Predict turnover, margin, and the impact of every assortment / promo / pricing change
ML-driven sales forecasting and what-if scenarios for FMCG operations. Plan with predicted impact in place of historical averages.
Discuss a pilot →Quick facts
- Business size
- FMCG brands, retail chains, CPG distributors with > 100 SKUs
- Timeline
- 6-10 weeks test, 3-5 months pilot, 6-9 months production
- Budget range
- Pilot from €45k.
- Hardware
- Cloud-based data warehouse and forecasting service.
- Data needed
- 12+ months sales history, product catalog, promo calendar, optional: competitor / weather / external signals.
- Evolution
-
- Genesis
- Custom-built
- Product
- Commodity
A vendor sells this result ready-made. We set it up and tune it to you.
What this scale means
Further right means more proven and cheaper. Further left means newer and riskier. Here is the test for each step.
- Commodity
- You could get the result yourself from a ready service, with almost no work. We rarely take these on.
- Product
- A vendor already sells this result turnkey, like shelf recognition from Trax or document reading from ABBYY. If one of them fits you, use it. You come to us when it does not: when it has to run on your own servers, cost less, or fit systems the product cannot reach.
- Custom-built
- Vendors sell only the parts. A tool like Tableau hands you charts, but the dashboards and metrics for your business still have to be built. That build is the work, and that is us.
- Genesis
- The approach exists but does not work reliably yet. You are betting on it maturing, so it costs more and carries more risk.
Expected outcomes
The Problem
FMCG operations live and die on demand forecasting. Inventory management, production planning and promo scheduling all depend on predicting turnover. Historically this is done with spreadsheet heuristics: take last year and adjust for trends. The result is chronic overstock on slow movers and stockouts on fast movers. Promo plans ignore cannibalization: one product’s promo eating sales from its shelf neighbors.
The data exists. Modern FMCG operations have rich sales history, promo calendars, weather and competitor signals, and consumer-trend data. Stitching it into operational forecasts takes ML infrastructure most FMCG operations haven’t built.
What the Solution Does
A forecasting and what-if scenario engine for FMCG operations.
- Forecast: turnover at the SKU, region and week level.
- Decompose: split the forecast into baseline trend, seasonality, promo impact and external signals.
- What-if scenarios: see how forecasts move if you change assortment, promo schedule or pricing.
- Inventory and production planning: feed forecasts into upstream planning.
- Continuous improvement: the model retrains as new data arrives, so accuracy improves over time.
Where It Fits
This makes sense if you…
- Operate FMCG / CPG / retail with > 100 SKUs
- Have 12+ months sales history with structured promo data
- See real cost from forecast misses (waste, stockouts, mis-allocated production)
- Want scenario-based planning that goes past plain trend extrapolation
This is probably not the right time if you…
- Operate at small scale where spreadsheets work
- Lack historical data (12+ months minimum)
- Don’t have organizational capacity to act on forecasts (the forecast is useless if planning ignores it)
Business Value
Forecast accuracy. Typical MAPE is 10-20% at the SKU-region-week level once the model matures. It is better at higher aggregation (category-region-month) and worse on highly seasonal or low-volume SKUs.
Waste reduction. FMCG perishables see around 15-30% less inventory waste with accurate forecasting. That is a major economic driver for fresh, dairy and bakery lines.
Stockout reduction. Stockouts drop by around 20-40% when forecasts inform inventory and production planning.
Scenario-driven planning. Promo planning, assortment changes and pricing experiments all become hypothesis-then-measure cycles.
How It Works
The architecture has six parts.
1. Data foundation
Sales history (12+ months ideal), product catalog, promo calendar, optional external signals (weather, competitor pricing, consumer trends). Loaded into a warehouse with Datapipe.
2. Feature engineering
Per SKU-region-week features: baseline trend, seasonality, recent velocity, promo flag, neighboring-SKU activity (cannibalization signals), external context.
3. Forecasting models
Hierarchical ML: gradient-boosted trees with seasonality decomposition. We’ve shipped XGBoost, LightGBM and Prophet variants depending on data structure.
4. What-if engine
Given a proposed change (different assortment, new promo, price shift), the engine re-forecasts under the modified conditions. Planners compare scenarios.
5. Inventory / production integration
Forecasts feed downstream planning systems (SAP, custom, etc.).
6. Continuous improvement
Forecasts are tracked against actuals. Models retrain on new data, so accuracy improves. Structural changes (new SKUs, new regions) are handled gracefully.
Stack
Datapipe, ClickHouse / BigQuery / Snowflake, XGBoost / LightGBM / Prophet for forecasting, Python for feature engineering, Metabase / PowerBI for planning dashboards.
What You Need to Make This Work
Data. Sales history, product catalog, promo calendar. Optional external signals.
Integrations. Source data systems (ERP, sales-data feeds). Output to planning systems.
Hardware. Cloud-based.
Team. Demand planning lead. Sales-ops lead. Finance contact. Data engineer.
Implementation Roadmap
1. Test (6-10 weeks)
Train a baseline model on a SKU subset. Validate against historical actuals. Output: a written report on accuracy and recommendations.
2. Pilot (3-5 months)
Production deployment for one category or region. Wire up planning-system integration. Build dashboards. Output: working forecasting with measured business impact.
3. Production (6-9 months)
Full SKU / region coverage. Continuous retraining. Quarterly accuracy review.
Keep in Mind
- Low-volume and highly-seasonal SKUs are hard. Forecast accuracy drops on long-tail and seasonal items. We report per-SKU-category accuracy so you see where it weakens.
- Promo cannibalization matters. A promo on Product A boosts A and depresses adjacent products. Forecasts need to model this.
- New-SKU cold-start. The first 6-12 weeks of a new SKU have noisy forecasts. We use similar-SKU patterns to bootstrap.
- External signals add complexity. Weather, competitor, consumer-trend data improve accuracy in some categories and add noise in others. Test per category.
- Planning culture matters. Forecasts are useless if the planning team doesn’t trust or use them. Change-management is real work.
FAQ
Can this work for retail, beyond FMCG manufacturers?
Yes. The architecture carries over, with different signals. Retail demand forecasting is a major adjacent use case.
Promo planning specifically, can the system recommend promo schedules?
Yes. What-if scenarios let planners compare alternative promo schedules. Full optimization (auto-recommend the best promo) is possible. We usually deploy it as a recommendation that a planner approves, so the system advises and a human decides.
What’s the smallest useful scope?
One category and one region with 12+ months of data. Roughly 1000+ SKU-weeks of training data minimum.
Multi-country support?
Yes. Per-country models account for local seasonality, promo patterns, demand drivers.
Ready to Discuss?
If you operate FMCG / retail / CPG at meaningful scale and forecasting is currently a spreadsheet-driven exercise, this is a worthwhile pilot. We’ll walk through your sales data and your planning process, and tell you what to expect.