What makes it special
In a time series, observations are ordered in time and each point often depends on recent past points. This breaks the usual assumption that rows are independent, so methods and validation must respect time.
Classic and modern approaches
- Naive forecasts repeat the last value or last season as a baseline.
- Moving averages smooth recent points to project ahead.
- Autoregressive models predict the next value from previous values.
- Machine learning models use lagged features, rolling stats, and calendar features.
Validation must respect time
- Never shuffle rows, since that leaks the future into training.
- Use a rolling or expanding window where you always train on the past and test on the next slice.
- Report error on truly future periods to mimic deployment.
Useful features
- Lags of the target, such as the value one day or one week ago.
- Rolling means and standard deviations over recent windows.
- Calendar signals like day of week, month, and holidays.
Key idea
Time series forecasting respects the order of time, builds lag and calendar features, and validates with rolling windows so the future never leaks into training.