Predictive Inventory Management: Eliminating Stockouts and Excess for a Global FMCG Distributor

The Challenge

A global FMCG distributor faced significant challenges in managing its inventory across a sprawling network of warehouses and distribution centres. Frequent stockouts led to missed sales opportunities, while excess inventory resulted in increased storage costs and wastage. Traditional methods of inventory management, relying heavily on historical data and manual adjustments, could not keep pace with dynamic market demands and rapid product turnover unique to the FMCG sector.

The Solution

Adyantrix deployed a cutting-edge predictive inventory management solution tailored to address these complex challenges. Our team harnessed the power of advanced data analytics and machine learning algorithms to transform raw inventory data into actionable insights. By integrating real-time data feeds from sales, warehouse operations, and market demand sensors, we developed a predictive model capable of accurately forecasting inventory needs.

Our solution included several key features:

1. Demand Forecasting: Leveraging machine learning, we automated demand prediction using patterns drawn from historical sales data, seasonal trends, and real-time market fluctuations.

2. Prescriptive Analytics: By recommending optimal order quantities and timings, the system mitigated the risk of stockouts and reduced unnecessary inventory holding.

3. Dynamic Replenishment: Automated inventory replenishment was driven by real-time insights, ensuring products were redistributed efficiently across various locations.

4. Dashboard Visualisation: A comprehensive dashboard provided at-a-glance visibility into inventory levels, demand forecasts, and order tracking, allowing managers to make informed decisions swiftly.

Key Results

The implementation of the predictive inventory management system revolutionised the distributor's logistical operations, leading to tangible improvements:

• Stockout Reduction: The accuracy of the predictive models decreased stockouts by 38%, significantly enhancing product availability and improving customer satisfaction.

• Wastage and Excess Inventory: Reduction in excess inventory led to cost savings of approximately 25% in storage-related expenses.

• Inventory Turnover Ratio: The improved demand forecasting and inventory control increased the inventory turnover ratio by 24%, indicating a more efficient inventory management process.

• Operational Efficiency: The intelligent automation of inventory management processes freed up managerial bandwidth, allowing executives to focus on strategic growth initiatives.

• Improved Responsiveness: With the ability to adapt dynamically to market changes and consumer behaviour, the distributor achieved a more agile and resilient supply chain.

Through the deployment of our predictive inventory management solution, Adyantrix enabled this global FMCG distributor to not only align its logistics operations with contemporary challenges but also position itself as a pioneer in logistics efficiency. This initiative underscores the transformative power of harnessing technology and data analytics in resolving age-old logistical dilemmas.

Technical Approach

The solution was architected as a modular data platform built on Google Cloud Platform (GCP), chosen for its managed data services and the distributor's existing GCP footprint. The pipeline ingested data from four primary sources: the ERP system (SAP S/4HANA) via SAP Data Services connectors, point-of-sale telemetry from retail partners via REST API webhooks, warehouse management system (WMS) stock movement events via Kafka streams, and external market signals including weather data, promotional calendars, and regional economic indicators via third-party APIs.

Data was landed in raw form in Google Cloud Storage, transformed and modelled in BigQuery using dbt (data build tool), and served to forecasting models via a feature store built on Vertex AI Feature Store. This architecture provided a single version of truth for both operational reporting and ML inference.

For demand forecasting, we evaluated four candidate model architectures on holdout historical data:

• SARIMA (baseline): Mean Absolute Percentage Error (MAPE) of 19.4% on 28-day forward forecasts.
• Facebook Prophet: MAPE of 14.2%, with strong handling of seasonal decomposition and holiday effects.
• LightGBM with lag features: MAPE of 11.8%, benefiting from the ability to incorporate external regressors such as promotional flags and competitor pricing indices.
• Temporal Fusion Transformer (TFT): MAPE of 9.3% on the full SKU-location matrix, selected as the production model for its interpretability and ability to quantify forecast uncertainty bounds — a critical feature for safety-stock calculations.

The TFT model was trained on 36 months of historical data at SKU-by-distribution-centre granularity, covering approximately 14,000 active SKUs across 22 distribution centres. Weekly retraining was automated via Vertex AI Pipelines, with model performance monitored using custom drift detection logic that triggered alerts when MAPE degraded by more than 3 percentage points on rolling 4-week windows.

Replenishment recommendations were generated by a downstream optimisation layer using linear programming (PuLP library) that translated probabilistic forecasts into order quantities, taking into account supplier lead times, minimum order quantities, storage capacity constraints, and target service levels per product category.

Implementation Highlights

The 16-week implementation programme surfaced challenges that are characteristic of large-scale FMCG data projects.

Data quality was the most time-consuming workstream. The initial data audit revealed that approximately 23% of historical sales records contained warehouse location codes that had been deprecated and replaced over the preceding three years without back-filling the historical records. This made it impossible to construct continuous SKU-location time series from the raw ERP data without a significant data remediation exercise. We invested four weeks in data cleaning and location mapping, building a lookup table validated against physical goods-receipt records, before any modelling could begin.

