Demand Forecasting with AI: A Practical Guide for Retail and CPG

AI demand forecasting uses machine learning to process far more variables than traditional statistical models. It can pull in things like point-of-sale data, promotions, weather, pricing, and other signals, and then keep updating the forecast as new data comes in instead of waiting for the next planning cycle. For retail and CPG leaders, that can mean fewer stockouts, less pressure to mark down products, and inventory decisions based on what’s happening now, not assumptions from the last quarter.

Key Takeaways

• AI forecasting earns its value by combining more data sources and recalibrating continuously, but it still depends on clean, well-governed data to perform.
• Demand sensing and predictive forecasting solve different problems and work best as complementary layers within one planning environment.
• SKU segmentation, which decides which products need which modelling approach, is an important driver of forecasting performance because different demand patterns require different modelling approaches. 
• UK and EU obligations apply selectively: GDPR matters where models use personal data, while sector rules such as EU food traceability law and the incoming Digital Product Passport requirements apply to specific categories.

A cold snap in one region, an unplanned promotion in another, a viral moment on social media: demand today moves faster than a monthly S&OP cycle can track. 

Retailers and CPG manufacturers running AI-driven forecasting can have a headstart ahead of those still relying on legacy systems, and the gap widens each planning cycle. This guide covers machine learning in demand forecasting, predictive demand planning, retail demand forecasting, AI supply chain forecasting, inventory forecasting, and how AI demand forecasting works, where it delivers genuine value across retail and CPG, what it requires to succeed, and how to judge readiness.

Where Traditional Forecasting Falls Short in Modern Retail and CPG

Spreadsheet-based and classical statistical forecasting remain genuinely useful, particularly for stable, high-volume categories, and many mature planning teams still run them as one layer within a broader ensemble approach. The gaps tend to appear in specific conditions rather than in all cases.

IHL Group's 2025 research puts the annual global cost of retail inventory distortion, the combined toll of out-of-stocks and overstocks across all root causes, at $1.73 trillion, roughly equivalent to South Korea's entire GDP. Supply chain disruption accounts for over $300 billion of that total, with the remainder spread across personnel issues, supplier problems, and planning and execution gaps, of which forecasting is one part. (Source)

Time-series models built on moving averages and seasonal indices work well in predictable environments. Where these models come under more pressure is in categories with heavier promotional complexity, frequent supply disruption, or demand shifts driven by social trends with little precedent in historical sales data. In those conditions, a purely statistical approach benefits from being supplemented rather than replaced outright.

Four failure modes tend to recur across retail and CPG planning teams:

• Out-of-stocks on fast-moving items during unplanned demand spikes, often the result of forecasts built on shipment history rather than true consumer offtake.
• Overstock on slower-moving items ties up working capital and drives markdown costs.
• Promotional misfires, where uplift is either badly overestimated or badly underestimated because cannibalisation and pull-forward effects sit outside the model.
• Retailer fill rate failures, where a supplier's internal forecast diverges from what retail partners actually need on the shelf.

The financial impact of forecasting badly extends well past a missed sales target. It shows up in working capital tied to excess stock, in customer satisfaction scores when shelves run empty, and increasingly in sustainability reporting, given how much waste stems from overproduction and markdown clearance.

What AI Demand Forecasting Actually Does

AI demand forecasting processes a wider set of signals simultaneously, and by learning the interactions between them, something classical statistical models were never built to do.

Where a traditional model might factor in historical sales and a seasonal index, a machine learning model can draw on POS data, promotional calendars, weather, pricing, macroeconomic indicators, social sentiment, and competitor activity, all at once. Crucially, it learns the relationships between these variables: how promotional response differs by region, how price elasticity varies by SKU, and how a promotion on one product pulls volume from a substitute on the next shelf.

Unilever's ice cream division can be an apt real-life example. By feeding weather data into its AI-driven forecasting models, the team improved forecast accuracy by 10% in Sweden, giving production planners a clearer view of expected monthly volumes and allowing production lines and cold-chain logistics to be adjusted ahead of demand spikes rather than reacting to them. (Source).

