Demand forecasting helps businesses estimate what customers are likely to buy, when they are likely to buy it and in what quantities. For companies managing large product ranges, seasonal demand, changing customer behaviour or long supplier lead times, those estimates can have a direct impact on inventory availability and cost.
Good demand forecasting does not mean predicting the future perfectly. It means using the data available to create a reliable view of future demand, measuring how well that forecast performs and adjusting it as demand changes.
Modern forecasting combines statistical methods, artificial intelligence (AI), machine learning and human expertise. The result gives inventory teams a stronger basis for deciding how much stock they need, when they may need to replenish it and where future demand could create inventory risks.
This guide looks specifically at demand forecasting, including how forecasts are created, the different forecasting methods available, how accuracy is measured and how AI is changing the process.
What is demand forecasting?
Demand forecasting is the process of estimating future customer demand for a product or service.
It typically uses historical demand data alongside factors such as trends, seasonality, promotions and known changes in the market. Depending on the forecasting method, businesses can also incorporate additional information that helps explain why demand has behaved in a particular way.
The purpose is not to produce a number that will always be exactly right. Forecasting provides an informed estimate of future demand that can be used as the basis for inventory and supply chain decisions.
For an inventory team, that could mean understanding:
- How much demand an SKU is likely to experience next month.
- Whether an item follows a seasonal pattern.
- Which products are likely to experience increasing or declining demand.
- Whether recent sales represent a genuine change in demand or a temporary fluctuation.
- How much demand could occur during a supplier’s lead time.
- Where actual demand is consistently higher or lower than forecast.
Demand forecasting is therefore an analytical process. Demand planning is the broader process of deciding how the business should respond to expected demand.
That distinction is important. A forecast tells you what demand is expected to look like. A demand plan combines that forecast with commercial and operational information to determine what actions the business should take.
Read more about what demand planning is.
How does demand forecasting work?
Demand forecasting uses patterns in historical demand and other relevant data to estimate future demand.
Historical data provides the starting point. A forecasting model analyses that history to identify characteristics such as the underlying level of demand, trends, recurring seasonal patterns and changes in recent behaviour.
The appropriate forecasting method depends on the data and the behaviour of the item being forecast. A product with relatively stable demand may suit a different statistical model from one with strong seasonal peaks. Products affected heavily by promotions or other external variables may require a causal approach.
The forecast can then be compared with actual demand as new data becomes available. Measuring this difference helps businesses understand how well forecasts are performing and whether forecasts consistently run too high or too low.
This creates an ongoing cycle:
Historical demand → identify patterns → generate forecast → compare with actual demand → measure error and bias → update forecast
Modern forecasting software can automate much of this work across thousands of items. Instead of applying one forecasting method across an entire product range, technology can analyse demand at item level and help determine which approach best reflects each item’s behaviour.
Demand forecasting methods
There is no single forecasting method that works equally well for every product.
A highly seasonal item behaves differently from a steady seller. A newly launched product may have little or no historical demand of its own. A promotion can temporarily change a sales pattern without representing a permanent increase in underlying demand.
Forecasting methods generally fall into two broad categories: qualitative and quantitative.
Qualitative demand forecasting
Qualitative forecasting relies on judgement and market knowledge rather than historical data alone.
Inputs can include:
- Sales team estimates
- Expert judgement
- Market research
- Customer feedback
- Management expectations
- Comparable products
- Information about upcoming market changes
Qualitative forecasting can be particularly useful when historical data is limited, such as for a new product, or when a known event is expected to affect demand in a way that historical patterns cannot capture.
Its main limitation is subjectivity. Human knowledge can add valuable context, but it can also introduce bias. Qualitative input is therefore often most useful when combined with quantitative forecasting.
Quantitative demand forecasting
Quantitative forecasting uses numerical data and statistical techniques to identify patterns and estimate future demand.
