Skip to content

Time Series & Forecasting

Learning from data that unfolds over time — predicting the future and spotting the unusual.

Imagine guessing tomorrow's weather by looking at the last week — if it's been getting warmer each day, tomorrow is probably warm too. That's the idea behind time-series data: numbers recorded in order over time, like daily temperatures, your heart rate, or a shop's sales. Because the points come one after another, the past holds clues about the future, so computers can learn the pattern and forecast what happens next. They also watch for an anomaly — a sudden break from the usual pattern, like your heart rate spiking while you sit perfectly still. Spotting these early can warn us that something is wrong, whether it's a machine about to break, a fraudulent payment, or a storm on the way.

The main ideas

  • Forecasting — Predicting future values — demand, prices, load — from historical sequences.
  • Classical models — ARIMA, exponential smoothing, and seasonal decomposition.
  • Deep forecasting — Temporal convolutions and transformer forecasters like N-BEATS and TFT.
  • Anomaly detection — Flagging outliers and change points in streams and logs.
  • Foundation models for time series — Pretrained, zero-shot forecasters that generalize across domains.
  • Evaluation & backtesting — Rolling-origin validation, horizons, and avoiding lookahead leakage.

Machine Learning · Deep Learning · Applications & Industry

Want to make things?

Head to AI School — AI camps where kids build their own games.