Non-Stationarity and Retraining
"Your model was trained on ten years of data. How do you deal with the fact that markets change?" "Retrain it often" is not enough. A complete answer names the kind of change and treats every training window as a bias-variance choice. It puts a number on how much data a weighting scheme really uses. And it says how you would detect that the model has stopped working.
Two kinds of change
A supervised model learns from the joint distribution of features and target:
Either factor can move.
- Covariate shift: changes and stays the same. Volatility doubles, so features take values that were rare in training, but the relationship holds. The model is extrapolating. The remedies are inputs that stay in range, such as features scaled by current volatility or ranked across assets, and training weights that favour conditions like today's.
- Concept drift: changes. A signal gets crowded and its payoff shrinks, or a relationship flips sign in a new rate regime. No rescaling of the inputs fixes this. Only newer data can.
The rest of this lesson is for subscribers
Unlock every lesson in Machine Learning for Quantitative Research, and every other premium course.
Subscribe to continueTest your knowledge
Questions are only available to subscribers.
Keep reading Machine Learning for Quantitative Research
27 lessons in this course, and every other premium course, on one subscription.
- Every lesson in every course, with the worked examples and interactive simulators
- Graded questions on every lesson, with explanations for the wrong answers as well as the right one
- The trainers, timed assessments and brainteaser library that go with them