FMCG & Retail | Supply Chain & Demand Forecasting
fewer missed calls LLM + Twilio telephony handling 24/7 emergency calls
Reduction in stockouts within the first quarter of deployment
About
Crestfield Foods is a mid-sized FMCG manufacturer operating across 12 regional markets, with a portfolio spanning over 200 SKUs across snack foods, beverages, and packaged goods. Selling through modern trade, general trade, and direct-to-consumer channels, the company had built consistent revenue growth — but its supply chain continued to operate reactively, with forecast errors quietly eroding margin across the business.
Industry
FMCG & Retail | Supply Chain & Demand Forecasting
Company size
1,000 – 5,000 employees
Founded
2008
The Company
A growing FMCG business held back by a reactive supply chain
Crestfield Foods is a mid-sized FMCG manufacturer with a portfolio spanning over 200 SKUs across snack foods, beverages, and packaged goods. The company operates across 12 regional markets, selling through modern trade, general trade, and direct-to-consumer channels — each with distinct demand profiles, seasonal patterns, and distributor dynamics.
Despite consistent revenue growth, the supply chain team had been struggling for years with a demand planning process that was fundamentally backward-looking. Forecast accuracy hovered below 40% for non-core SKUs, and the business was simultaneously experiencing stockouts in high-velocity markets and excess inventory in slower-moving regions. The cost of this imbalance — in lost sales, expired stock write-offs, and tied-up working capital — had reached a point where it was visibly affecting margin.
The challenge
Good data, but the wrong model
The core problem was not a lack of data. Crestfield Foods had years of historical sales data, distributor sell-out records, and regional performance reports. The problem was that their forecasting model — built in Excel and managed by a small planning team — was almost entirely reactive. It used historical sales averages with simple seasonal adjustments, with no mechanism to incorporate the external signals that actually drive demand in FMCG.
Festivals, local events, regional weather patterns, and competitor promotional activity were all known to move SKU-level demand significantly — but none of these signals were systematically built into the forecast. The result was a model that performed reasonably on stable, predictable SKUs and failed badly on anything subject to external influence.
The consequences were cascading. High-demand periods ended in stockouts that damaged retailer relationships and cost the business sales it could not recover. Post-promotion periods left warehouses overstocked with product approaching expiry. And the finance team was making working capital decisions based on forecasts they privately acknowledged they did not trust.
The Solution
A hybrid forecasting architecture built for FMCG complexity
Seven Billion began with a Phase 0 discovery session to map the existing forecasting process end-to-end — identifying which data sources existed, how they were being used, and where the largest forecast errors were occurring. The analysis confirmed that the highest-impact opportunity was not rebuilding the core model, but integrating the external signals the team was currently ignoring.
A hybrid forecasting architecture was designed combining three layers of signal: internal historical data including SKU-level sales, distributor sell-out records, and promotional calendars going back three years; regional contextual signals covering local festival calendars, weather data by geography, and retail traffic indicators; and market signals including category-level demand indicators and competitive promotional activity sourced through retail intelligence feeds.
XGBoost was selected as the primary forecasting engine for its ability to handle the high dimensionality of the combined signal set. Azure ML managed model training, versioning, and scheduled retraining cycles. Azure Data Factory orchestrated the pipeline, pulling from eight source systems and delivering clean, consolidated inputs to the model on a daily basis. Output was delivered through a demand planning dashboard showing SKU-level forecasts at regional granularity, alongside confidence bands and the primary drivers behind each forecast — allowing planners to apply informed judgment where needed, and trust the model where it was performing reliably.
The Results
Measurable impact from the first quarter
Demand forecast accuracy improved from sub-40% to an average of 64% across the product portfolio, with the highest gains on festival and promotion-period SKUs where external signal integration had the most influence. Inventory stocking efficiency improved by 86%, measured as the reduction in instances where stock availability deviated significantly from the optimal level given actual demand.
Stockouts dropped by 32% in the first quarter, with the most significant reductions in the highest-velocity SKUs in metro markets. Rs. 5 Cr in inventory capital was freed over the first six months as the reduction in safety stock requirements translated directly into lower warehouse inventory levels.
The supply chain team's weekly firefighting meetings — previously used to manage active stockout situations — were replaced by proactive planning sessions working from the forecast dashboard. The shift from reactive to predictive planning was felt immediately across the business.
The difference after the Seven Billion model went live was immediate — not just in the numbers, but in how the team operates. We plan now. We do not just react.
Head of Supply Chain Planning, Crestfield Foods
KEEP READING
Explore more studies on AI, analytics, and enterprise intelligence.
