Say I have a portfolio, $X_t$, using a leverage of $f$, such that the dynamics are given by \begin{equation} dX_t = \mu f X_t dt + \sigma f X_t dW_t \end{equation} I want to optimize the expected utility after some time $T$, $E[U(V_T)]$, and find the optimal leverage $f$. With the utility function $U(x)=\frac{x^\gamma}{\gamma}$ this is fairly easy. The SDE can be solved and the expected utility is maximized with $f^* = \frac{\mu}{\sigma^2 (1-\gamma)}$. With $\gamma=0$ and log-utility this is just the Kelly criterion.

But what if I also have to pay a constant cost $C$ such that the dynamics are \begin{equation} dX_t = (\mu f X_t - C) dt + \sigma f X_t dW_t, \quad X_t > 0 \end{equation} and $dX_t = 0$ when $X_t=0$ (i.e., I go bankrupt). The utility function would need to be altered to account for the non-zero probability of bankruptcy, so $U(x)=\frac{(x + b)^\gamma}{\gamma}$ with some $b>0$ so that the utility is bounded at bankruptcy.

Is there any way I can formulate the problem such that I can get an expression for the optimal $f^*$ that maximizes the expected utility $E[U(V_T)]$ when costs are included?


I hope my computations are correct.

Let $u(t,x)=\max_{(f_s)_{s\geq t}}\mathbb{E}[(b+X^{f_.}_T)^\gamma]$.

Using HJB (you have to prove that it is ok to use it).

$$0=\max_{f}\partial_t u(t,x)+(\mu f x - C)\partial_x u(t,x)+\frac{\sigma^2}{2}f^2x^2\partial_{xx}u(t,x)$$ Since $\partial_{xx}u(t,x)<0$ (prove it), maximum is hit at $f=\frac{\mu x \partial_xu(t,x)}{\sigma^2 x^2\partial_{xx}u(t,x)}$

$$0=\partial_t u(t,x)- C\partial_x u(t,x)-\frac{\mu^2}{2\sigma^2}\frac{(\partial_x u(t,x))^2}{\partial_{xx}u(t,x)}$$

We postulate $u(t,x)=f(t)(b+x)^{\gamma(t)}$

one has : $$ 0 = f'(t)(b+x)^{\gamma(t)}+\gamma'(t)(b+x)^{\gamma(t)-1} - C \gamma(t) (b+x)^{\gamma(t)-1}-\frac{\mu^2}{2\sigma^2}\frac{(b+x)^{\gamma(t) } }{\gamma(t)-1}$$

And finally we have : $$0=\gamma'(t) - C \gamma(t)\text{ and }\gamma(T)=\gamma$$ and $$f'(t)= \frac{\mu^2}{2\sigma^2(\gamma(t)-1)}\text{ and }f(T)=1$$

which leads to:

$$u(t,x)= (b+x)^{\gamma e^{-C(T-t)}}\left(1-\int_{t}^T\frac{\mu^2}{2\sigma^2(\gamma e^{-C(T-s)}-1)}ds\right)$$

Post in comments, if you see computation mistakes.

  • $\begingroup$ This looks like a great way to do it. Note however that $\gamma<0$ (usually) since higher risk-aversion will lead to more concavity, so maximizing $\mathbb{E}[(b+X^{f_.}_T)^\gamma]$ would imply minimizing the expected utility. Do you have any reference on using HJB in similar settings? $\endgroup$ – Freelunch Feb 16 '17 at 7:49
  • $\begingroup$ I would have said $0<\gamma <1$, but $\gamma<0$...? $\endgroup$ – MJ73550 Feb 16 '17 at 14:16

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.