A model can predict historical returns well because its fitting procedure has seen information unavailable at the forecast date. A valid simulation must reproduce the model's learning schedule as carefully as its trading schedule. This lesson turns label availability into training eligibility and separates parameter estimation, model selection and final confirmation.
Learning objectives. By the end of this note you should be able to:
- assign training, validation and test data distinct roles;
- implement an expanding or rolling walk-forward schedule;
- exclude unfinished labels at each fit time;
- keep preprocessing and hyperparameter selection inside their proper windows;
- distinguish a historical replay from a nonchronological scenario evaluation.
In this lesson
- Three roles rather than three filenames
- Training eligibility follows information time
- Expanding and rolling windows
- Fit the whole pipeline inside the window
- Purging, embargoes and final tests
- What to freeze before opening the test
Before you start
The full lesson, with worked examples and exercise solutions, is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or browse the Quantitative Research, Data and Backtesting syllabus.