Demand forecasting with external data

Internal sales history only goes so far; the best demand forecasts blend it with external signals. This guide covers which external data improves forecasting and how to use it without tripping over common pitfalls.

Available across the EU. DataSupplier sources and delivers this data in all 27 European Union countries — including Germany, France, Spain, Italy, the Netherlands and Poland — and across the EEA, in the format and cadence you need.

Why external data improves forecasts

Demand is driven by factors outside your own data, weather, mobility, prices, events, the economy. Adding the right external predictors can sharpen forecasts materially, especially for volatile or seasonal demand.

Useful external predictors

• Weather: a powerful driver for retail, energy and food.
• Mobility and footfall: activity near locations.
• Macro and prices: economic and cost context.
• Events and calendars: holidays, events, school terms.

Feature engineering

External data rarely helps raw; it needs aligning to your time grain and geography, lagging appropriately, and transforming into features. Time alignment and location matching are where much of the work is.

Avoiding leakage and overfitting

Two traps dominate: using data that would not have been available at forecast time (look-ahead bias), and adding so many predictors that the model overfits. Point-in-time data and disciplined validation guard against both.

Sourcing considerations

You need history deep enough to train and a live feed matching it for production. Synthetic or historical backfill can bootstrap development. Cadence must match the forecast cycle.

In a managed model

A managed partner can source aligned external predictors with deep history and a matching live feed, ready for forecasting.

Feature engineering and alignment

External predictors rarely help in raw form. They need aligning to your forecast’s time grain and geography, lagging appropriately (yesterday’s weather, last week’s footfall), and transforming into features the model can use. Time alignment and location matching are where most of the engineering effort, and most of the lift, comes from.

Avoiding leakage and overfitting

Two traps dominate: using information that would not have been available at forecast time (look-ahead bias), and adding so many predictors that the model overfits noise. Point-in-time data guards against the first; disciplined validation and parsimony against the second. Match the live feed’s cadence to the forecast cycle, and bootstrap development with historical or synthetic data.

Sources & further reading

• Academic literature on forecasting with exogenous variables.
• Eurostat and ECB: macroeconomic indicators.
• National meteorological services: weather predictors.
• Internal practice: DataSupplier predictor sourcing.