Demand Forecasting for FMCG That Beats the 60% Wall

Demand Forecasting for FMCG That Beats the 60% Wall

The forecast accuracy ceiling for fast-moving consumer goods has not moved in twenty years. Most $1M-$10M FMCG operators run their statistical forecast off retailer shipment history, then quietly accept a MAPE in the 50s and 60s as the cost of doing business. That accuracy floor is not a law of physics. It is the predictable output of feeding a model the worst possible signal: what retailers ordered, two weeks after they decided what to order, four weeks after consumers actually walked the aisle.

Your Shipment Forecast Is the Bullwhip Talking to Itself

Pull up your demand forecast right now. If the input data set starts with retailer purchase orders, shipments, or warehouse withdrawals, you are forecasting the bullwhip, not consumer demand.

Industry analysis surfaced by ToolsGroup shows CPG forecast accuracy at the unit and location level commonly sits in the 50s and 60s on MAPE, with very few brands clearing the 70s (CPG forecasting ceiling). Aggregate-level statistical planning rarely lifts forecast accuracy above 60% even for fast-moving items. That is not a small problem. It is the difference between a Coles or Woolworths buyer trusting your replenishment schedule and the buyer placing safety-stock orders that lock up working capital you cannot get back.

The reason for the ceiling is structural. Statistical forecasts in ERP and APO systems take retailer shipment history as the demand truth. Shipments are not demand. Shipments are the distorted echo of demand after retailers, distributors, and your own DC have all applied their own buffer logic. By the time a Coles DC receipt shows up in your forecasting model, the consumer who triggered that purchase did so somewhere between fourteen and forty-five days earlier. The signal you are forecasting against is the bullwhip at its most amplified point.

Practitioners who measure MAPE rigorously already know this (Imperia MAPE guide). They see the variance between aggregate-level forecasts and item-location forecasts. They watch promotional weeks blow out the model. They sit through monthly S&OP meetings where every department blames a different number. The instinct, encouraged by every supply chain software vendor, is to buy a smarter model. ML, AI, neural nets, demand sensing add-ons. The pitch is that better algorithms will rescue you from a bad signal.

That pitch sells. It also misses the point. A more sophisticated algorithm trained on the same shipment history produces a more confident wrong answer. The bullwhip distortion is in the input layer, not the modelling layer. You do not fix it by spending six figures on a forecasting platform that trains on the same retailer purchase order data your spreadsheet has been chewing on for three years.

I have seen this pattern across multiple FMCG brands in the $1M-$10M band. Founders sign a contract with a forecasting vendor, pay setup fees that run into five figures, and watch their MAPE drop by three or four points before plateauing. Three points of MAPE improvement on a base of 58% is rounding error. It is not the number that earns better service levels with retailers or cuts the safety-stock burden on your warehouse. It is the comforting illusion that you have done something while the underlying input set stays the same.

The contrarian read is the opposite. The fastest way to lift forecast accuracy is to throw out the shipment-history default and replace it with consumption signals. That means the data telling you what consumers are actually buying off the shelf, in real time, before the retailer purchase order even gets cut.

The Consumption Signal Engine: A Different Input Stack

There is a name for the alternative input stack. I call it The Consumption Signal Engine. It is a reweighting of the forecast inputs around what consumers are doing, not what retailers ordered. The mechanics are blunt, the wins are real, and a $1M-$10M brand can stand the first version up in a spreadsheet inside thirty days.

The Engine has four input streams, ranked by recency and proximity to the consumer.

Stream 1 is POS sell-through. Daily or weekly point-of-sale data pulled either through a syndicated data partner like Circana or NielsenIQ, or directly through retailer vendor portals such as Coles Connect, Woolworths Quantium, or Amazon Vendor Central. This is the closest you get to actual purchase events without leaving the retail channel.

Stream 2 is DTC and direct-channel orders. Your own Shopify, Amazon Seller Central, or marketplace order stream. Lower volume than retail for most FMCG brands, but the temporal lag is zero. A unit sold on your DTC site at 8pm Tuesday shows up in the order table at 8.01pm Tuesday.

Stream 3 is household panel data. Syndicated panel coverage from NielsenIQ Homescan, Kantar, or Circana that reports household-level purchase frequency and basket composition. Slower to arrive (typically four-week lag) but corrects retailer-channel bias and reveals demand you do not see in your own data.

Stream 4 is external demand signals. Search trends, social conversation volume, promo calendars, weather, and macro consumption indicators. Leading rather than lagging. A spike in Google searches for sunscreen the week before a heatwave is a forecast input. A retailer purchase order placed three weeks later is not.

The architecture is the spine. Demand sensing as a category has matured around exactly this principle of pulling consumption signals forward in the input stack (RELEX demand sensing, Circana demand sensing). The Consumption Signal Engine is the operator-grade version of that idea. You do not need a vendor licence to begin. You need a spreadsheet, a Circana feed or a retailer portal export, and the discipline to keep the shipment-history forecast running in parallel for the first quarter. That parallel run is what proves the variance on your own data.

