Structured data

Time series forecasting

demand, traffic, revenue, capacity

Future values of many related series at a fixed horizon.

Probability & Bayes

Hypothesis testing

Foundations

Autocorrelation, stationarity, and prediction intervals all rest on probability and hypothesis testing.

Frame the problem

Frame

Horizon, granularity, point or distribution. Decide the unit: per SKU, per store-SKU, per region.

Exploring the data

Hunting for leakage

Data & labels

Seasonality, trend, holidays, structural breaks; plot before modelling. Any feature that uses future information, such as month totals, leaks.

Create features

Represent

Lag and calendar features are the representation. Holidays and promotions as exogenous inputs.

Temporal splits & backtesting

Split

Rolling-origin backtest. Never random.

Dumb baseline

Per-series classical

ETS or ARIMA when there are few series. Prophet when holidays dominate.

Classical time series

One global model

Gradient boosting on lag and calendar features across all series. Usually the winner at scale.

Boosting on lag features

Deep, if warranted

Only with many related series and a need for probabilistic output.

RNN / LSTM / GRU

Deep forecasting

Model

Naive and seasonal naive first. Many series never beat them.

Train

Nothing unusual here.

Forecasting metrics

Evaluate offline

MASE against naive, per horizon. Check interval calibration.

Evaluate online

Nothing unusual here.

Serving & release

Monitor & retrain

Ship & monitor

Retrain on a schedule. Watch for regime changes.