Zansa › Answers › Forecasting and data science
How much historical data do I need to build a useful forecast?
It depends on the method and seasonality, but a common rule of thumb is at least two full seasonal cycles, so about two years for yearly seasonality, to estimate seasonal patterns. Simple methods can work with less. Complex machine-learning models generally need more history, more observations or many similar products to learn from.
With only a year of data you cannot tell whether a December peak is seasonal or a one-off. With a few months you can still forecast short-term using averages and recent trends, but you should expect wide uncertainty.
More important than length is relevance and quality. Five years of data from a different business model may be less useful than eighteen months of clean recent data. Pooling similar products can also help new or slow items borrow strength from the group.
Want this applied to your business? Zansa builds dashboards and forecasts for small and mid-sized companies worldwide. Start with a free intro call.
Book a call