Bakery Demand Forecasting for Beck Roman In progress

Executive Summary #

Client: Beck Roman A traditional bakery in Canton Schwyz with multiple branches.
Goal: Replace gut-feeling daily bake decisions with per-product demand forecasts, reducing waste without losing sales.
Tech Stack: Python, Scikit-learn (HistGradientBoosting, quantile loss), Chronos-2 (foundation model), Pandas, DVC, GraphQL APIs (POS + ERP), Weights & Biases, Open-Meteo, Docker.

The Challenge #

Every morning, the bakery decides how much to bake of each product. Bake too little and customers leave empty-handed; bake too much and the surplus becomes waste. The existing decision relies on experience: robust but unscalable, and blind to subtle drivers like weather, school holidays, and local events.

Current Status: In Progress #

This is an active engagement. A detailed write-up will follow once the production rollout is complete.

Data and Models #

• End-to-end data pipeline from the bakery’s POS system (GraphQL API) into a reproducible DVC-managed dataset spanning ~4 years of transaction history.
• Quantile regression for seven products: the framing directly captures the asymmetric cost (a lost sale hurts differently than a return).

Measured Wins #

• The return rate drops from roughly 15-25% to roughly 10-15%. Validated on a 3-month frozen test window.
• The tradeoff is deliberate: the model bakes slightly less, but the savings on unsold ingredients clearly outweigh the missed margin. Net effect: more profit.

A Model Bake-Off, Scored in Francs #

• Four model families competed, measured on the actual cost of stockouts and returns rather than abstract error metrics.
• The winner: a pretrained foundation model (Chronos-2) with no training on the bakery’s data at all. Close to 9% lower mismatch costs than the tuned local models, across all seven products.
• The losers are just as instructive: fine-tuning on the bakery’s own data made results worse, and elaborate neural networks lost to the simple models.

Production Operation #

• A containerised system produces daily forecasts and writes the recommended bake counts directly into the bakery’s ERP / production-planning system via its API.
• An automated daily email additionally keeps the bakehouse in the loop.
• Guarded by sanity bounds, an audit log, and a kill switch. Currently running in supervised test mode.

The full write-up with charts and architecture details will be published here once the rollout across all products and branches is complete.