Time Varying Coefficients vs Rolling Estimation What are the practical differences for forecasting from fitting a model with time varying coefficients vs. estimating a model with fixed parameters over rolling windows? Intuitively it seems that these approaches both attempt to capture time variation in parameters, but the former model does this explicitly whereas the latter model handles this implicitly. I am trying to understand the tradeoffs with the two approaches, in particular

*

*Besides various cross validation methods, are there other approaches to determine which model would be suitable for a given use case? Things like AIC and BIC come to mind but in the rolling case, since we are actually refitting many models I'm not sure how these would be applicable?

*Are there certain situations where one class of model is generally preferred? e.g. Is quantifying things like forecast uncertainty more tractable via time varying parameter models?

Within the context of statespace models, the models I am referencing have the following form. Quoting from Time Series Analysis by State Space Methods by J. Durbin and S.J. Koopman Section 3.6,

$$ y_t = X_t \beta + \epsilon_t \hspace{10mm} \epsilon_t \sim N(0,
> H_t)  \hspace{10mm} (3.30)$$ for $t = 1,...,n$ where $X_t$ is the $1 \times k$ regressor
vector with exogenous variables, $\beta$ is the $k \times 1$ vector of
regression coefficients and $H_t$ is the known variance that possibly
varies with $t$. This model can be represented in the state space form (3.1) with $Z_t = X_t, T_t = I_k$ and $R_t = Q_t = 0$, so that $\alpha_t = \alpha_1 = \beta$
...
Suppose that in the linear regression model (3.30) we wish the coefficient vector
$\beta$ to vary over time. A suitable model for this is to replace $\beta$ in (3.30) by $\alpha_t$ and to
permit each coefficient $\alpha_{it}$ to vary according to a random walk $\alpha_{i,t+1} = \alpha_{it} + \eta_{it}$.
This gives a state equation for the vector $\alpha_t$ in the form $\alpha_{t+1} = \alpha_t + \eta_t$.

The referenced form in (3.1) is given by
$$
\begin{align}
y_t &= Z_t \alpha_t + \epsilon_t, \hspace{10mm} \epsilon_t \sim N(0, H_t), \\
\alpha_{t+1} &= T_t \alpha_t + R_t\eta_t, \hspace{5mm} \eta_t \sim N(0, Q_t), \hspace{8mm} t=1,...n, 
\end{align}
$$
 A: 
Besides various cross validation methods, are there other approaches to
determine which model would be suitable for a given use case? Things like AIC and BIC come to mind but in the rolling case, since we are actually refitting many models I'm not sure how these would be applicable?

Time-varying coefficients will capture nonlinearity in fine-grain level, i.e., roughness/smoothness in time-series literature, Gaussian Process literature, see Lecture notes of Zoubin Ghahramani for how exponent generates different roughness.  Measuring roughness for time-series can be used in deciding suitable model. Higher roughness indicates higher nonlinearity, so time-dependent coefficients would be more suited. See How to measure smoothness/roughness of a time series. AIC/BIC may not be able to capture this, they are more of complexity measure.

Are there certain situations where one class of model is generally preferred? e.g. Is quantifying things like forecast uncertainty more tractable via time varying parameter models?

Higher the roughness, explicit modelling is preferred.