SKU proliferation presented a scalability challenge. The distributor managed over 14,000 active SKUs, with significant variation in demand patterns — from high-volume staple products with stable, predictable demand to promotional lines with sparse, intermittent transaction histories. A single model architecture could not optimally serve both. We implemented a SKU segmentation framework using ABC-XYZ analysis to classify each SKU by volume tier (A/B/C) and demand variability (X/Y/Z), then applied different forecasting strategies per segment: TFT for high-volume stable lines, Croston's method for intermittent demand SKUs, and a simpler weighted moving average for very low-volume tail products.

Integration with the replenishment execution workflow required building a two-way interface with the SAP ERP purchase order module. Rather than automating order creation directly (which the client was not yet comfortable with), we surfaced recommendations as draft purchase orders in SAP that buyers could review and approve with a single click. This "human-in-the-loop" design significantly accelerated user adoption compared with a fully automated approach.

Measurable Outcomes

Against the pre-implementation baseline established during a 12-week measurement period:

• Stockout rate: Reduced from 6.2% of SKU-days to 3.8% — a 38% improvement, translating to an estimated £2.1 million in recovered revenue from previously lost sales in the first year of operation.
• Excess inventory value: Average on-hand excess stock (inventory held beyond 30-day demand cover) fell by 25%, releasing approximately £8.4 million in working capital that had previously been tied up in slow-moving stock.
• Inventory turnover ratio: Improved from 8.1x to 10.1x annualised, a 24% uplift reflecting faster stock velocity across the network.
• Forecast accuracy (MAPE): The production TFT model achieved a steady-state MAPE of 9.8% on 28-day forward forecasts against a pre-deployment baseline of 22.4% from the manual forecasting process — a 56% improvement in forecast accuracy.
• Write-off costs: Perishable product write-offs (stock reaching best-before date before distribution) fell by 31%, driven by the system's improved visibility into stock ageing across the warehouse network.
• Buyer time on forecasting tasks: Reduced by approximately 60%, as manual spreadsheet-based forecasting was replaced by reviewing system-generated recommendations — freeing senior buyers to focus on supplier relationship management and category strategy.

Lessons Learned

Invest heavily in data quality before building models. The four-week data remediation sprint consumed nearly a quarter of the overall project timeline, yet it was the single most important investment made. Models trained on clean, well-labelled historical data consistently outperformed those trained on raw ERP exports, regardless of architectural complexity. In future engagements of this type, we now schedule a dedicated data audit and remediation phase as a fixed workstream before the modelling sprint begins.

Segment your SKU portfolio before choosing a modelling approach. Applying a single sophisticated model uniformly across all 14,000 SKUs produced results that were excellent on high-volume lines but poor on intermittent demand products. The ABC-XYZ segmentation framework allowed us to match modelling complexity to data richness — using simpler, more robust methods where data was sparse, and reserving the computationally expensive deep learning approach for products where it delivered genuine accuracy gains.

Change management is a parallel workstream, not an afterthought. The transition from manual spreadsheet forecasting to algorithm-driven recommendations required buyers to change deeply ingrained working habits. Structured training, a phased rollout starting with the most analytically confident team members, and a transparent "explainability" layer in the dashboard that showed buyers why the system was recommending a particular order quantity were all critical to achieving genuine adoption rather than superficial compliance.

Why This Approach Worked

The solution succeeded because it was designed around the practical realities of FMCG distribution rather than theoretical data science ideals. Several deliberate choices made the difference.

Using an uncertainty-aware model (TFT) rather than a point-forecast model meant that safety-stock buffers could be calibrated to actual forecast confidence rather than to arbitrary rules of thumb. For high-uncertainty SKUs, the system automatically recommended higher safety stock; for stable, predictable products it recommended leaner buffers. This dynamic calibration, invisible to end users but embedded in the optimisation layer, was responsible for the simultaneous improvement in both stockout rates and excess inventory — outcomes that traditionally trade off against each other.

Keeping humans in the loop at the replenishment approval stage preserved buyer expertise and institutional knowledge whilst capturing the efficiency gains of algorithmic recommendation. Buyers frequently identified exceptions — an upcoming promotion not yet flagged in the system, a supplier reliability issue based on informal knowledge — that improved recommendation quality when fed back as adjustments. These adjustments were logged and used as additional training signals in subsequent model retraining cycles, creating a virtuous feedback loop between human expertise and machine learning that continued to improve system accuracy beyond the initial deployment.

Speak with our Data Engineering team at Adyantrix to find out how we can support your next project.

Work with Adyantrix

If you are looking to tackle a similar challenge, Adyantrix has the expertise to help across the full project lifecycle. Our data engineering practice covers pipeline design, streaming, and data infrastructure. Our data analytics practice covers BI reporting and self-serve analytics platforms. Our business intelligence practice covers BI strategy, platform selection, and dashboard delivery. Our AI & machine learning practice covers ML model development, MLOps, and intelligent automation. Get in touch to discuss your requirements — no commitment required.