Skip to main content

Forecasting for Data Sciences

Forecasting is predicting future values of a time series such as sales or traffic. The unit covers trend and seasonality, baseline methods, ARIMA and exponential smoothing, machine learning with lag features, and backtesting with rolling origin splits, since random train-test splits are invalid for ordered data.

Stage: Experimentation and Causal Thinking · Part of the Data Science roadmap · Last reviewed 19 November 2026

What to learn in Forecasting

Forecasting interview questions

Why can't you use a random train-test split for time series?

A random split puts future observations in the training set and past ones in the test set, leaking information and giving unrealistic scores. Time series needs splits that train on earlier data and test on later data, ideally repeated with a rolling origin.

What is a seasonal naive forecast and why use it?

It predicts each value as the value from the same point in the previous season, for example last week's same weekday. It is a cheap baseline, so a complex model must beat it to justify its added effort and maintenance.

Where Forecasting fits in the Data Science roadmap

Forecasting belongs to the Experimentation and Causal Thinking stage (intermediate level), which takes about 4 hours. What actually works.

Back to the full Data Science roadmap