by Lukas Gonon, Niklas Weber, Thilo Meyer-Brandis
Continue on SelfMadeQuant
Suggested foundations for risk measurement and stress and financial machine learning. Specialist prerequisites vary by paper.
Represent losses as a distribution rather than one forecast.
Distinguish a loss quantile from the average loss in the tail.
Preserve time order and avoid future information in training.
Establish an interpretable baseline before a flexible model.
Compare risk estimators and their assumptions.
Check forecasts against realised losses.
Account for dependence and overlapping observations.
Connect diversification, common risk factors, investor views, and the properties of a risk measure.
Choose a reproducible strategy, specify the trading rule, and test what remains after costs and estimation uncertainty.