Two operating principles run through the Engine.

First, weight signals by recency and consumer proximity. POS beats shipment, DTC beats POS for the SKUs you sell direct, panel beats nothing if you can afford it, external beats all four for new product launches with no sales history.

Second, do not chase a single calibrated forecast on day one. Run the Consumption Signal Engine forecast as a shadow. Let the planners see both numbers in their weekly review. Make them defend the variance. The forecast that wins is the one that was closer to retailer scan data the previous month, measured at item-location.

Phase 1: The POS Overlay (Days 0-30)

The Phase 1 build is a POS overlay. The point is to put a consumption signal alongside your existing shipment-history forecast, run them in parallel for thirty days, and quantify the variance per SKU. You are not replacing the forecast. You are exposing how wrong it has been.

Week 1 is data plumbing. Pull a daily or weekly POS extract from your two largest retail customers. In Australia that means Coles Connect, Woolworths Quantium, or Metcash IRI. In the US that is Amazon Vendor Central, Walmart Retail Link, or a syndicated Circana feed (Circana forecasting solutions). For most $1M-$10M FMCG brands, two retailers cover 60% to 80% of revenue. That is enough coverage to start. Do not wait for full-coverage syndication.

Week 2 is the alignment build. Map your internal SKUs to retailer item codes. The mapping table is small, finite, and the single most common reason POS overlay projects stall. Build it once in a spreadsheet, version-control it in a shared drive, and assign one supply planner as the owner. Without this owner the table goes stale inside two weeks.

Week 3 is the variance report. For each SKU at each retailer, compute the four-week rolling variance between forecasted demand (your existing shipment-history number) and observed POS sell-through. Flag any SKU where the variance exceeds 25% in either direction. Do not try to fix the forecast yet. Just measure the gap.

Week 4 is the first weighted blend. For your top twenty SKUs by revenue, run a blended forecast: 60% POS sell-through, 40% existing shipment forecast, four-week trailing average. Compare both to the next four weeks of actual scan data. The blended number will be closer for roughly three out of four SKUs. The exceptions tell you something useful about promotional cadence or distribution gain that needs a second adjustment.

Tooling for Phase 1 is deliberately boring. A Google Sheet or Excel workbook is enough. The data volumes for two retailers and 100-300 active SKUs sit comfortably in a spreadsheet. The temptation to go straight to a Looker dashboard or a forecasting SaaS platform should be resisted. The Phase 1 win is not a polished tool. It is the visible variance number on the weekly supply review agenda.

KPI for Phase 1: weighted MAPE at item-location level for your top twenty SKUs at your top two retailers. Set the baseline in Week 4. Track monthly. The Phase 1 lift is usually 8 to 12 MAPE points across the top SKUs, and it is the cheapest lift you will ever buy.

Phase 2: DTC, Panel, and Promo Layering (Days 31-90)

Phase 2 widens the input set. The POS overlay running in Phase 1 covers what your two largest retailers sold last week. Phase 2 adds the data that fills the gaps: your DTC orders, household panel coverage, and the promotional calendar that was distorting the forecast all along.

Start with DTC because it is free and the latency is zero. Pipe your Shopify, Amazon Seller Central, BigCommerce, or marketplace orders into the same forecasting workbook on a daily refresh. For the SKUs that move on both retail and DTC, you now have two consumption signals to weight against shipment history. The blending rule that works assigns DTC a weight proportional to its share of total category demand for the SKU. Cap the weight at 30% to avoid distorting the forecast for retail-heavy items.

Panel data is the second layer, and the one most $1M-$10M FMCG operators write off as too expensive. A scaled-down NielsenIQ or Circana panel subscription, focused on a single category, runs $30,000-$60,000 AUD a year. For a brand with $5M in revenue, that is a 0.6% to 1.2% expense line. The return shows up in two places. The first is faster identification of demand softening in households you do not directly serve. The second is better calibration of new-product launches where you have no shipment history to feed the model. Build a quarterly review of panel-versus-shipment variance into the S&OP cadence.

The promotional calendar is the third Phase 2 input. Most ERP forecasts apply a flat percentage uplift to a base forecast for any promoted week. That works on stable mature SKUs and breaks on everything else. The Phase 2 fix is a per-SKU promo elasticity table built from the previous twelve months of POS scan data. For each SKU, capture the average lift on weeks with a 20% off promotion, on weeks with feature-and-display, and on weeks with co-promotion against a competing brand. Plug the elasticity values directly into the weighted blend. This single addition typically cuts promo-week forecast error from 35-40% MAPE down to 15-20% (Drivepoint CPG forecasting).

