Forecasting projects spend their energy on model choice and their trouble on data. A more elaborate algorithm rarely rescues unreliable inputs; it learns the unreliability. A short set of checks, run before any modeling, catches a surprising share of the problems that later get blamed on the model.
The five gates
- Completeness — are there gaps in the time series, missing fields or blank values that follow a pattern? A system outage that left a week of zeros looks, to a model, like a collapse in demand.
- Consistency — do definitions stay the same over time and across sources? A metric that changed meaning when a product launched or a system was replaced creates a break that isn't real.
- Uniqueness — are there duplicates inflating totals? Merged systems and retried imports are common sources.
- Plausibility — do values fall in sensible ranges, with outliers explained? A price of zero, a negative quantity or a date in the future needs a decision, not silent inclusion.
- Freshness and lineage — how recent is the data, and can you trace each figure to its source? A forecast built on last quarter's snapshot, from an extract nobody owns, is a risk.
Write down every adjustment you make — removed outliers, filled gaps, reconciled definitions — and who approved it. When a number is challenged later, the record is the difference between an explanation and an argument.
Check the business events, not just the numbers
Ask the people who run the process what changed during the history you are using: a price change, a promotion, a system move, a policy shift. These events explain patterns the data alone can't, and they decide which periods are comparable.
Make it a gate, not a one-off
Automate these checks so they run every time the data refreshes, and stop the forecast with a clear message when one fails. It is better for a dashboard to say 'data not updated, forecast withheld' than to show a precise-looking number built on a broken feed. Trust in forecasting is earned mostly by what the system refuses to publish.