Different models suit different demand patterns.
| Method | How it works |
| Moving average | Calculates an average using demand from a defined number of previous periods. It can smooth short-term fluctuations and reveal the underlying level of relatively stable demand. |
| Exponential smoothing | Gives more weight to recent observations while still incorporating older demand. This allows a forecast to respond to recent changes without ignoring the longer history. |
| Time-series forecasting | Analyses demand over time to identify patterns such as trends and recurring behaviour. |
| Seasonal forecasting | Identifies demand patterns that repeat at regular intervals, such as annual summer peaks or increases around Christmas. |
| Regression or causal forecasting | Examines relationships between demand and other variables, such as price, promotions or economic conditions. |
| AI and machine learning | Analyses item-level demand patterns and can automate tasks such as model selection, helping forecasting adapt across varied product ranges. |
These methods are not necessarily alternatives that a business must choose between permanently. A product range can contain thousands of items with very different demand characteristics.
This is one reason AI and machine learning have become increasingly useful in demand forecasting. They make it possible to analyse item-level behaviour at scale rather than relying on a single forecasting method for every SKU.
For more detail on forecasting calculations and techniques, see our guide to demand forecasting best practices and formulas.
How AI and machine learning improve demand forecasting
AI is increasingly becoming part of the forecasting process itself rather than a separate forecasting method used only in particular circumstances.
Machine learning can analyse large volumes of historical demand data, identify patterns and help determine which forecasting approach is most appropriate for an individual item.
This matters because demand rarely behaves consistently across an entire assortment.
One SKU may have stable year-round demand. Another may experience a strong annual peak. A third may sell intermittently, while a fourth has only recently been introduced and has little history of its own.
Applying the same forecasting logic to all four can produce poor results.
AI-supported forecasting can help by:
- Analysing item-level sales histories.
- Identifying trends and seasonal patterns.
- Detecting unusual observations and outliers.
- Automating the selection of appropriate forecasting models.
- Processing large numbers of SKU and location combinations.
- Supporting forecasting when historical data is sparse.
- Using relevant comparable items to support new item forecasting.
This automation does not remove the planner from the process. Commercial knowledge still matters, particularly when something is about to happen that historical data cannot know about.
The advantage is that planners do not need to manually analyse every SKU or spend their time repeatedly selecting and testing models. They can focus more attention on exceptions, unusual changes and decisions where their knowledge adds value.
Read more about AI in inventory management.
Demand vs. sales forecasting: What’s the difference?
Demand forecasting and sales forecasting are closely related, but they answer different questions.
Demand forecasting estimates the demand customers are likely to have for a product. Sales forecasting estimates the sales a company expects to make.
That distinction matters because actual sales do not always represent actual demand.
Imagine a retailer sells 80 units of an item before running out of stock. Recorded sales are 80 units, but customer demand may have been 100 units or more. Using sales alone without accounting for the stockout could underestimate future demand.
| Demand forecasting | Sales forecasting | |
| Focus | Expected customer demand | Expected company sales |
| Purpose | Support inventory and supply decisions | Support revenue and commercial planning |
| Typical data | Historical demand, trends, seasonality and external factors | Historical sales, pipeline and sales performance |
| Affected by | Customer behaviour, market conditions and seasonality | Availability, sales activity and commercial targets |
| Common users | Inventory, supply chain and procurement teams | Sales, finance and leadership teams |
| Primary output | Expected future demand | Expected future revenue or sales volume |
Both have value, but using them interchangeably can obscure what customers actually wanted to buy.
How demand forecasting supports inventory decisions
A demand forecast becomes particularly valuable when it informs inventory decisions.
The forecast itself does not determine an entire demand plan. Instead, it provides an estimate of future demand that can be combined with inventory data, supplier lead times, safety stock policies and other variables.
Inventory requirements
Forecasts help businesses understand how much demand they may need to cover over a given period.
This gives inventory teams a stronger basis for balancing product availability against the cost of carrying unnecessary stock. If expected demand rises, the business can prepare for that increase. If demand is declining, purchasing can be adjusted before excess inventory builds up.
Replenishment
Forecast demand can feed directly into replenishment calculations.
Expected demand during a supplier’s lead time is particularly important. Combined with current inventory, open orders, safety stock and ordering constraints, the forecast helps determine when more inventory will be needed.
This creates a much more responsive approach than relying only on fixed reorder quantities or historical averages.
Supplier visibility
Longer-range forecasts can also provide suppliers with greater visibility into expected future requirements.
This does not mean every forecast becomes a fixed purchase commitment. It gives suppliers and purchasing teams a common indication of likely future demand, which can support conversations about capacity, lead times and upcoming requirements.
