hold to account for forecast uncertainty) is directly proportional to your forecast error. Cut forecast error by 10%, and you reduce the safety stock you need to hold for the same service level.

For companies with large assortments and long supply chains, this isn't a rounding error. It's six figures in freed-up capital.

3. Production Planning and Procurement

For manufacturers, better demand forecasting has a different flavor. It's not just about what ends up on the shelf. It's about what you need to produce, when, and with which raw materials.

At a Belgian delicacy producer, disjointed planning systems were causing delays in raw material procurement and inefficiencies in production scheduling. Building a demand-driven planning layer allowed them to order the right ingredients at the right time, reduce stress on the production floor, and minimize costly downtime from changeovers and shortages.

The principle is the same: if you know what demand is coming, you can plan backwards. If you don't, you're always reacting.

4. Promotions and Pricing

This one is often overlooked. When you run a promotion, you need to separate two things: the baseline demand (what you would have sold anyway) and the promotional uplift (what you sold because of the discount or campaign). Without an accurate baseline forecast, you can't make that distinction.

The result? You either over-order and end up with promotional leftovers you discount further, or you under-order and leave revenue on the table. Worse, you don't know which one happened.

Better demand forecasting is the foundation of smarter promotional planning. It's also the bridge to the next frontier: understanding price elasticity at the product level, and using that to optimize your discount strategy. More on that in a future post.

Demand Classification: Why One Forecast Model Cannot Fit Your Entire Assortment

Here is where most forecasting implementations go wrong. A company buys a forecasting tool, runs it across their assortment of 10,000 SKUs, measures average accuracy at 72%, and declares success.

But averages lie.

Demand patterns are not uniform across an assortment. Some products sell every day in predictable quantities; forecasting those accurately is relatively easy. Other products sell sporadically, in bursts, with no discernible pattern. Forecasting those accurately is, in many cases, mathematically impossible.

If you're averaging accuracy across both types, you're hiding the real picture.

At Crunch, we classify every SKU into one of four demand types before we build a single model:

Smooth demand: Regular timing, stable quantity. These products sell consistently week over week. They're easy to forecast and respond well to standard ML models. This is where you should achieve high accuracy and use that accuracy aggressively to reduce safety stock.

Intermittent demand: The timing is irregular (the product doesn't sell every period), but when it does sell, the quantity is fairly consistent. Harder to forecast, but manageable with the right techniques.

Erratic demand: The product sells regularly, but the quantity jumps around. You can predict when demand will occur, but not how much. Accuracy will be lower, and safety stock needs to account for that.

Lumpy demand: Irregular timing and irregular quantity. Honestly? These products are largely unforecastable at the individual SKU level. That's not a failure of the algorithm. It's a property of the product. The smart response is to aggregate: forecast at a higher level (category, region, or time bucket) and disaggregate down.

Why does this matter? Because your response to each type should be different. Tight safety stock on smooth products. Higher buffers on lumpy ones. Aggregated forecasts where SKU-level forecasting breaks down.

In that engagement, classifying products by demand type was a core part of the work. It changed not just the forecast, but the inventory strategy, because different products need to be managed differently.

Common Demand Forecasting Challenges: Product Chaining, Lost Sales, and New SKUs

Demand classification gets you further than most companies go. But even with the right framework in place, several practical challenges trip up forecasting models that look good on paper.

Product chaining and reformulations

When a product is discontinued and replaced by a new SKU, the model loses all historical context. The new product code has zero history. Without intervention, the model treats it as a brand-new product and takes months to build reliable forecasts, even if it is essentially the same product under a new name or packaging.

Smart forecasting handles this with product chaining: identifying that Product A v2 is the successor to Product A v1 using name similarity, product characteristics, and sales pattern overlap, then combining their histories into a single continuous signal. It sounds like a detail. In practice, for retailers with frequent range refreshes or seasonal reformulations, it has a meaningful impact on accuracy from day one.

