openskills.info
Course catalog
Time Series Analysis and Forecasting logoCourse Preview

Time Series Analysis and Forecasting

Time series analysis and forecasting is the practice of studying data recorded in time order, such as sales per day or requests per second, to describe its structure and estimate its future values. It splits a series into trend, repeating seasonal cycles, and unexplained variation, then fits a model that projects the next values with a range of uncertainty around them. It matters since methods that assume independent, shuffleable rows mislead when each observation depends on the ones before it.

itData engineering and analytics

Don't Panic — Time Series Analysis and Forecasting

Here is the honest version. A time series is data with the clock left in: units sold each day, requests each second, revenue each month. Analysis is the job of naming the structure hiding in that order. Forecasting turns the structure into a guess about where the series heads next, bundled with a frank admission of how far off the guess could be.

Why it needs its own field: most statistics assumes you can shuffle the rows without losing anything, and here you cannot. Tuesday resembles last Tuesday. This minute resembles the minute before. The older move was to eyeball the last few points and extend the line, which is fine until a curve shows up. Almost every technique that follows is a way of not cheating against the clock.

Three ideas carry the weight. The first is decomposition: split the series into a slow-moving trend, a repeating component tied to the calendar, and whatever noise is left once those are subtracted. The second is stationarity: several classic models want a series whose behavior is the same in every window, and the usual way to get one is differencing, which swaps each value for the amount it changed. The third is the humble benchmark. Repeat the reading from exactly one season back and you have a seasonal naive forecast; it is irritatingly accurate on real data, and a fancier model that fails to beat it has bought you nothing.

The counterintuitive bit: bigger models are not a free upgrade. Two large open contests, known as M4 and M5, found that off-the-shelf machine learning trailed classical methods such as exponential smoothing and ARIMA when the task was a broad set of separate series, and only took the lead once there were thousands of related series to learn from together. So "which model" really means "how many series, and do they rhyme with each other".

One more thing the jargon hides: a single predicted number, on its own, is nearly useless for planning. What a planner can act on is the prediction interval, the band the true value should sit inside with some stated probability. That band grows wider the further out you look, because the unknowns pile up, and it leans on assumptions about the leftover noise. Break those assumptions and the band shrinks below where it should be, and the forecast starts sounding surer than it has any right to.

Where to go from here. Slides and Intro walk through the model families and how they connect. The Cheatsheet is the reference card: the ETS letter codes, the ARIMA parameters, the error-metric formulas, and a table of warning signs for a model that has quietly gone off. Practice Reference and Exercise turn all of it into a Python routine, from a first decomposition plot to a rolling backtest scored against that stubborn seasonal naive. Field Notes is where the bruises are recorded: why the benchmark is the actual deliverable, why digging up what the data looked like at the time beats arguing about models, and why a clean-looking Prophet chart has not earned your trust until you have tested it.

Where this skill leads

Relevant careers

See how this topic contributes to broader role-level skill maps.

Sources