| |
Abstract:
In this paper, we enhance the Q-learning algorithm for
optimal asset allocation proposed in (Neuneier, 1996). The new
formulation simplifies the approach and allows policy-iteration
without the necessity for having a model of the system. The new
algorithm is tested on real data of the German stock market.
Furthermore, the possibility of risk management within the
framework of Markov decision problems is analyzed. This leads to
different return functions, which bridge the gap between classical
portfolio management and asset allocation strategies derived by
reinforcement learning.
|
