MultiregimeTVTP: Dynamic Regime Switching

• MultiregimeTVTP is a class of Markov-switching models with time-varying transition probabilities that capture regime-dependent mean and volatility dynamics via both endogenous and exogenous inputs.
• It employs flexible TVTP specifications—including endogenous, exogenous, and score-driven mechanisms—to enable likelihood-based inference, simulation, and filtering.
• The framework facilitates accurate regime recovery and forecasting in economic and financial time series, although identifiability challenges may arise with score-driven implementations.

MultiregimeTVTP denotes a class of Markov-switching (MS) models augmented with general time-varying transition probabilities (TVTP) across multiple regimes, specifically tailored to capture regime-dependent mean and volatility dynamics in time series driven by both endogenous and exogenous information. This framework supports $K$ regimes, arbitrary TVTP specifications, and provides likelihood-based inference, robust simulation, and filtering tools, notably as implemented in the R package multiregimeTVTP. MultiregimeTVTP models are crucial for accurately characterizing evolving regime structures, especially when regime switching is influenced by both observed covariates and latent state dynamics (Modée et al., 14 May 2026).

1. Formal Framework of MultiregimeTVTP

The multiregimeTVTP model is defined for an observed univariate sequence $y_1, ..., y_T$ with a latent regime process $\{z_t\}$, where $z_t\in\{1, ..., K\}$. The observation equation under regime $z_t$ is: $$y_t \mid (z_t = i) \sim \mathcal{N}(\mu_i, \sigma_i^2)$$ and regimes evolve as a first-order Markov chain with state-dependent, time-varying transition matrix $P_t = [\pi_{ij,t}]$. Transition probabilities are mapped via a (row-wise) logistic transformation of a latent process $\psi_t$: $$\pi_{ij,t} = \frac{\exp(\psi_{ij,t})}{\sum_{\ell=1}^K \exp(\psi_{i\ell,t})}$$ The evolution of $\psi_t$ encapsulates a variety of specifications:

• Model I (endogenous, lagged outcome): $y_1, ..., y_T$0
• Model II (exogenous covariate): $y_1, ..., y_T$1
• Model III (GAS/score-driven): $y_1, ..., y_T$2, with $y_1, ..., y_T$3, $y_1, ..., y_T$4.

The generic regime-switching filter propagates forward and computes the one-step predictive density and associated log-likelihood: $y_1, ..., y_T$5 with log-likelihood $y_1, ..., y_T$6 (Modée et al., 14 May 2026).

2. TVTP Mechanisms and Model Variants

The modeling strength of multiregimeTVTP arises primarily from the flexibility in specifying $y_1, ..., y_T$7—the driver of regime transitions—which may depend on lagged endogenous variables (Model I), exogenous covariates (Model II), or be directly driven by the score of the conditional density (Model III, GAS). The latter allows transition probabilities to respond to the most recent information via a parameterized updating rule: $y_1, ..., y_T$8 where $y_1, ..., y_T$9 are diagonal $\{z_t\}$0 matrices and $\{z_t\}$1 is scaled by the Fisher information (Modée et al., 14 May 2026). However, the identifiability of the score-driven (GAS) mechanism is severely compromised in practice, as discussed below.

3. Statistical Inference and Software Implementation

Robust maximum likelihood estimation for multiregimeTVTP proceeds via multi-start optimization, likelihood filtering, and profile likelihood analysis. The R package multiregimeTVTP implements all core operations:

• Simulation: Functions such as simulate_data_const, simulate_data_tvp, simulate_data_exo, and simulate_data_gas generate synthetic regime-switching series under prescribed specifications.
• Filtering: Functions like rfiltering_const, rfiltering_tvp, rfiltering_exo, and rfiltering_gas implement the forward-backward filtering, producing filtered regime probabilities and log-likelihood evaluations.
• Estimation: Estimation routines estimate_const_model, estimate_tvp_model, estimate_exo_model, estimate_gas_model provide multi-start maximum likelihood estimation, yield coefficient vectors, standard errors (delta method), filtered regimes, latent process $\{z_t\}$2, and numerical Hessians at the optimum.
• Forecasting: One-step forecasts are computed as $\{z_t\}$3 (Modée et al., 14 May 2026).

Computationally, the package leverages both pure-R and compiled C routines for speed (up to 50x faster for filtering), and parallelizes multi-start estimation via future/future.apply.

