AI glossary · Basic · Classic ML, recsys & search
Time series forecasting
Time series forecasting analyzes past patterns in data collected in time order (daily sales, hourly electricity use, etc.) to anticipate future values.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Data, ops & platforms
- Added
- 2026-10-08
How it works
Time series data usually contains a long-term upward or downward movement (trend), patterns that repeat at regular intervals (seasonality), and noise that is hard to explain. A forecasting model finds these components in past data, assumes they will continue, and calculates future values. Traditionally, statistical methods that model trend and seasonality mathematically were used; today, machine learning such as gradient boosting and deep learning models are also widely used. Adding outside information such as weather, holidays and sales promotions often makes forecasts more accurate.
Branches of time series forecasting methods
Statistical methodsTrend and seasonality as formulas
- Usable even with little data
- Results are easy to explain
Machine learning and deep learningLearn patterns from data
- Easy to add outside information
- More data is better
Time series foundation modelsPretrained general-purpose models
- Try applying directly to new data
- Need validation on business data
공통 All assume past patterns will continue into the future
Why it matters
Being able to see even part of the future lets you prepare in advance. Retailers use demand forecasts to decide how much stock to hold, and power companies use usage forecasts to adjust generation. It is also used for call center staffing, server capacity planning and financial risk management. Even a small improvement in forecasts can reduce wasted inventory or supply shortages, so it is a field in high demand among businesses.
Things to know
Forecasting rests on the assumption that the future will unfold much like the past. When something unprecedented happens, such as the spread of an epidemic or a sudden policy change, forecasts can be badly wrong; such changes should be treated as a form of data drift, and the model should be checked regularly. For this reason, presenting a forecast as a range, such as "it is likely to fall within this range," is often more useful in practice than a single number. Recently, time series foundation models pretrained on large numbers of time series and able to forecast new data right away have appeared, but you still need to verify for yourself whether they fit your business data.
Example
A convenience store chain combines past sales by store and product with day of the week, weather and local event information to forecast sales for the next few days, and uses the results to set order quantities automatically. A cloud service operations team forecasts traffic by time of day and adds servers before users flood in. It is also used together with anomaly detection: when the forecast and the actual value diverge widely, this is treated as an anomaly signal and an alert is raised.