Online Learning Applied to Autonomous Valuation of Financial Assets

Abstract

In the context of Artificial Intelligence, Online learning is focused on environments that are not independent and identically distributed, i.e. the environment may changes its behavior as time goes by. Blum proposed a famous algorithm to this problem, which was called randomized weighted majority algorithm. In this paper, we propose an adaptation of such algorithm to autonomous valuation of financial assets. Our approach is based on learning from expert's advices, in order to create a more adaptable solution and reuse some results achieved for other researchers. We also briefly review some papers in the field. The proposed approach is materialized through an online learning algorithm that defines an analysis derived from many different analyses performed by autonomous analysts. Such analysts may be created using techniques from finance or machine learning fields. Our algorithm is able to take into account different costs of analysis errors. We believe that this skill in fundamental to an efficient analyst. We implemented the algorithm and tested it using several different techniques from finance and one (very simple) algorithm from machine learning area. This implementation was tested and the achieved results are analyzed and discussed. Furthermore, we proved that our algorithm's cost of error is limited by an expression of the cost of error of the best autonomous analyst. We believe that this algorithm may contribute to development of better systems that intend to estimate the price of financial assets in an autonomous way.