Predicción de índices bursátiles mediante un sitema híbrido basado en modelos ocultos de Markov y redes neuronales artificiales

Abstract

The mixture of expert approaches includes a great variety of models, whose philosophy consists in breaking down a time series in various states, so that each state is modeled through an expert, which results in capturing the patterns of the time series in an efficient way. The Time-Line Hidden Markov Experts (THMEs) represents one of the most advanced methods based on this approach, which are just hybrid models based on Artificial Neural Networks (ANNs) and Hidden Markov Models (HMMs). A distinctive characteristic of the THMEs is that the state transitions of the time series is modeled through a HMM, whose state transition matrix is time-variant. The aim of this research consisted in assessing the performance of the THME in the prediction of Stock Market Indexes, and to compare it with the results obtained through pure ANNs. The experiments were carried out using 15 stock market indexes, and the results obtained were that the THMEs substantially overcomes the ANNs, both in accuracy as in its capacity to detect patterns.