The Martingale Method for utility maximization (seen in e.g. Björk's book) is based on separating the optimization problem $E^\mathbb{P}[U(X_T)]$ over a class of admissible strategies into the static problem of determining the optimal wealth profile $\hat{X}_T$. If the inital capital is $x$ the constraint is $E^\mathbb{Q}[X_T]=x$ hence the problem turns into maximizing the Lagrangian

\begin{equation} E^\mathbb{P}[U(X_T) - \lambda(L_T X_T - x)] \end{equation}
where $L_T$ is the Radon–Nikodym derivative and $\lambda$ is the lagrange multiplier. Solving this gives the optimal wealth as some function of $L_T$, $X_T = X_T(L_T)$, and since the martingale dynamics is $d\hat{X}_t = \hat{u}_t dW_t^\mathbb{Q}$ (for some predictable $\hat{u}_t$) all that is left is to determine $\hat{u}_t$ (e.g. by Ito's lemma).

Now to my question: I have noticed that for the problems I have encountered $\hat{u}_t$ is also a $\mathbb{Q}$-martingale. For example it might be easy to determine $\hat{u}_T$ close to the terminal date $T$, and then it is evident that $\hat{u}_t = E^\mathbb{Q}[\hat{u}_T | \mathcal{F}_t]$ after the problem has been solved. Is it possible to prove that this is true? If $\hat{u}_t$ is known this is obviously easy to check, but what can be said about the general case?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.