Forecasting and Transfer Function Noise Model
Minimized MSE Forecasts for ARMA models
Consider a stationary ARMA model (casual and invertible)
Rewrite as a MA process
Then, for \(t = n + h\),
Suppose we have observations till \(X_n, X_{n-1},...\) and wish to forecast \(h\) step ahead of future values \(X_{n+h}\) as a lin.comb. of the observations. Then, we define the mean square error forecaster
where \(\hat\psi_i\) are parameters to be determined.
Then MSE of the forecast is
Rule of calculating conditional expectation
Using the fact that \(a_i\)'s are uncorrelated
For \(h > 0\),
Example
Consider a AR(1) model \(X_t = 0.5X_{t-1} + a_t\). Then
Transfer Function Noise model
Consider the model that \(X_t = f(X_{t-1}, X_{t-2}, ... , Z_t, Z_{t-1},...)\), \(X_t\) is a linear combination of past observations and an external variable.
A TFN model is a time series regression that predict values of a dependent variable based on both the current and lagged values of one or more explanatory variables.
Procedure of building the single input TFN model
- Preliminary identification of the impulse response coefficients \(v_i\) (prewhitening)
- Specification of the noise term \(n_t\)
- Specification of the transfer function using a rational polynomial in B if necessary
- Estimation of the TFN specified
- Model diagnostic checks
Rational distributed lag model \(v(B)\) can be approximated by a ratio of polynomials
and then, \(y_t = \delta(B)x_t / \theta(B) + n_t\)
ARMAX
\(y_t = v(B)x_t + n_t = v_0 x_t + v_1 x_{t-1}+v_2 x_{t-2}+...+n_t\)
where \(v(B) = \sum^\infty v_j B^j\), and \(x_t,n_t\) are independent.
The coefficient \(v_0, v_1,...\) are referred as the impulse response function of the system.
To make such equation to be meaningful, \(\sum^\infty |v_j| = g<\infty\), which the system is stable and \(g\) is called the stead-state gain. \(g\) represents the impact on \(Y\) when \(X_{t-j}\) are held constant over time.
Properties
\(x_t \sim ARMA(p,q)\), \(v_i = \phi^i (1-\phi)\), \(y_t = \sum^\infty v_i x_{t-i}+a_t\)
Pre-whitening
Consider \(x\sim\)ARMA, i.e. \(\Phi_x(B)x_t = \Theta_x(B)a_t\)
Apply \(\Phi_x(B)/\Theta_x(B)\) on TFN model
Let \(\tau_t = \frac{\Phi_x(B)}{\Theta_x(B)} y_t, n_t = \frac{\Phi_x(B)}{\Theta_x(B)} \epsilon_t\), we get
To get \(\gamma_{a\tau}(0)\), multiply both sides by \(a_t\) and take the expectations.
To get \(\gamma_{a\tau}(1)\), multiply both sides by \(a_{t-1}\)
Therefore, \(\gamma_{a\tau(k)} = v_k\sigma^2\)
Since \(\rho_{a\tau}(k) = \gamma_{a\tau}(k) / \sigma_a\sigma_\tau\)
Box-Tiao Transformation
Similarly, since \(n_t\sim\) ARMA, i.e. \(\Phi_n(B)n_t = \Theta_n(B)a_t\)
Which \(\frac{\Phi_n(B)}{\Theta_n(B)}n_t = a_t\)
Then, apply \(\frac{\Phi_n(B)}{\Theta_n(B)}\) to both sides of the equation.
is called the Box-Tiao Transformation
Steps of The Estimation Procedure
The steps of the estimation procedures
- Run the OLS regression on \(y_t = \sum_{j=1}^s v_j x_{t-j} + e_t\) to collect the residuals \(\{\hat e_t\}\)
- Identify an ARMA model for \(\hat e_t\)
- Apply Box-Tiao transformation to filter \(y_t, x_t\)
- Run regression on the transformed equation
- check the correlation of regression residuals