←Back

Time Series & Regression Supply Forecasting Model

2021 · Time Series · Regression · Our World in Data

Impact

  • Graded 8.5/10 as an MSc thesis (Transport, Infrastructure & Logistics, Delft University of Technology); the underlying model was adopted by Heineken beyond the thesis scope
  • 15% lower MAPE than the incumbent projections, which had systematic upward bias from stakeholders hedging against shortage risk
  • Presented results to the Chief Procurement Officer; extended from cans to multiple product groups after the initial rollout
  • Confidence intervals were integrated directly into the stochastic supply allocation model

Business Problem

This started as an MSc thesis, done during a Data Science internship at Heineken. Heineken was systematically overforecasting can supply needs. The motive was understandable: avoid any shortage risk. But the result was excess inventory across the supply chain, which is expensive.

The harder challenge was that the business believed specific events would meaningfully impact demand, and wanted those effects modelled in isolation so they could run scenario simulations. A simple time series forecast was not enough.

Solution

For the thesis, I combined time series models with a regression-based event impact layer into a hybrid forecasting approach. Historical supply delivery data was processed through exponential smoothing for the base trend, then passed through a regression layer that modelled event effects separately: COVID-19 restrictions using external signals from Our World in Data and Google Maps mobility data, and internal promotions from the calendar.

Output was supply forecasts with confidence intervals, fed into a stochastic supply allocation model and used directly for scenario simulation per event type.

Architecture

Supply forecasting system architecture
architecture_v1.svg

Technical Challenges

Regime shifts from COVID. Standard time series models assume stationarity. Lockdowns and reopenings created sudden, non-gradual breaks in the delivery series that the time series component alone couldn't anticipate. The regression layer was added specifically to handle these: external signals like hospital admissions and mobility data acted as leading indicators the time series component couldn't see on its own.

Isolating event effects. Promotions, COVID restrictions, and macro trends all move the series simultaneously. Separating each effect required careful feature design in the regression layer to avoid collinearity and ensure the isolated impacts were interpretable for scenario simulation. If the features weren't independent, the scenario outputs became meaningless.

Overforecasting as the baseline. The existing projections had systematic upward bias as a hedge against shortage risk. Evaluating model quality meant accounting for this: the target wasn't simply matching historical actuals but producing forecasts that were materially closer to actuals than an incumbent approach designed to be deliberately conservative.

Status

  • MSc thesis delivered in 2021, graded 8.5/10, done during a Data Science internship at Heineken
  • Deployed for cans; extended to multiple product groups in collaboration with the analytics department
  • Confidence interval output handed off to the stochastic supply allocation model as a one-off integration

Next Steps

  • Live CI integration: replace the one-off handoff with a live input to the stochastic supply allocation model so forecast updates automatically propagate to allocation decisions

Code is proprietary; happy to walk through the architecture in detail.