Volatility Forecasting Flashcards
Which model should I adopt for the conditional mean, conditional variance and innovation distribution?
Conditional mean: doesn’t matter Innovation distribution: doesn’t matter Conditional variance: matters, should use QL and MSE losses to rank model options.
How should I evaluate different models?
Average out of sample of the loss
Why are the MSE and QL the “right” loss functions?
The MSE and QL give the same ranking for the proxied and true variances
Why do the rankings provided by the MSE and QL loss functions for the proxied variance coincide with the rankings for the true variance?
If you compare two forecasting methods a and b and difference their losses, the term responsible for ranking errors goes to 0 in probability as the number of forecasts increase. (slide 17 for derivation)
Is QL or MSE preferred?
QL
Why do we prefer QL to MSE?
QL is based on multiplicative errors, which are homoskedastic. MSE is based on additive errors, which are heteroskedastic.
Equal Predictive Ability Test
Assesses if the predictive ability of two forecasts is significantly different.
Problem with EPA test
The difference in losses is not iid.
What do we do about the problem with EPA?
Use an estimator of the difference in losses that is robust to serial dependance.
Eg, Newey-West estimator.
How often should I update my ARCH parameters?
Roughly weekly or biweekly.