Recalibration is the other structural shift. Depending on the use case, the data available, and how the system is architected, AI forecasting can update daily, weekly, or on a trigger basis as new data arrives, rather than sitting fixed for a full monthly or quarterly cycle. The right update frequency depends on the decision it supports: a long-range production plan requires less frequent recalibration than a short-term replenishment order.

Where AI Forecasting Still Needs Human and Data Foundations

AI models perform best when the inputs feeding them are solid. Forecast quality depends heavily on data quality, consistent SKU hierarchies, complete promotional records, and planner confidence in the outputs. Most demand forecasting software today offers some form of ensemble or multi-model capability, so the meaningful differentiation between implementations tends to come from elsewhere: how well suited the data foundation is to the SKU mix, how carefully models are selected and tuned by segment, how cleanly systems integrate, and how well forecast outputs actually connect to replenishment, allocation, and production decisions.

Demand Sensing vs Predictive Demand Forecasting: Understanding the Difference

Predictive demand forecasting projects demand weeks, months, or seasons ahead, using historical sales, promotional plans, and seasonal patterns to support strategic decisions such as production scheduling and long-range inventory positioning. Demand sensing works on a much shorter horizon, typically days to a few weeks, drawing on current signals such as recent POS data, local weather, and web traffic, updated as frequently as the data feeding it allows, to adjust that baseline as conditions change.

Demand sensing and predictive demand forecasting answer different planning questions, and confusing the two is one of the more common mistakes teams make when evaluating AI forecasting tools.

In a mature retail or CPG planning environment, the two work together. Predictive forecasting sets the baseline plan guiding supplier commitments, capacity, and budget. Demand sensing then continuously adjusts that baseline as real conditions unfold, closing the gap between what was planned and what is actually happening on the shelf.

But remember, demand sensing tools are sometimes positioned as a full replacement for forecasting, even though their short data window offers limited visibility into long-range strategic decisions. Treating sensing as a layer on top of a solid predictive baseline produces stronger results than treating it as a substitute.

Where AI Demand Forecasting Delivers Value in Retail

Retailers see the clearest returns from demand forecasting where SKU counts are high, promotional activity is frequent, and the cost of getting inventory wrong compounds across stores.

• SKU-level forecasts: predicting demand at an individual product/store level helps identify patterns that could be masked if you’re only working with category- or region-based averages, leading to better replenishment.
• Promotion forecasts: using machine learning models allows businesses to distinguish between incremental volumes actually driven by a promotional push versus volumes moved over from other timescales or cannibalised by competing items.
• Seasonal forecast optimisations: AI models can identify changing ‘windows’ of seasonality (which makes sense given climate change and how consumers spend).
• New product introduction: Where there’s no historical sales data to go off of, these models use analogous product performance as well as early indications from sellers (sell-through) to develop a forecast more quickly than would otherwise be possible through manual means. New product introduction
• Inventory forecasting for e-commerce: Online channels generate a lot of real-time behavioural data, giving retailers more frequent signals to work with. This helps forecasting models respond faster and supports quicker fulfilment and inventory decisions.

Use case: A leading retailer was using spreadsheets for demand forecasting, which made the process slow and inconsistent. Tredence built more than 10,000 ML models tailored to different product types. The result was a 600-basis-point improvement in forecast accuracy, 50% less time spent by planners, and a 6% improvement in seasonal inventory costs. (source)

Where AI Demand Forecasting Delivers Value in CPG

CPG manufacturers face a different planning challenge to retailers: they sit one or more steps removed from the end consumer, which makes demand signal quality and retailer collaboration central to forecasting performance.

• Launch planning: Forecasting demand for a product with no sales history calls for models that draw on comparable launches, category trends, and early distribution data rather than historical time series alone.
• Supply alignment: Connecting the demand forecast to production scheduling and raw material procurement keeps manufacturing responsive to real market conditions.
• Trade promotion planning: AI models help quantify promotional lift and cannibalisation effects across a portfolio, providing clearer visibility of genuine ROI to support trade spend decisions.
• Retailer collaboration: Sharing forecast and POS data with retail partners, where commercial agreements allow, improves alignment on replenishment and narrows the gap between shipments and true shelf demand.
• Demand volatility management: For categories exposed to weather, viral trends, or economic sensitivity, continuous recalibration keeps forecasts closer to current reality than a static monthly model.

