Portfolio Allocation and Testing in Markov Switching Models

Constructing a Real-Time Regime Indicator for Asset Allocation: Modeling regimes directly from multiple asset class returns is a numerically challenging exercise. Here, we present an alternative approach to classifying regimes for a large number of assets through the construction of a single real-time regime indicator. The indicator is based on a dynamic factor model, using multi-frequency macroeconomic and market data. A three-state Markov-switching model is found to appropriately  t the extracted latent factor. Its regimes can be interpreted as normal markets, mild recession and severe recessionary periods. These regimes are subsequently shown to capture conditional volatility and correlation changes for six asset class returns. In out-of-sample asset allocation testing, the trading strategy based on this indicator yields an improved portfolio risk pro le relative to the static asset-mix, and compares well relative to the multivariate regime model derived from asset-class returns. Extending the application of the indicator to a broad set of asset class sector returns asserts that the regime classi cation is also useful in capturing sector returns' conditional volatility and correlation characteristics.

Evaluating the Applicable Number of Regimes in Markov-Switching Models via Regression Techniques: While regime switching models have gained great popularity as an approach to model the conditional properties of many  nancial time series, the tests used to determine the number of applicable regimes remain a challenge. At the heart of the problem is the presence of nuisance parameters under the null hypothesis which in essence, invalidates the standard likelihood ratio-based tests. In recognition of this, several authors have proposed alternative testing procedures. However, these are often computationally complex or may require signi cant programming, thus rendering them less accessible to the everyday time series analyst. Here, we examine the plausibility of two simple regression-based tests. Our results show that the original tests have poor size against a Markov-switching model alternative. Once we account for size, the tests have reasonable power for data-generating processes with regime dependent means or intercepts. Nevertheless, the results also reveal that while the test is simple, the empirical critical values vary quite di erently depending on the alternative model being tested. Unlike the two versus one regime results, results for higher number of regimes are promising, and show signi cantly improved power with sample size.