Weekly-retrained demand forecasting for an energy retailer

A UK energy retailer was over-provisioning generation capacity against a demand forecast built on a five-year-old spreadsheet model, with no stated uncertainty. Planning decisions were being made against a single point number nobody fully trusted anymore.

The team had six weeks of budget and needed the new model integrated into an existing planning process, not a parallel system nobody would actually use.

We built a forecasting model validated against two years of historical meter data, deployed it behind an internal API, and set it to retrain weekly against live data. Every forecast ships with a stated confidence interval instead of a single number.

A production forecasting service, an internal dashboard showing forecast vs. actual with the uncertainty band plotted, and a weekly retraining job with drift alerting.