Why Your Demand Forecast Is Wrong Before the Model Even Runs

Sales teams inflate demand forecasts to ensure they never face a stockout. Finance teams cut those forecasts to protect working capital targets. Supply chain teams add safety stock because they trust neither input. The AI demand planning model then trains on the output of that negotiation process - numbers that reflect organisational politics, not market demand.

This is the core problem with demand forecast accuracy in FMCG, and it is not solved by a better algorithm. It is a behavioural problem embedded in the incentive structure. When a sales manager is measured on service level - never running out of stock - they have every rational incentive to over-forecast. When a finance director is measured on inventory turns, they have every rational incentive to cut that forecast. The model does not fix this dynamic. It learns it.

The consequence is a forecast that is systematically biased in predictable ways. High-margin products are over-forecast because the sales team protects them. New products are over-forecast because launch optimism is never penalised. Promotional periods are under-forecast because the promotional uplift sits in a separate system that nobody connected to the demand plan.

Forecast bias compounds across planning cycles. An over-forecast in month one leads to overproduction in month two, which leads to excess inventory in month three, which leads to a finance-driven forecast cut in month four - which leads to underproduction in month five and a stockout in month six. The forecast error is not random. It is systematic and directional, and it creates a supply chain that oscillates between excess and shortage rather than tracking actual demand.

The financial cost is significant: excess inventory ties up working capital and creates write-off risk, while stockouts create lost sales and customer service failures. Both exist simultaneously in most FMCG supply chains - not because demand is volatile, but because the forecast is structurally biased.

Improving forecast accuracy by 10 percentage points in a £100M FMCG business typically reduces inventory by £3–5M and reduces lost sales by 1–2% of revenue. Those improvements do not require a better AI model. They require removing the systematic bias from the input data.

The practical approach is to create two parallel forecasts: a statistical baseline that models historical demand patterns without human intervention, and a consensus forecast that incorporates sales, marketing, and finance input. The gap between the two is the measure of forecast bias - and it should be tracked and reviewed in every S&OP meeting.

When the consensus forecast is consistently 15% above the statistical baseline and the actual demand follows the statistical baseline, the organisation has quantified its forecast inflation and can have an evidence-based conversation about changing the incentive structure. That conversation is uncomfortable, but it is the only one that actually improves forecast accuracy.

The AI demand planning tool adds value on top of a clean statistical baseline - incorporating external signals (weather, promotions, competitor activity) that a statistical model alone cannot capture. But the foundation must be a clean signal, not a politically adjusted number. The tool cannot fix the input. Only the organisation can.

The statistical baseline everyone agrees they need only exists if the data to build it is in one place. In most FMCG businesses it is not: shipment history sits in the ERP, distributor sell-out and modern-trade POS arrive as spreadsheets, promotions live in a trade-marketing system, and none of it is joined. The first build is unglamorous and decisive — land SAP S/4HANA or Microsoft Dynamics 365, distributor sell-out, POS, and promotional calendars into one governed store on Microsoft Fabric, with OneLake as the single copy and Azure Data Factory moving the feeds.

On that foundation a statistical baseline can run untouched by negotiation, and a Power BI Direct Lake semantic model holds one definition of demand, forecast accuracy, and bias that S&OP reads from. Sell-out — not shipments — becomes the demand signal, because shipments measure what the supply chain pushed, not what the market pulled. That single connection, distributor sell-out into the planning layer, is the most common gap we close, and it usually moves forecast accuracy more than any change of algorithm.

Only then does AI earn its place. With a clean baseline and connected external signals, predictive analytics can sense demand shifts and lift the SKUs and regions where the statistical model is weakest. The order matters — unify the data, then predict — and it is the order most failed demand-planning tools skipped.

No data foundation fixes an incentive that rewards cover. If the sales team is measured purely on service level and never on bias, the consensus forecast will keep drifting above the baseline regardless of how clean the signal is. The platform can quantify the inflation — the persistent gap between consensus and statistical baseline — but leadership has to act on it in S&OP. The number makes the conversation possible; it does not have the conversation for you.

New products are the second hard case: there is no history for a statistical model to learn from, so launch forecasts stay judgement-led and over-optimistic. The honest approach is analogue modelling against similar past launches plus tight early-sell-out tracking, not a confident AI number that implies a certainty the data cannot support.

And promotions remain the perennial blind spot. If promotional uplift lives in a system nobody connected to the demand plan, the forecast will under-call every promoted period no matter how good the baseline is. Connecting that data is a data integration job, not a modelling one — which is exactly why it gets skipped.

The mechanism that exposes bias is simple to describe and uncomfortable to run: maintain two forecasts side by side. The statistical baseline models historical sell-out with no human adjustment. The consensus forecast layers in the judgement of sales, marketing, and finance. The gap between them, tracked by SKU and region in every S&OP cycle, is your bias — quantified, not debated. When consensus runs persistently 15% above baseline and actuals follow the baseline, the inflation is no longer an opinion.

Surfacing that gap is a reporting job the governed foundation makes trivial. A Power BI semantic model holds both forecasts and the variance, so the S&OP meeting opens with the bias on screen rather than with each function defending its own number. The conversation shifts from "whose forecast is right" to "why is this category structurally over-called, and what changes." That is the meeting most supply chain analytics programmes are trying to reach and rarely do.

Over a few cycles the discipline changes behaviour, not just measurement. When over-forecasting a high-margin line is visible and attributed, the incentive to pad it weakens. Forecast accuracy improves because the input improves — and only then does predictive analytics, sensing demand shifts from weather, promotions, and competitor activity, add lift on top of a baseline it can trust rather than a negotiated figure it would simply learn to repeat.

None of this is a tooling upgrade. It is an operating rhythm — two forecasts, one tracked gap, reviewed every cycle — sitting on a foundation where both numbers are governed and reconcilable. The businesses that hold that rhythm are the ones whose fill rate and inventory both improve; the ones that buy another forecasting engine and skip the discipline see neither move.

The return on this is not a cleverer forecast — it is a supply chain that stops oscillating between excess and shortage. When the baseline is clean and the bias is visible, inventory and fill rate move in the right direction at the same time: a 10-point improvement in forecast accuracy in a £100M business typically frees £3–5M of working capital and recovers 1–2% of revenue lost to stockouts.

And it starts in weeks, not quarters. A six-week Discover and Foundation build can connect the priority systems, stand up the statistical baseline, and put the consensus-versus-baseline gap in front of S&OP — first value in 6 weeks, not a 50-slide roadmap. The AI layer is added once the foundation is trusted, not before.

The uncomfortable diagnosis most supply chain heads already half-know: the forecast was wrong before the model ran, because the input was a negotiated number. Fix the input and the accuracy the tool was bought to deliver finally shows up.