Robust statistical jump models

Abstract:
Statistical Jump Models are a class of regime identification models,
similar to Hidden Markov Models, with the added advantage of facilitating
large-scale integration of exogenous variables that influence the
latent regime-switching process. Recent advancements in SJMs include
the introduction of automatic feature selection techniques, which eliminate
irrelevant exogenous information and enhance the model’s ability to
identify regimes more effectively.
We generalize the regular Statistical Jump Model by considering robust
distances measures, while preserving the feature selection. This is
important for financial data, as these are often heavy-tailed.
The simulation study confirm that the proposed modified algorithm
is outperforming the regular statistical jump model on heavy-tailed data
but not on light tailed data, as predicted by statistical theory. We also
find that the corresponding latent regime process is more persistent when
estimated on index returns than comparable estimates derived using Hidden
Markov Models or the regular statistical jump model, an obvious
advantage for any trading strategy based on the latent regime process.