Optimal retirement planning under partial information (Q2291758)

A retiree has an annuity which pays at a rate \(a\). She has further an amount \(x\) to invest. For the investment there is a riskless asset \(d B(t) = r B(t) \;d t\) and a risky asset \(d S(t) = S(t) \{ (r + \theta \sigma)\;d t + \sigma \;d W(t)\}\), where \(W\) is a standard Brownian motion. The remaining lifetime \(\tau\) of the retiree is independent of the Brownian motion and has a deterministic force of mortality \(\zeta(t)\). The market price of risk \(\theta\) (and thus the drift of the risky asset) is unknown and can take \(m\) different values \(\vartheta_1, \ldots, \vartheta_m\). Writing \(Y(t) = W(t) + \theta t\) the process \(Y(t) = \sigma^{-1} \{\log (S(t)/S(0)) - (r-\frac12 \sigma^2)t\}\) is observable. Bayes' rule yields \(\mathbb{P}[\theta = \vartheta_k \mid \mathcal{F}_t^S] = \mathbb{P}[\theta = \vartheta_k \mid Y(t)]\), which are explicit expressions. The retiree can decide on the investment strategy \((\pi(t))\) and on the consumption rate \((c(t))\), which need to be adapted. There is a minimal consumption \(\underline{c}\) that the retiree needs for living. Then the current wealth process fulfils \[ d X(t) = (r X(t) + a - c(t)) \;d t + \pi(t) \sigma \;d Y(t) \;.\] The goal is to maximise the expected utility of the consumption process \[ \sup_{(\pi,c) \in \mathcal{A}(x)}\mathbb{E}\Bigl[ \int_0^\tau e^{-\rho t}\frac{(c(t) - \underline{c})^{1-\gamma}}{1-\gamma}\;d t\Bigr]\;,\] where \(\gamma \in (0,\infty) \setminus\{1\}\) and \[\mathcal{A}(x) = \Bigl\{(\pi,c): X(0) = x, c(t) \ge \underline{c}, X(t) \ge \frac{\underline{c}-a}r, \int_0^\infty \pi(t)^2 \;d t < \infty\Bigr\}\;.\] The condition \(X(t) \ge (\underline{c}-a)/r\) is in order that there is enough wealth for the minimal consumption \(\underline{c}\). The problem is solved by considering \(d \hat W(t) = d Y(t) - \mathbb{E}[\theta \mid Y(t)]\;d t\) and changing the measure, such that \((\hat W(t))\) becomes a standard Brownian motion. This makes the observation complete. The solution is then the solution to Merton's problem with a correction term caused by the random life time. In the same way, logarithmic utility (\(\gamma = 1\)) is considered and it is shown that the strategy is just the limiting strategy as \(\gamma \to 1\). Some numerical examples and extensions to the model are discussed.