New products with no history

Truly new products present a harder version of the same problem. There is no history to chain to. Standard time-series models have nothing to train on.

The right approach is to use demand patterns from similar existing products as a proxy, matched on category, price point, and sales channel. Combined with any available leading indicators such as pre-orders or planned marketing spend, this gives the model a starting point. New product forecasting is a solvable problem, but it requires a deliberate strategy, not just running the same model you use on established SKUs.

Lost sales corrupting your training data

This is the demand forecasting version of a data quality problem. When a product is out of stock, your system records zero demand. But zero demand is not the same as zero need. Customers who could not find the product went elsewhere. Over time, if you train a model on uncorrected historical data, it learns that certain products have lower demand than they actually do, and systematically under-orders them.

The fix is lost sales imputation: reconstructing what you would likely have sold during stockout periods, and correcting the historical signal before the model ever sees it. Skipping this step means your "better model" is learning from a dataset that already has the wrong answer baked in.

Leading and trailing products

Not all products drive their own demand. Some products are leading products: they bring customers to the store or to the category. Others are trailing products: bought because the customer is already there. Think paint and brushes, or a printer and ink cartridges.

The forecasting implication is that demand for a trailing product is partly a function of demand for its leading product, not just its own sales history. Treating them identically in a model misses that dependency and can produce structurally wrong safety stock recommendations. The full commercial consequences of this distinction, for discounting and promotional strategy, go beyond forecasting, but recognising the pattern in your assortment is a useful starting point.

External factors: real impact, limited predictability

Demand does not happen in a vacuum. Weather, competitor promotions, public holidays, and broader economic conditions all move the needle. The impact is not marginal. When two major competitors run simultaneous promotions in the same category, sales might drop 20 to 30% in the affected week. A sustained cold spell shifts demand for seasonal categories by a comparable magnitude. These are not edge cases - they are recurring events that your forecast needs to account for.

The value of incorporating external factors is primarily retrospective: correcting your historical sales data so the model learns from actual demand, not demand suppressed by a competitor promotion you forgot to account for. Where external signals are available in advance - a competitor promotion already advertised, a public holiday on the calendar, a weather forecast for the next few days - they can also improve short-term predictions.

But this is where honest expectations matter. You cannot predict the exact weather two weeks from now, and you cannot predict what your competitors will decide to do next month. What you can do is build a model that recognises the patterns from past events and responds appropriately when similar conditions arise. That is not the same as forecasting the future with certainty. It is a better starting point than ignoring external reality entirely.

What a Modern Demand Forecasting Solution Delivers

A proper demand forecasting solution in 2026 is not a black box you plug in and trust. It's a system with clear inputs, transparent outputs, and measurable performance.

Here's what it delivers:

Speed of deployment. Our forecasting solution can be trained on your historical data in a matter of days and running in production soon after. You don't need a 6-month data science project. You need your data in the agreed format and a clear definition of what you want to optimize.

Granularity and aggregation. Predictions at the SKU level, automatically aggregated to any category, channel, or region you care about. The same model powers the operational team's daily decisions and the commercial team's monthly review.

Continuous improvement. The model retrains automatically as new sales data comes in. It gets smarter with every transaction. Unlike a static spreadsheet model, it adapts to shifts in demand patterns over time.

Honest performance measurement. Before deployment, you see how the model performs on held-out historical data and how it compares to your current forecasting approach. After deployment, performance is tracked continuously. You know, at all times, whether the model is adding value.

Context-awareness. Where data is available, the model can incorporate external signals: weather, promotional calendars, price changes, holidays. The more context the model has, the better it handles exceptional periods.

At a large food manufacturer, the challenge was improving the "On Time, In Full" (OTIF) rate for their customers, meaning the percentage of orders that arrived complete and on time. By building a forecasting solution that accounted for their product lines, customer ordering patterns, and supply chain variability, Crunch improved OTIF performance meaningfully. More customers got what they needed, when they needed it.