Reduce the risk of the bullwhip effect
Forecast quality also affects what happens further upstream.
The bullwhip effect occurs when relatively small changes in customer demand become progressively larger changes in orders as they move through the supply chain.
Poor demand signals, delayed information and overreaction to short-term fluctuations can all contribute to this effect.
Better forecasting helps businesses distinguish meaningful changes in demand from ordinary variation, reducing the likelihood of reacting too aggressively to short-term changes.
What are the benefits of demand forecasting?
The value of demand forecasting is not simply having a more accurate number on a screen. It comes from making better decisions using that information.
Effective forecasting can contribute to:
- Better product availability: Inventory teams can anticipate expected demand and prepare stock accordingly.
- Lower excess inventory: Purchasing can respond when expected demand falls rather than continuing to order based on outdated assumptions.
- Better use of working capital: Less money needs to be tied up in inventory that is unlikely to sell soon.
- More effective replenishment: Future demand becomes part of the calculation of what needs to be ordered and when.
- Improved supplier visibility: Forecasts can provide an earlier indication of future purchasing requirements.
- Faster planning: Automated forecasting reduces the amount of manual analysis required across large product ranges.
- More informed decisions: Planners have a consistent, data-driven starting point while retaining the ability to apply their own knowledge.
The quality of those decisions still depends on the quality of the forecast. That makes measuring forecast performance an important part of the process.
What are the risks of inaccurate forecasting?
No forecast will be perfectly accurate. Problems arise when forecast errors are large, persistent or poorly understood.
Forecasting too high can result in excess inventory, higher carrying costs and capital being tied up in stock that is not needed.
Forecasting too low can contribute to stockouts, lost sales and lower service levels.
Persistent errors can also affect purchasing decisions and supplier requirements. If forecasts repeatedly overstate demand, for example, the business may continue ordering too much inventory even when the underlying demand does not justify it.
The aim should therefore not be to eliminate forecast error entirely. Businesses need to understand the size and direction of forecast errors so they can improve their models and make better decisions.
How to measure demand forecast accuracy
Forecast accuracy metrics compare predicted demand with what actually happened.
No single metric tells the whole story. Different measures reveal different aspects of forecast performance, which is why businesses often monitor more than one.
Mean absolute percentage error (MAPE)
Mean absolute percentage error expresses forecast error as a percentage of actual demand.
A simplified calculation is:
MAPE = average of |Actual demand − Forecast demand| ÷ Actual demand × 100
A lower MAPE generally indicates that forecasts are closer to actual demand.
Its percentage format makes it relatively easy to interpret, but MAPE becomes problematic when actual demand is zero or very low.
Mean absolute error (MAE)
Mean absolute error measures the average size of forecast errors in the original unit being forecast.
If the MAE for a product is 20 units, for example, forecasts differ from actual demand by an average of 20 units.
This makes MAE intuitive at SKU level, although it can be harder to compare across products with very different sales volumes.
Weighted mean absolute percentage error (WMAPE)
WMAPE measures forecast error while accounting for the relative volume of demand.
That makes it useful when assessing forecast performance across a product portfolio where one SKU might sell thousands of units while another sells only a handful.
Forecast bias
Accuracy tells you how large the error is. Forecast bias tells you which direction the error tends to go.
A forecast can appear reasonably accurate overall while still consistently forecasting too high or too low.
Persistent positive or negative bias can lead to repeated inventory problems. Tracking bias alongside forecast error helps identify whether the forecasting process systematically overestimates or underestimates demand.
Common demand forecasting challenges and how to address them
Even sophisticated forecasting models depend on the information they receive and how they are used.
Poor historical data
Missing transactions, incorrect product information and inconsistent data can distort the patterns forecasting models are trying to identify.
Clean, integrated data provides a stronger foundation for forecasting. This is particularly important when businesses are pulling information from multiple systems or relying heavily on manual spreadsheet processes.
Read more about building better inventory and supply chain data.
Stockouts distorting demand history
Sales history does not always equal demand history.
If an item was unavailable for two weeks, low sales during that period may reflect a lack of stock rather than a lack of customer demand. A forecasting model that treats those sales as normal demand could underestimate future requirements.
