Tutorials#
Each tutorial is one problem worked start to finish, with every command and the output it produces. Work through one in a terminal and you end up with a fitted model and a scored forecast.
The user guide covers one command group at a time. These pages follow a single problem through, which is where you see how the commands fit together.
Fetch a real price series, hold out a year, backtest four forecasters against a baseline, and score the winner on data it never saw.
Six channels of smartwatch data, four activities. Use registry tags to rule out estimators that can’t read it, then fit and predict.
Install the CLI and its bundled skill file, check the agent sees it, then hand it a forecasting task.
Which one first#
If you have never run sktime-cli, start with the
Quickstart. It is shorter and covers the commands these
pages assume.
After that:
- Forecast crude oil prices
Read this one if you read only one. It covers the full loop, from a downloaded CSV to a scored forecast, and spends most of its time on holding data back, checking a baseline, and reading a backtest.
- Classify motion from wearable sensors
Read this if your data is many short series rather than one long one. It runs offline in under a minute, and shows how registry tags answer “will this estimator work on my data” before you spend a run finding out.
- Set up an agent to drive the CLI
Read this if you would rather describe the task than type the commands. It is also the quickest way to see the JSON contract in use.