Phase 2 also forces a discipline conversation. Demand planners and brand marketers usually live in different systems. The promo calendar lives in a marketing folder. The forecast lives in a supply planning tool. Closing that gap requires a single source of truth for the promo calendar, a shared spreadsheet or a planning tool like Anaplan or Pigment, that both teams update before any forecast cycle runs.

KPIs for Phase 2: a weighted MAPE drop of an additional 5 to 10 points for the SKUs receiving DTC and panel signal, and a halving of promo-week forecast error on the elasticity-adjusted SKUs. Both metrics are tracked monthly in the same workbook that produced the Phase 1 baseline.

Phase 3: External Signals and the S&OP Cadence (Quarter 2 Onward)

Phase 3 is where the Engine becomes a forecasting practice rather than a project. The first two phases proved the variance and built the input set. Phase 3 brings in the leading signals that catch demand turning before it hits retailer scan data, and ties the whole stack into a monthly S&OP rhythm that the entire business runs on.

External demand signals do not need a data science team to add value. Four pay back fastest for a $1M-$10M FMCG brand. Google Trends and social conversation volume on TikTok and Instagram cover near-term shifts. Weather data matters for category-sensitive SKUs like sunscreen, soup, ice cream, and allergy medication. Macro indicators round out the set: consumer confidence, fuel prices for impulse purchase categories, and ATO BAS data for cash-flow-sensitive segments. The connection point is a weekly indicator dashboard that the demand planner reviews before publishing the next forecast cycle. Anomalies in the indicator dashboard trigger a manual override on the affected SKUs, with a written rationale logged in the workbook.

Demand sensing as a discipline has codified the seven capabilities that move a brand from shipment-based to consumption-based forecasting (ToolsGroup demand sensing). The capabilities are not technical features. They are organisational ones: data freshness, signal weighting, exception management, scenario planning, S&OP connectivity, learning loops, and governance. Phase 3 is where you start building those capabilities deliberately rather than backing into them.

The S&OP cadence is the binding ring. Run a monthly Sales and Operations Planning meeting with three fixed agenda items. First, review last month's weighted MAPE at item-location. Second, review the variance between consumption-based and shipment-based forecasts for the next ninety days. Third, agree on the demand plan that goes into the supply plan. Sales presents demand. Supply confirms producibility. Finance signs off on the cash and inventory implications. The meeting takes ninety minutes and the discipline is non-negotiable. The output is a single demand plan that sales, supply, and finance all signed.

The macro trajectory is set. Gartner predicts that 70% of large organisations will adopt AI-based supply chain forecasting to predict future demand by 2030 (Gartner forecasting research). The brands that get there first are not the ones that bought the most expensive platform. They are the ones that built consumption-signal discipline into the input layer first, then layered AI on top of a clean signal stack. AI on top of shipment history is still AI on top of the bullwhip.

Phase 3 is also when you stop running the shipment-history forecast in parallel. Once the Consumption Signal Engine has produced a lower weighted MAPE for three consecutive months across more than 70% of your active SKU base, retire the shadow shipment forecast. Keep the historical numbers archived for variance analysis, not for decision-making.

The New North Star: Weighted MAPE at Item-Location

The metric you measure determines the discipline you get. The single most common forecasting failure mode in FMCG is reporting aggregate-level MAPE, the company-wide forecast accuracy number that gets quoted in the board pack, while running operations off item-location-level demand. The two numbers are not comparable. Aggregate MAPE flatters because high-volume SKUs absorb low-volume errors. A brand with 60% aggregate MAPE often has 35-40% MAPE on the long tail, and the long tail is where stock-outs and write-offs actually happen.

Weighted MAPE at item-location is the new north star. It is a SKU-store (or SKU-DC) level error metric, weighted by revenue or volume contribution, calculated monthly, and reviewed at S&OP. The exact formula is well documented across forecasting practitioners (Forecast accuracy benchmarks). The discipline is the harder part. It requires the demand planning team to defend a number that gets worse before it gets better, because you are now measuring at the level where the truth lives.

The before-and-after picture for brands that switch is consistent. Aggregate MAPE moves modestly, often from the high 50s to the high 60s. Weighted MAPE at item-location moves more sharply, frequently from the low 50s to the high 60s or low 70s. Stock-outs at retailer DCs drop by 15-25% over six months. Returns to the warehouse from retailer over-orders fall by a similar amount. The combined working capital effect for a $5M FMCG brand sits in the $150,000 to $400,000 range, recurring annually, with no software spend and no headcount addition.

That is the prize. Not a fancier algorithm, not a more polished dashboard, not a vendor logo on the supply chain deck. The prize is forecasts that match what consumers are actually doing, measured at the level where the operations team can act, reviewed in a cadence the whole business is signed up to. Build that, and the rest of the supply chain stops fighting the forecast and starts trusting it.