Recognising these distortions is therefore an important part of preparing historical data.
Outliers
An unusually large order or one-off event can create a spike that is unlikely to repeat.
If that observation is treated as normal demand, it can distort future forecasts. Identifying and appropriately handling outliers helps preserve the underlying demand signal.
New items with little or no history
Traditional forecasting relies heavily on historical data, which creates an obvious challenge when a product is new.
Comparable product histories can provide a useful starting point. AI can support this process by identifying relevant items and using their historical behaviour to help establish an initial forecast.
As the new product generates its own sales history, the forecast can increasingly reflect its actual demand pattern.
Seasonality
Seasonal demand is not simply a temporary spike.
Products may follow patterns that repeat annually, monthly or at other regular intervals. Recognising those patterns allows the forecast to anticipate recurring peaks and troughs rather than reacting after they happen.
Using one forecasting model for every item
Different products behave differently.
Applying one model across an entire assortment can therefore work well for some products and poorly for others. Item-level model selection allows the forecasting approach to reflect the actual demand characteristics of each SKU.
This is one area where specialised forecasting software can go beyond the more basic forecasting functionality available in many ERP systems.
Read more about the inventory management gaps an ERP may leave.
Excessive manual overrides
Planner knowledge can improve a forecast when teams know something the historical data does not.
Problems arise when forecasts are routinely overridden without evidence or without checking whether those changes improve results.
Tracking forecast accuracy before and after overrides can help businesses understand where human input adds value and where the statistical forecast performs better on its own.
Demand forecasting example: Muuto
Muuto is a Copenhagen-based Scandinavian design company with a complex product portfolio, global supplier network and a mix of retail and contract demand.
Before using AGR, forecasting and replenishment relied heavily on Excel spreadsheets and manual exports from its ERP system. The process required significant manual work and increased the risk of forecasting errors.
With AGR, Muuto automated more of its forecasting and replenishment process and gained a clearer view of future purchasing requirements. Forecasts also became useful for supplier collaboration, helping the team communicate expected demand further ahead.
Weekly stock ordering that previously took around two hours was reduced to approximately 15 minutes. Muuto also reported improved forecast accuracy and availability for important high-volume products, while category managers gained more time for supplier negotiations, product quality and assortment management.
As Michael Wong Min at Muuto explains:
“We use AGR for developmental forecasts and purchase planning. My forecasts are nearly 100% accurate every time.”
Read the full Muuto case study.
Choosing the right demand forecasting software
Forecasting across a small number of stable products can be relatively straightforward. Complexity grows quickly when a business manages hundreds or thousands of SKUs across multiple locations, channels or suppliers.
Demand forecasting software should make that complexity easier to manage rather than simply producing more data.
Useful capabilities to look for include:
- Item-level forecasting
- Multiple statistical forecasting models
- AI and machine learning capabilities
- Automated best-fit model selection
- Seasonality detection
- New item forecasting
- Forecast accuracy and bias monitoring
- ERP integration
- Exception management
- Clear reporting and forecast visibility
- Scalability across products and locations
Ease of use also matters. A sophisticated forecast provides little practical value if planners cannot understand it, investigate exceptions or see how it relates to the inventory decisions they make every day.
AGR’s demand forecasting capabilities
AGR combines statistical forecasting, AI and machine learning to help inventory teams forecast demand across complex product ranges.
Rather than applying the same forecasting model to every item, AGR analyses sales history at item level and applies best-fit statistical forecasting models according to the demand pattern.
AI supports the forecasting process further by helping clean sales history, identify unusual activity, recognise demand patterns and support forecasting for products with limited history. For new items, AGR can use relevant comparable sales histories to establish a forecast before enough product-specific data has accumulated.
AGR’s forecasting capabilities include:
- AI-supported demand forecasting
- Best-fit statistical models
- Item-level forecasting
- Seasonality analysis
- New item forecasting
- Forecast accuracy monitoring
- Safety stock calculations
- Inventory optimisation
- Automated order proposals
- Exception-based planning
- ERP integration
The goal is not to remove human judgement from forecasting. It is to automate the analysis that software can perform at scale while giving planners visibility and control over the resulting forecasts.
