Almost every organization runs on questions about the future. How many units will we sell next month? How much electricity will the grid draw at 7 p.m. tomorrow? How many support tickets should we staff for? How many servers must be warm before the traffic spike? These are all time series forecasting problems, and in 2026 the tools to answer them span everything from a one-line seasonal-naive benchmark to pretrained foundation models that forecast unseen series with no training at all. This course teaches you to move confidently across that whole range and, just as importantly, to know which tool a given problem actually deserves.
Educational note: This course is for learning. Any data you use must respect data-protection law such as the GDPR when it contains personal data, dataset licenses, and confidentiality obligations. Where we touch financial series, treat every forecast as analysis and never as investment advice — no model in this course, or anywhere else, guarantees returns. Forecasts are probabilistic statements about the future, not certainties, and communicating that uncertainty honestly is part of the job.
What makes time series different
A time series is a sequence of observations indexed by time, usually at regular intervals: hourly, daily, weekly, monthly. That ordering is not decoration. It is the whole point. In an ordinary supervised-learning table you may shuffle the rows freely, because each row is assumed independent and identically distributed. In a time series the rows are not independent: today's value is strongly related to yesterday's, and the very act of shuffling destroys the signal you are trying to model.
This single fact has deep consequences that we return to again and again:
- You cannot shuffle for cross-validation. Randomly splitting rows leaks future information into the past. Evaluation must respect the arrow of time, which is why this course dedicates an entire module to temporal cross-validation and backtesting.
- Autocorrelation is the signal. The correlation of a series with its own past (its lags) is often the strongest predictor you have. Where a fraud model looks for informative columns, a forecaster looks first at the series' own history.
- The data-generating process drifts. Trends grow, seasonal patterns shift, promotions and pandemics reshape demand, and regimes change. A model trained on last year may quietly go stale, which is why deployment and monitoring are part of the curriculum and not an afterthought.
A concrete illustration: suppose you model daily coffee sales at a chain of cafes with a standard regression, shuffling rows into train and test. Monday rows from June 2026 land in training while Friday rows from March 2026 land in test — the model has effectively seen the neighborhood of every test point. Offline error looks tiny. Deployed, the model faces a genuinely unseen future week and the error triples. Nothing was wrong with the algorithm; the evaluation was dishonest. Avoiding that dishonesty is a recurring theme of this course.