4. Identifiability Issues and Monte Carlo Results

Empirical and simulation-based evidence shows marked contrasts between specification types:

• Identifiability for GAS (Model III): The GAS score coefficient $\{z_t\}$4 is typically not identifiable, presenting a flat (ridge-like) log-likelihood surface along $\{z_t\}$5, resulting in MLEs for $\{z_t\}$6 collapsing to zero for nearly all random initializations. Consequently, the fitted model reverts to constant TVTP (i.e., all transitions are time-invariant). This issue is robust to sample size and persists despite extensive random multi-start searches (Modée et al., 14 May 2026).
• Parameter recovery: Regime means $\{z_t\}$7, variances $\{z_t\}$8, and average transition probabilities $\{z_t\}$9 are consistently well recovered for correctly specified models in Monte Carlo studies. Coverage rates for regime means and variances are near nominal for $z_t\in\{1, ..., K\}$0; for $z_t\in\{1, ..., K\}$1, transition probability coverage is slightly sub-nominal (≈75–80%).
• Forecasting robustness: One-step-ahead forecasts are virtually insensitive (within ±1%) to TVTP specification errors, since forecast accuracy dominantly reflects regime mean estimation.
• Filtered probabilities: Accurate specification of the TVTP process is indispensable for recovering the true regime path, especially as $z_t\in\{1, ..., K\}$2 increases, as misspecification induces significant increases in mean-squared and mean-absolute errors of state recovery (Modée et al., 14 May 2026).

5. Empirical Applications in Economics and Finance

In applications to U.S. Treasury zero-coupon yield changes (1961–2024), a $z_t\in\{1, ..., K\}$3 regime model with exogenous TVTP (Model II) driven by the lagged yield level achieves the best empirical fit across all maturities, as evidenced by the lowest log-likelihood and information criteria (AIC, BIC) values. This configuration distinctly outperforms models with constant transition probabilities or those driven only by lagged changes (Model I). The order of estimated regime means and variances aligns with economic intuition regarding volatility regimes.

The GAS model failed to converge for nearly all random starts, reproducing simulation-based identifiability problems. Paths of filtered regime probabilities in the exogenous TVTP fit closely track business-cycle-like episodes and display richer structural alignment than those obtained from constant transition probability models (Modée et al., 14 May 2026).

6. Contrasts with Alternative Multiregime Approaches

Within the broader spectrum of multiregime and time-varying parameter models, multiregimeTVTP distinguishes itself by:

• Allowing for flexible, non-parametric modeling of transition probabilities with exogenous, endogenous, or score-driven dynamics.
• Providing straightforward likelihood-based inference, reproducible simulation, and efficient filtering for arbitrary $z_t\in\{1, ..., K\}$4, directly extensible to both univariate and multivariate contexts as a building block for larger TVP-VARs or MS-GARCH variants (Hauzenberger et al., 2019, Fischer et al., 2021, Chaudhary, 4 Jun 2026).
• Explicitly addressing the practical pitfalls of GAS identification, which differ from Bayesian mixture-based TVP approaches that naturally handle multi-modality in parameter evolution through nonparametric or shrinkage priors.

7. Practical Implementation Guidelines and Software Usage

The core workflow for applied work using multiregimeTVTP encompasses simulation, filtering, estimation, and diagnostic plotting. The package supports high-dimensional filtering and estimation through C-accelerated routines and multi-core parallelization. Model selection across TVTP specifications is efficiently conducted via comparative information criteria (AIC, BIC), and the multidimensional structure of TVTP offers transparency for empirical identification of regime drivers. Example R code illustrates a typical simulation-estimation-visualization loop as described in (Modée et al., 14 May 2026). The software exposes all underlying objects used in presented Monte Carlo and empirical analyses, facilitating full reproducibility and benchmarking by external researchers.

---

A table summarizing the key model specifications implemented in multiregimeTVTP:

Model Name	$z_t\in\{1, ..., K\}$5 Specification	Typical Application Source
Constant TVP	$z_t\in\{1, ..., K\}$6 constant vector	Baseline Markov switching
Model I	$z_t\in\{1, ..., K\}$7	Endogenous transition
Model II	$z_t\in\{1, ..., K\}$8	Exogenous driver (macro/fin)
Model III (GAS)	$z_t\in\{1, ..., K\}$9	Score-driven (GAS), typically not identified in practice

---

The multiregimeTVTP class offers a unified, efficient platform for simulation, estimation, filtering, and applied analysis of multiregime time series with general, potentially high-dimensional time-varying transition probabilities, exposing both methodological strengths and empirical limitations directly through its open-source implementation (Modée et al., 14 May 2026).