Chapter 2 · Machine learning basics
Forecasting over time
Forecasting predicts future values from patterns in past ones: trend, seasons and noise.
Chapter 2 · Machine learning basics
Forecasting predicts future values from patterns in past ones: trend, seasons and noise.
Sales over time look wobbly…
Most real series are a mix of a slow trend, a repeating season and random noise. A forecast learns the first two and admits it cannot know the third.
A time series is one measurement taken again and again: daily sales, monthly rainfall, hourly power use. The order matters, so we cannot shuffle it.
The line is 36 months of invented sales. The forecast learns from it and predicts the next 12 months. Tick the box to see the real pattern behind the noise.
Simplified teaching version: a straight line plus a 12-month wave. Violet is the past, magenta the forecast, the pale band its likely range.
Sales rise every December, year after year. Which part is that?
Why does a forecast band get wider further into the future?
What can break even a very good forecast?