Supply Forecasting System
2021 · Prophet, Exponential Smoothing, Python, SQL, Our World in Data (COVID regression)
Problem
Breweries were systematically overforecasting supply needs — cans, glass, carton — to avoid any risk of shortage. The result was excess inventory across the supply chain, which is expensive.
The harder challenge: breweries believed the trend would be meaningfully impacted by specific events (promotions, macro trends, COVID-19 restrictions), and wanted those effects modelled in isolation so they could run scenario simulations. A simple time series forecast wasn't enough.
Solution
Hybrid forecasting approach combining time series models with regression-based trend impact. Supply delivery data was read from brewery Excel exports, processed through a time series forecast, then passed through a regression layer that modelled event impacts separately (using external data sources including Google Maps mobility data, hospital admissions, and other COVID-19 signals from Our World in Data). Output was supply forecasts with confidence intervals, synced with a stochastic supply allocation model.
Built initially for cans, designed from the start to scale to glass and carton with the analytics department.
Architecture
Impact & Scale
- Trend forecast was closer to actuals than historic brewery projections
- Built for cans; worked with the analytics department to ensure the approach scaled to glass and carton
- Received 8.5/10 for the research and results; the company started investing in scaling the solution
- The team integrated the confidence intervals directly into their stochastic supply allocation problem
Next Steps
Integration of the confidence interval output into the stochastic supply allocation model as a live input rather than a one-off handoff.