Forecasting demand for fast-moving SKUs across channels

External factors matter. Heat waves lift demand for deodorant and cold drinks. Rain shifts shopping online. A traditional model that ignores weather is blind to these swings.

The result: a naive model that averages past sales will always be surprised by exactly the events that matter most.

What signals AI blends together

The core idea is simple. Instead of using only past sales, an AI model learns the relationship between demand and many signals at once. Here are the main ones for FMCG.

POS data (Point of Sale). This is the scanner data from checkouts, showing what actually sold, where, and when. Retailers often share it through platforms or data pools. POS is more accurate than shipment data because shipments tell you what left the warehouse, not what the shopper bought.

Promotion calendar. A structured record of every planned price cut, display, coupon, and feature, by SKU, store group, and week. This is often the single most predictive input for fast movers.

Weather. Temperature, rainfall, and forecasts. Especially powerful for seasonal or weather-sensitive categories (ice cream, sunscreen, soup).

Pricing. Both your price and competitor prices. Demand is sensitive to relative price, not just absolute price.

Search and web signals. For e-commerce, search volume and site traffic lead sales by days or weeks.

Calendar events. Holidays, paydays, school terms, major sporting events.

The AI model finds patterns like "when we run a BOGO promo AND the temperature is above 30 degrees AND it is the week before payday, deodorant sales jump 55%." No human can hold all these interactions in their head. A model can.

How the model actually works

Most modern FMCG forecasting uses gradient boosted trees (a machine learning method that builds many small decision rules and combines them) or deep learning models designed for time series. You do not need to know the math. You need to know the shape of the workflow.

Here is a simplified view of the feature table that feeds the model. Each row is one SKU in one store (or channel) for one week.

The model trains on years of these rows, learning how the target responds to each feature. To forecast next week, you fill in the known future values (the promo is planned, the weather is forecast, the calendar is fixed) and the model predicts units sold per SKU per channel.

For a deeper but still accessible primer on the underlying techniques, Google's Machine Learning Crash Course is free and well structured.

Forecasting across channels

The shampoo brand does not need one forecast. It needs a forecast for each channel because each behaves differently and each has different lead times.

Grocery and drugstore (retail). Demand driven by promotions and shelf presence. Lead times are longer, so forecasts feed into replenishment orders weeks ahead. Errors here cause empty shelves or overstock that ties up cash.

E-commerce (own site and marketplaces). Demand driven by search, ads, and delivery promises. Faster reaction, shorter lead times, but also more volatile. A single influencer post can spike demand overnight.

Quick commerce (rapid delivery apps). Very short shelf life on decisions. Demand is hyper-local and time-of-day sensitive.

A good practice is hierarchical forecasting: predict at the fine level (SKU by store by day) and at aggregate levels (SKU by region by week), then reconcile them so the numbers add up. This prevents the common failure where store-level forecasts sum to something wildly different from the national plan.

The promotion spike problem, solved

Back to our shampoo brand. The 40% miss happened because the old model treated a BOGO week like a normal week.

With an AI model that includes the promotion calendar and promo type as features, the model has learned from dozens of past BOGO events that this specific promo, on this specific SKU, in these stores, lifts demand by roughly 45%. It also learned the pantry loading effect: the two weeks after a promo dip below baseline because shoppers already stocked up.

The practical payoff is not perfect accuracy. It is narrowing the error band and, crucially, quantifying uncertainty. Good models output not just "140,000 units" but a range like "120,000 to 160,000 with 80% confidence." Supply chain teams use that range to set safety stock intelligently instead of guessing.

Common pitfalls when deploying

Garbage promo data. The promotion calendar is often maintained in spreadsheets by sales teams, riddled with errors and last-minute changes. If the calendar says no promo but a store ran one anyway, the model learns the wrong lesson. Clean promo data is the highest-leverage fix in most FMCG forecasting projects.

Ignoring the cold-start problem. New product launches have no history. AI cannot learn from data that does not exist. For new SKUs, use analogs (forecast based on similar past launches) and blend in human judgment.

Over-trusting the model during rare events. A pandemic, a supply shock, or a viral moment falls outside the training data. Keep humans in the loop for anomalies. The model is a co-pilot, not an autopilot.

Forecasting the wrong metric. Optimizing for lowest average error can hide big misses on your most important, highest-margin SKUs. Weight your accuracy measures by value, not just volume.

Measuring whether it worked

Two metrics dominate FMCG forecasting.

MAPE (Mean Absolute Percentage Error): the average size of your errors as a percentage. Lower is better. Fast movers with heavy promotions often sit in the 20% to 40% range for weekly SKU-store forecasts even with good models, so context matters.

Bias: are you consistently over or under forecasting? Bias is often more damaging than random error because it systematically drains cash into excess inventory or starves shelves.

Track these by channel and by promo versus non-promo weeks separately. An overall MAPE that looks fine can hide a disastrous promo-week performance, which is exactly where the money is.

Key Takeaways

• Fast-moving SKUs break simple history-based forecasting because promotions, channels, and weather cause the swings that matter most.
• AI models earn their value by blending POS data, the promotion calendar, weather, pricing, and search signals to learn interactions no human can track manually.
• Forecast each channel separately (retail, e-commerce, quick commerce) and reconcile across a hierarchy so store-level numbers add up to the national plan.
• Clean promotion data is usually the single biggest driver of accuracy for FMCG; fix the spreadsheet before blaming the model.
• Judge accuracy with MAPE and bias, split by promo versus non-promo weeks, and weight by product value so your key SKUs get the attention they deserve.