Use case: Colgate-Palmolive used Walmart’s Scintilla insights for their holiday oral-care gift packs. They combined data on past sales, trends, and future demand to find the best stores. This smart plan led to huge sales jumps, massive unit growth, and a 90 to 95 per cent sell-through rate. (Source)

The SKU Segmentation Problem Most Implementations Get Wrong

Chasing a single accuracy number across every SKU in the catalogue can be a costly mistake in AI demand forecasting implementations, and it is a mistake worth addressing before model selection begins.

Stable, high-volume SKUs forecast tightly almost regardless of method, since their demand patterns repeat with reasonable consistency. Long-tail items, new product introductions, and heavily promoted SKUs behave very differently, and holding them to the same accuracy bar sets teams up for disappointment and quietly erodes planner trust.

The more useful question is narrower: on the SKUs where a better forecast genuinely changes a costly decision, how accurate does the forecast need to be, and is current performance good enough to act on?

Measuring Forecast Accuracy the Right Way

Measurement works best at two levels, and treating them separately gives planning teams a clearer view of whether the forecasting programme is actually working.

The first level is statistical: forecast accuracy metrics such as MAPE or WAPE and forecast bias, which tracks whether a model systematically over- or under-forecasts rather than simply missing by a random margin. Both should be measured by SKU segment, since a strong aggregate figure can mask a stable core performing well while promotional or long-tail SKUs quietly underperform.

The second level is operational: whether improved statistical accuracy translates into better business outcomes. Useful indicators include stockout rate and overstock levels by segment, inventory turns and working capital tied up in stock, service level to retail partners, and promotion performance measured as genuine incremental lift rather than shifted volume. A forecast can look statistically strong and still fail to move these numbers if it does not connect cleanly to replenishment and allocation decisions, which is why both levels matter together.

A practical segmentation approach groups SKUs into categories such as:

• Stable core items with predictable, high-volume demand.
• Promotional SKUs where uplift and cannibalisation dominate the forecast.
• Seasonal products with recurring but variable demand windows.
• New product launches with limited or no history.
• Long-tail items with low, intermittent demand.

Each segment benefits from a different modelling approach and a different accuracy expectation. Defining this segmentation before selecting a forecasting platform, rather than after, keeps the technology decision grounded in the actual planning problem.

What Your Data and Infrastructure Must Support

AI forecasting models are only as strong as the data pipeline feeding them, and this foundation matters more than most vendor conversations suggest.

Internal data requirements typically include historical sales, current inventory levels, pricing history, promotional calendars, and consistent SKU master data across systems.

External data requirements extend to POS and retailer sell-through data, weather feeds, economic indicators, and competitor activity where it can be sourced reliably.

Integration requirements connect the forecasting layer to ERP, warehouse management systems, trade promotion management platforms, and retailer data-sharing portals, so forecast outputs translate into operational action.

One structural issue that can occur is the coordination latency problem, the gap between a demand signal appearing and the operational decision that responds to it. When that gap stretches to the next S&OP cycle rather than closing within hours, supply chains lose money regardless of forecast accuracy. A data readiness assessment, covering data quality, integration maturity, and governance, should precede platform selection.

How to Get Started: A Practical Roadmap

Moving from evaluation to a working AI forecasting capability tends to go more smoothly when organisations follow a sequence rather than a full-scale rollout from day one.

1. Define the decision. Identify the specific inventory, production, or allocation decision the forecast needs to improve, rather than starting with "better forecasting" as a vague goal.

2. Assess data readiness. Audit historical data quality, SKU hierarchy consistency, and integration maturity before committing to a platform.

3. Establish baseline metrics. Measure current forecast accuracy and bias by SKU segment, so improvement tracks against a genuine starting point.

4. Start with high-value, stable SKUs. Prove the approach where data quality is strongest, and the business case is clearest, then expand.

5. Connect outputs to operational systems. A forecast creates value once it reaches replenishment, allocation, or production planning systems.

