AI demand forecasting for consumer goods and electricity grids
A global FMCG group's spreadsheet forecasts left out weather, retailer scan data, social trends, inflation and FX. Forecasting every SKU in every market with signals like these, and rebalancing stock to match, cut inventory 18% at that group and stockouts 22% at a pan-African one, with shelf availability up at both.
How does AI demand forecasting work?
AI demand forecasting runs several models for every SKU and market, from statistical methods for stable lines to machine-learning and deep-learning models for promotions and volatile lines, and reweights them weekly by how each has performed against actuals. It adds outside signals such as weather, FX and retailer scan data, and feeds forecasts with confidence intervals into the planning system teams already use. The same approach forecasts national electricity demand, down to a 5-minute nowcast.
- A global FMCG group cut inventory 18% across 8,000 SKUs in 31 markets, as forecast error (MAPE) fell from 18% to 12%.
- A pan-African FMCG group cut stockouts 22% and overall inventory 14% across 18 countries, as MAPE fell from 16% to 8%.
- At the global FMCG group, ice-cream and beverages, where weather data mattered most, saw MAPE improve 31% and 22% and inventory fall 24% and 19%.
- A European national grid operator's day-ahead demand forecasts reached about 4% MAPE, from a 7-9% baseline, and system-balancing costs fell about 18%.
Demand forecasting, step by step
Bring the demand data together
Sales history, inventory positions, in-transit shipments and promotion calendars come from the ERP and planning systems, alongside outside data: weather, FX, retailer scan data, port and customs status. The pan-African FMCG group built a forecasting-ready dataset per country, with a data-quality layer handling each country's specific problems.
Forecast with several models at once
Each SKU and market gets an ensemble: statistical models for stable lines, gradient-boosted trees for promotional uplift and a deep-learning model for volatile lines. Weights are re-estimated weekly from forecast-versus-actual results. The output is a 26-week forecast with confidence intervals. The grid operator's ensemble models heat-pump, EV-charging and industrial demand separately.
Rebalance stock to the forecast
An optimiser weighs the forecasts against current stock, in-transit shipments, transfer costs, lead times and safety-stock policies, and recommends where to move stock. At the pan-African FMCG group, cross-country rebalancing was the largest single contributor to its 14% inventory reduction. Recommendations come out weekly and are reviewed with the regional planning teams.
Feed the plans teams already run
At the pan-African FMCG group, forecasts and recommendations go into the existing planning system, which planners still use as the system of record. The global FMCG group's forecasts are written back into its ERP twice a day. At the grid operator, forecasts feed the day-ahead cycle, the intra-day refresh and a 5-minute nowcast.
Track accuracy and catch shifts early
Forecast error is tracked continuously by SKU and country, and planners see the causes of the largest misses with suggested model adjustments. At the grid operator, anomaly detection on the live demand stream flags emerging patterns, such as heatwave-driven shifts in their early stages, before the model has learned them, and feeds its recalibration.
Planners keep the decisions. At the pan-African FMCG group, forecasts go into the planning system its teams already use, weekly rebalancing recommendations are reviewed with regional planners, and category managers use the new patterns to adjust each country's range and promotions. At the grid operator, the operations team sees the forecast, its confidence interval and recommended balancing actions on dashboards built around each role.
Results from 3 deployments
Every figure below comes from the case study it links to.
Global FMCG
Demand Forecasting + AP Automation
- 8,000 SKUs forecasted
- 12% forecast MAPE (was 18%)
- <48h invoice TAT (was 14 days)
Pan-African FMCG
- 14% overall inventory reduction
- 18 african countries on-platform
- 16 → 8% forecast MAPE improvement
National Grid Operator
Grid Demand Forecasting
- 18% system-balancing cost reduction
- 5-min real-time refresh cadence
The accelerators behind it
Pre-built accelerators do the work, configured to your documents, rules and systems. Delivery took 28 to 36 weeks in the case studies above.
Demand forecasting: the questions buyers ask
How accurate is AI demand forecasting?
At the global FMCG group, group-level forecast error (MAPE) fell from 18% to 12% on a rolling 13-week measure; at the pan-African FMCG group it fell from 16% to 8%. The national grid operator's day-ahead demand forecasts reached about 4% MAPE, from a 7-9% baseline. The biggest gains came on volatile and weather-sensitive lines.
Doesn't cutting inventory lead to more stockouts?
Not in these deployments. The pan-African FMCG group cut stockouts 22% while overall inventory fell 14%, and on-shelf availability improved. The global FMCG group held 18% less inventory and on-shelf availability still rose 1.4 percentage points. Both moved from buffer-driven to forecast-driven stocking.
What data does it use besides sales history?
At the global FMCG group: weather for ice cream and beverages, retailer scan data for fast-moving lines, social trends for personal care, and inflation and FX in volatile markets. The pan-African FMCG group added FX, weather, and port and customs status. The grid operator uses substation metering, weather forecasts, distribution-network and EV-charging data.
Do we have to replace our planning system or ERP?
No. At the pan-African FMCG group, forecasts and recommendations feed the existing planning system, which regional teams still use as the system of record; the platform gives it better inputs. The global FMCG group's forecasts are written back into its ERP twice a day. The grid operator's forecasts flow into its existing grid-operations framework through standard APIs.
Can the same approach forecast electricity demand?
Yes. A European national grid operator uses an ensemble with separate models for heat-pump, EV-charging and industrial demand. Day-ahead error fell to about 4% MAPE from 7-9%, system-balancing costs fell about 18%, and a 5-minute nowcast supports real-time decisions. Production forecasting is served on the operator's CPU infrastructure.
Where does the data sit?
Inside the company's own environment. The grid operator runs the platform on its private cloud, and its operationally sensitive demand data stays inside its perimeter. The pan-African FMCG group keeps data in-country in markets with data-residency rules and consolidates processing elsewhere. Both FMCG platforms run in the groups' own cloud tenants.
How long does a deployment take?
Delivery took 28 weeks at the national grid operator and 32 weeks at the global FMCG group, across 8,000 SKUs in 31 markets and alongside an accounts payable project. The pan-African FMCG group took 36 weeks across 18 countries, starting with consolidating data from each country's systems.
Get the MindMap capability deck
The platform, the accelerator library and how an engagement runs, in one deck. Our team emails it within one business day.
More use cases
KYC and account opening →
4 hrs onboarding at a UK challenger bank, down from 5 days
AML and fraud detection →
71% fewer AML alerts to analysts at an Indian public-sector bank
Regulatory reporting and compliance →
7 days regulatory close at an ASEAN bank, down from 28 days
See it running on your own process
A 20-minute walkthrough with the engineers who build these deployments: your documents, your systems, the accelerators running.