Skip to content
Predict.ai

Fundamentals

Time series forecasting, explained

Time series forecasting predicts the next values of something that was recorded in order — sales by day, load by hour, tickets by week — by learning from its own past, and from anything that regularly moves with it.

Updated Aug 18, 2026·8 min read

What a time series is

A time series is a value recorded in order. Tuesday’s orders sit next to Monday’s on purpose. Shuffle those rows and you still have an average. You no longer have a Monday.

Businesses run on series even when the BI tool pretends they run on pivots: load by hour, claims by week, visits by day. If the next decision is “what happens next on this clock,” you are in forecasting country. If the decision is “which accounts will churn,” you want a score. Different clock, different metric.

One series, four layers

Time series decomposed into trend, season, and leftoverObservedTrendSeasonLeftover
Most business series are a slow drift, a repeating clock, and leftover noise. Models differ in how explicitly they split these. You still want to see the split.

Trend, season, leftover

Most series you will actually ship are a slow drift, a repeating clock, and leftover noise. The café that sells twice as much coffee at 8am as at 3pm has seasonality. The app that added users every month for a year has a trend. The Tuesday a celebrity tweeted your URL is leftover — unless you plan to hire that celebrity on a schedule.

You do not always have to split these by hand. Some models swallow the raw path. You still want to look at the split, because a trend you did not notice will get credited to a promo you did.

Why order matters

Random train/test splits are how image models work. They are how time-series models cheat. Train on a Wednesday to predict the Tuesday before it and your notebook will look like a miracle. Live serving cannot go backwards. Walk the origin forward. Hide the future.

Gaps and timezones are part of the series. A DST change, a store that closes Mondays, a feed that arrives a day late — those are not “data quality tickets.” They are the clock. Make them explicit or the model will invent a season that is actually a file delay.

How models read a series

The core is lags: yesterday, last week, last year. Rolling means summarize recent level. Calendars mark holidays. Covariates explain leftovers the lags cannot. Foundation models skip some of the handmade columns and read a longer raw window instead. They still cannot use a number you will only know tomorrow.

  • Univariate: the target’s own past.
  • With drivers: price, promo, weather, known ahead.
  • Panel: many series that share a clock (every store).

When not to treat it as a series

If you have twelve annual points and a regime change in the middle, you do not have a series. You have a story. If the “timestamp” is an invoice date that can be backdated, you have accounting, not a clock. And if the question is who, not when, stop wrapping a classifier in a line chart.

FAQ

What is time series forecasting?
It is predicting future values of a quantity that was measured over time, using the order of those measurements. Yesterday informs tomorrow in a way a shuffled table of the same numbers does not.
What is the difference between time series and panel data?
A single series is one thing over time (one store's daily sales). A panel is many series that share a clock (every store). Most business forecasting is panel data wearing a time-series costume.
Do I need stationarity?
Classical models often want a series whose mean and variance sit still. Many modern and foundation models cope with trend and season as-is. You still want to know which is which, because a trend you did not notice will look like a brilliant forecast until it bends.

Keep going

Try it on your data.

Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.