6. Build planner adoption. Give planners visibility into why a forecast changed, not just the new number, to build lasting trust.

7. Scale and expand. Extend to harder SKU segments, categories, or markets once the core capability proves out.
   

UK and EU Considerations for Data Governance

Retail and CPG companies operating across the UK and EU face data governance requirements that vary by the type of data a forecasting model uses, so the right starting point is understanding which rules apply to which inputs.

Most demand forecasting relies on aggregated sales, inventory, and pricing data, which sits outside data protection law. UK GDPR and EU GDPR become directly relevant only where a forecasting model draws on personal data, such as purchase history from individual loyalty programmes used to inform demand signals; in those cases, a clear lawful basis for processing and appropriate safeguards are required. Where forecasting relies on data shared between a CPG manufacturer and a retail partner, the underlying data sharing agreement needs to set out ownership, permitted use, and retention terms clearly, since these arrangements vary by retailer and market.

Traceability and product-level regulation add a further layer for specific categories, and the relevant rules depend on the sector. Food and beverage businesses operating in the EU carry traceability obligations under EU food law (Regulation (EC) No 178/2002 and related legislation), shaping what supply chain and provenance data needs capturing. Across a wider set of product categories, the EU's Ecodesign for Sustainable Products Regulation is progressively introducing Digital Product Passport requirements, which will require certain product and supply chain data to be captured and shared in a structured, auditable way, an obligation worth factoring into data architecture ahead of time for businesses in scope.

Mapping which of these obligations genuinely apply, rather than assuming the broadest possible set applies to every model, keeps the compliance effort proportionate and the forecasting architecture easier to scale across markets and retail partners.

Why Tredence for AI Demand Forecasting in Retail and CPG

Demand sensing and supply chain intelligence sit at the core of Tredence's retail and CPG practice, built on the view that forecasting technology delivers value only when it connects cleanly to the operational decisions it supports.

Tredence works with retail and CPG organisations to assess data readiness, design SKU segmentation strategies suited to the actual planning problem, and build forecasting architectures that integrate with existing ERP, WMS, and trade promotion systems, rather than operating as standalone tools. Grounded in data engineering depth and genuine retail and CPG domain experience, this approach turns a forecasting model into a system planners actually use, including the governance foundation UK and European organisations need to scale forecasting across markets with confidence.

Explore how Tredence's supply chain analytics and demand forecasting services help retail and CPG organisations turn AI forecasting from a pilot into an operational advantage. Get in touch with us to assess your data readiness and map out a practical path forward.

FAQs

1. What is AI demand forecasting, and how does it differ from traditional statistical forecasting?

AI demand forecasting uses machine learning to process a wide range of variables simultaneously, including promotional data, weather, and pricing, updating continuously as new data arrives. Traditional statistical forecasting relies mainly on historical sales patterns on a fixed update cycle, limiting its ability to respond to sudden demand shifts.

2. What data does a retail or CPG business need for AI demand forecasting? 

At the basic level, you need historical sales, inventory levels, pricing history, promotional calendars, and clean SKU data. You can also bring in signals like POS sell-through, weather, and competitor activity when that data is available. The more reliable and consistent the underlying data is, the more useful the forecasts tend to be. 

3. What is demand sensing and how does it differ from demand forecasting? 

Demand sensing focuses on short-term shifts using current, frequently updated data, typically over a horizon of days to a few weeks. Predictive demand forecasting projects further ahead, from weeks to well over a year, supporting strategic planning decisions.

4. How should retail and CPG companies segment SKUs for AI forecasting? 

Group SKUs into stable core, promotional, seasonal, new product, and long-tail categories, since each behaves differently and benefits from a distinct modelling approach and accuracy expectation.

5. What are the most common reasons AI demand forecasting implementations fail? 

Poor data quality, inconsistent SKU hierarchies, missing promotional data, and weak planner adoption undermine forecasting programmes far more often than the choice of algorithm.

6. How should retail and CPG companies measure the ROI of AI demand forecasting? 

Track forecast accuracy and bias by SKU segment against a pre-implementation baseline, alongside operational metrics such as stockout rate, overstock levels, and working capital tied up in inventory.