Skip to content
AI360Xpert
Core ML

Visual explainer

Time Series Forecasting (ARIMA)

Time series forecasting predicts future values by breaking past data into trend, seasonality, and residual noise.

Time series forecasting predicts future values based on past ordered observations.
Time series forecasting predicts future values based on past ordered observations.

Unlike standard data where each row is independent, time series data is a journey. To guess what happens next, we look at the sequence of steps taken so far.

Splitting the Signal

A time series can be decomposed into trend, seasonality, and residual noise.
A time series can be decomposed into trend, seasonality, and residual noise.

A raw time series is messy, but we can decompose it into three simpler parts: a general direction (trend), repeating cycles (seasonality), and random fluctuations (residual noise).

Predicting from the Past

The ARIMA model combines past values and past errors to predict the next value.
The ARIMA model combines past values and past errors to predict the next value.

Once trend and seasonality are handled, ARIMA focuses on the noise. It uses recent past values (Auto-Regressive) and recent past prediction errors (Moving Average) to estimate the next step.

The Forecast

The further into the future we predict, the wider the confidence band of uncertainty becomes.
The further into the future we predict, the wider the confidence band of uncertainty becomes.

Because every prediction carries a small error, forecasting further ahead means accumulating those errors. The model gives a best guess surrounded by a confidence band that widens over time.

Where It Breaks

Unpredictable external shocks can break the patterns, causing the forecast to fail completely.
Unpredictable external shocks can break the patterns, causing the forecast to fail completely.

Time series models assume the future will behave like the past. A sudden, unprecedented external shock—like a pandemic or a policy change—breaks this assumption, rendering the forecast useless.

The Quick Version

  • Sequence matters: Data points depend on the points that came before them.
  • Decomposition: Signals are split into trend, seasonality, and noise.
  • ARIMA: Models the remaining noise using past values (AR) and past errors (MA).
  • Uncertainty: Confidence bands grow wider the further ahead we look.
  • Fragility: Sudden structural breaks ruin the prediction completely.

What to Read Next