The following is a specific question that is useful for demonstrating a general idea.

Consider the following autoregressive model: $$ X_{t+1} = \alpha_0 + \beta_0 (X_t - \alpha_0) + W_{t+1}, $$ where $-1 < \beta_0 < 1$ and $W_{t+1}$ is distributed as a normal with mean zero and variance one.

How should I go about constructing the bivariate score process associated with the parameters $\alpha_0$ and $\beta_0$? How can I verify that it is a martingale?


I would begin by constructing the log-likelihood process as follows (conditioning on $X_0$): $$ \ell_t(\theta \mid \textbf X) = -\frac t2 \ln(2 \pi) - \frac 12 \sum_{j=1}^t (X_j - \alpha_0 - \beta_0(X_{j-1} - \alpha_0))^2. $$ Then, the score process can be constructed as $$ s_t(\theta \mid \textbf X) = \begin{bmatrix} (1 - \beta_0) \sum_{j=1}^t (X_j - \alpha_0 - \beta_0(X_{j-1} - \alpha_0)) \\ \sum_{j=1}^t (X_j - \alpha_0 - \beta_0(X_{j-1} - \alpha_0)) (X_{j-1} - \alpha_0) \end{bmatrix}. $$ Is this correct? How do I proceed?


The derivation of the score process is correct. To verify that the process is a Martingale, recall the definition. It becomes clear that if we substitute $W_{t+1}$ back into the equation $$ s_t(\theta \mid \textbf X) = \begin{bmatrix} (1 - \beta_0) \sum_{j=1}^t W_j \\ \sum_{j=1}^t W_j (X_{j-1} - \alpha_0) \end{bmatrix}. $$ Because $W_{t+1}$ are Normal with mean 0 and variance 1 (I'm assuming they're iid), then $$ E[s_{t+1} \mid s_t ] = s_t + E \left [ \begin{matrix} (1 - \beta_0) W_{t+1} \\ W_{t+1} (X_t - \alpha_0) \end{matrix} \middle | s_t \right ] = s_t $$ and we are finished.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.