L52 Regimes and Nonlinear Models
- Regimes = states (any two are not the same) (they are divided by thresholds)
Modelling Regimes
Threshold based regimes
\[
\begin{align}
\text{Regime 1:} & y_{t-d} \lt \gamma \\
\text{Regime 2:} & y_{t-d} \gt \gamma \\
\end{align}
\]
- \(y_{t-d}\) is the lagged value of (another/same) variable
- \(\gamma:\) threshold value
- Example: TAR, SETAR
Smooth Transition regimes
\[
G(z_{t-d}; \gamma,c) = \dfrac{1}{1+e^{ -\gamma(z_{t-d}-c) }}
\]
- \(z_{t-d}:\) threshold variable
- \(\gamma:\) slope parameter
- \(c:\) threshold value
- Example: STAR model
Latent state based regimes
\[
P = \begin{bmatrix}
p_{11} & p_{12} \\
p_{21} & p_{22}
\end{bmatrix}
\]
- \(p_{ij}:\) probability of transitioning from \(i\) to \(j\)
- Example: Markov switching models like MSAR
Benefits of Regime Models
- Improved forecasting
- accuracy for systems with abrupt changes
- we can capture the entire system properly if we differentiate the regimes by using different models
- Interpretability:
- Flexibility:
- can capture a wide variety of nonlinear behavior
Challenges of Regime Models
- Threshold Selection
- How to select \(\gamma\)?
- Computationally intensive
- Overfitting
- Too many regimes \(\implies\) overly complex
- generalize poorly
- Data sufficiency
- Each regime requires sufficient data points
TAR model
- A class of nonlinear time series models
- behavior switches between regimes
- depending on whether threshold variable exceeds a predefined value
- This structure is good for capturing regime shifts in TS data
Key Features
- Regime Switching
- Model assumes different dynamics in different regimes
- Each regime is modelled by its own AR process
- Threshold value: determines the regime (commonly, \(y_{t-d}\))
- Piecewise linearity: overall nonlinear, it is linear within each regime
- Flexibility: handles systems where relationship between variables change depending on the regime they are in
\[
y_{t} = \begin{cases}
\phi_{1,0} + \phi_{1,1} y_{t-1} + \dots \phi_{1,p} y_{t-p} + e_{t}, & \text{if } y_{t-d} \le \gamma \\
\phi_{2,0} + \phi_{2,1} y_{t-1} + \dots \phi_{2,p} y_{t-p} + e_{t}, & \text{if } y_{t-d} \le \gamma \\
\end{cases}
\]
- \(\phi_{i,j}:\) AR coefficient for regime \(i\) and lag \(j\)
- \(y_{t-d}:\) threshold variable
- \(\gamma:\) threshold value
Steps in Model Building
- Specify threshold variable \((y_{t-d})\)
- Determine lag structure (\(p\) in AR for each regime)
- Estimate threshold (\(\gamma\) using grid search, minimize residual variance or information criteria)
- Estimate parameters: (Fit the AR processes using methods like least squares, for each regime)
- Diagnostics
- Residual autocorrelation, stationarity
- Model accuracy
- Hypothesis test for nonlinearity (e.g., Hansen's test)
Advantages of TAR Models
- Interpretability
- Flexibity
- Stationarity in Regimes: Each regime can be stationary even if overall is not
Challenges
- Threshold estimation
- Overfitting
- Boundary issues (sparse data near \(\gamma\) can make estimation difficult)
Applicaations
- Economics
- Business cycles
- inflation dynamics with unemployment rates as \(\gamma\)
- Climatology
- Representing shift in climate variables
- Engineering
- operational thresholds, load-bearing structures or machinery