Solution Delivered
​
↳ Data Consolidation and Feature Engineering 
 
The foundation of the engagement was building a unified data platform that brought together every signal relevant to beer demand into a single, queryable layer. Plexteq designed and deployed an AWS-based data lake ingesting five years of historical sales records across all SKUs, daily POS sell-through data from the brewery’s three largest supermarket partners, weather data from the UK Met Office API correlated to the brewery’s distribution postcodes, a structured UK events calendar covering sporting fixtures, bank holidays, school terms, and festival schedules, and returnable packaging inventory records tracking kegs and crates across customer sites.
​​
Apache Spark running on AWS EMR was used to normalise and join these heterogeneous sources into a consistent time-series dataset at daily and weekly granularity. The feature engineering phase produced over 60 derived variables including lag features (sales 1, 2, 4, and 8 weeks prior), rolling mean and standard deviation windows, degree-day temperature indices, and binary event flags for each fixture in the UK sporting and hospitality calendar. The richness and specificity of this feature set became one of the primary drivers of downstream model accuracy.
 

↳ Demand Forecasting with Ensemble ML Models 

 

With a rich feature dataset established, the Plexteq data science team evaluated several modelling approaches against the brewery’s requirement for 90-day forward forecasts at SKU level. Three model families were developed and benchmarked:

​

• SARIMA - SARIMA (Seasonal AutoRegressive Integrated Moving Average) was implemented as the statistical baseline. It effectively captured the strong weekly and annual seasonality present in the sales data and provided interpretable decompositions of trend, seasonality, and residual noise. SARIMA performed reliably on stable SKUs with long sales histories but was less effective on newer craft lines and the irregular impact of major one-off sporting events.

• XGBoost - XGBoost gradient-boosted decision trees were trained on the full engineered feature set. The model excelled at capturing non-linear interactions — for example, the compounding effect of a bank holiday coinciding with a Premier League match week during a warm spell. Hyperparameter tuning was performed using Bayesian optimisation, and SHAP (SHapley Additive exPlanations) values were applied to surface the most influential features per SKU, providing the planning team with a transparent, human-readable explanation of each forecast.

• LSTM - LSTM (Long Short-Term Memory) recurrent neural networks were trained on the multivariate time-series data using TensorFlow. The LSTM architecture was particularly effective at learning medium-term demand trajectories for the brewery’s core lager lines, where sequential dependencies across 6–8 weeks were strong predictors of upcoming volume shifts. The network was trained with dropout regularisation and early stopping to prevent overfitting on the limited SKU-level history.

​

Final production forecasts were generated using a stacked ensemble that weighted the outputs of all three models dynamically per SKU - higher LSTM weighting on stable high-volume lines, higher XGBoost weighting on seasonal and event-sensitive products. The ensemble was managed and versioned through MLflow, enabling the team to track experiments, compare model generations, and roll back to previous versions when new data caused temporary performance degradation.

​

↳ Returnable Packaging Prediction 

​

A dedicated modelling workstream addressed the returnable packaging challenge. The brewery’s keg and crate inventory was modelled as a flow problem: units leave the brewery with outbound shipments and return on a probabilistic schedule that varies by customer segment. Supermarkets return packaging reliably and rapidly; independent hospitality venues do so inconsistently; some smaller customers had historically never returned materials at all. Plexteq built a customer-segmented return probability model using logistic regression and survival analysis, trained on five years of invoice and credit-note records. Combining the outbound shipment forecast with the return probability model yielded a 90-day net packaging availability forecast, enabling the procurement team to identify future shortfalls with sufficient lead time to negotiate with suppliers at lower cost — a core operational requirement identified at project initiation.

​

↳ Production Planning Integration and Dashboard 

​

The forecasting engine was integrated directly into the brewery’s production scheduling workflow via a purpose-built planning dashboard. The dashboard presented 4-, 8-, and 12-week volume forecasts at SKU level with confidence intervals, flagged upcoming high-demand events on a visual calendar, displayed projected packaging availability against planned production volume, and generated a recommended weekly brew schedule that respected tank capacity and lead time constraints.

 

The system was deployed on AWS using containerised microservices, with automated weekly model retraining triggered by each new data ingestion cycle. PostgreSQL served as the operational data store for forecast outputs and scheduling artefacts; the dashboard was built with a React front-end consuming a FastAPI backend, enabling secure access for both the planning team and senior commercial stakeholders.