Forecasting and Time-Series Basics

What is forecasting?

Forecasting estimates future values using historical observations and relevant context.

A forecast is an estimate with uncertainty, not a promise. Good forecasts include a range (prediction interval), not just a single number.

Understand the time series

Check before modelling:

• Time interval (daily, weekly, monthly)
• Missing periods
• Changes in how data was collected
• Holidays and special events

Start with a baseline

More advanced methods include exponential smoothing (e.g., Holt-Winters), ARIMA / SARIMA, Prophet, and machine-learning models with lag features.

A complex model is only useful if it beats a reasonable baseline on data it wasn't trained on.

Evaluation

• Compare methods over the same test periods.
• Use time-based splits or rolling-origin (walk-forward) validation.

Common mistakes

• Randomly shuffling time-series data: future information leaks into training. Always train on earlier periods and test on later ones.
• Ignoring one-off events that won't repeat.
• Presenting a single number without a range.
• Not comparing with a baseline.

Try it: Plot monthly sales over two years. Identify any trend or seasonality and suggest a simple baseline forecast.