Uma abordagem apoiada em portfólio de planejadores probabilísticos aplicada ao mercado financeiro

Abstract

In the context of planning, one of the recurrent problems is choosing the most suitable planner for a given problem. Besides that, for certain problems, several combinations of planners can be used. In this work, we describe an approach that seeks, from the specification of the planning problem, a planner that is suitable to solve the problem. In order to do so, a portfolio of probabilistic planners is used for problems specified as POMDPs (Partially Observable Markov Decision Process) and machine learning techniques, which is a general purpose approach, regardless of the type of planning problem. However, for evaluation purposes, the approach is applied in the selection of the most suitable planner for the financial market planning problem.

The stock market can be considered a domain non-deterministic and partially observable as investors never know all the information that affects prices and the results of an investment are almost always uncertain. Thus, this paper proposes the use of probabilistic planning in order to assist the active of trading in the financial market. It is proposed a new model to the market problem using MOMDP (Mixed Observability Markov Decision Process). Other points are treated, such as the use of multiple assets, the cost of brokerage and mounting investment portfolios.

For the financial market, offline and online planning can both be used, and the choice of planning type can vary according to the form of negotiation. Depending on the different existing planners, the present work explores the use of both types of planning, but in particular online planning in portfolio creation.