5.21 Time-series forecasting
You can forecast a real series and validate it without cheating.
Before:03. Data Handling & AnalysisUnlocks:06. Deep Learning10. Production & MLOps13. Capstones, Portfolio & Interviews
Forecasting is supervised learning where time is the axis and the rules change: stationarity, seasonality, and features that may only use the past. Classical ARIMA and boosted trees over lag features both remain competitive, which is itself informative. It sits here as the specialised craft classical ML feeds into. The non-negotiable is backtesting with a rolling origin — random splits or features that peek forward produce forecasts that only ever work in the notebook.
Work through these
Stationarity, trend, seasonality, differencing
Whether a series' behaviour is stable over time, and the trend and seasonal components that usually make it not. Differencing is the standard way to remove them, and knowing why is part of the item.
ARIMA/SARIMA and exponential smoothing
The classical statistical forecasting families, which remain strong baselines and are frequently hard to beat on short series. Skipping straight to machine learning here is a common mistake.
ML approaches: lag features, gradient boosting
Turning a time series into a table of past values and applying a general-purpose model. This is what wins on many practical problems, and constructing the features correctly is the whole trick.
Backtesting with rolling origin
Validation must respect time: train on the past, test on the future, and repeat as the origin moves forward. Any other split lets the model see the future and produces a fantasy score.
Sign in to keep your progress.
Free resources
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.