Cutting forecast error in half with a unified data pipeline

A mid-market distribution business was running one of its most important numbers — the demand forecast — across a dozen disconnected spreadsheets. Each region kept its own version, reconciled by hand, and by the time the figures reached leadership, nobody entirely trusted them.

The symptoms were familiar: stock in the wrong places, decisions delayed by "let me check that number," and a planning team spending more time assembling data than thinking about it.

The challenge

• Forecast inputs lived in siloed spreadsheets with no single source of truth.
• Manual reconciliation took days and introduced errors of its own.
• Because the numbers were distrusted, they were quietly overridden by gut feel.
• No one could trace how a given forecast figure had been produced.

What we did

We didn't start with a model. We started with the plumbing.

Unified the data

We consolidated the regional sources into a single governed pipeline — automated ingestion, validation, and a documented flow from raw input to final figure.

Made the numbers trustworthy

Reconciliation checks, freshness monitoring and tests now run automatically, so a broken input is caught before it reaches a decision rather than after.

Rebuilt the forecast on solid ground

Only once the data was reliable did we improve the model itself, with a clear evaluation harness so every change could be measured instead of argued.

The results

• Forecast error cut by roughly half — driven far more by clean, unified data than by a cleverer model.
• Reconciliation dropped from days to minutes, freeing the planning team to analyse instead of assemble.
• One trusted number that leadership stopped second-guessing — the real unlock.

The model got the credit. The data pipeline did the work.

The lesson generalises: most forecasting problems are data problems wearing a modelling costume. Fix the plumbing and the accuracy tends to follow.

Have a forecast nobody trusts? That's usually where we start.