دانلود رایگان مقاله انگلیسی پیش بینی شاخص سهام در بازار روزانه: یک رویکرد شبکه بیزی - الزویر 2018

Abstract

In this work, we investigate the feasibility of Bayesian Networks as a way to verify the extent to which stock market indices from around the globe influence iBOVESPA – the main index at the S˜ao Paulo Stock Exchange, Brazil. To do so, index directions were input to a network designed to reflect some intuitive dependencies amongst continental markets, moving through 24 and 48 hour cycles, and outputting iBOVESPA’s next day closing direction. Two different network topologies were tested, with different numbers of stock indices used in each test. Best results were obtained with the model that accounts for a single index per continent, up to 24 hours before iBOVESPA’s closing time. Mean accuracy with this configuration was around 71% (with almost 78% top accuracy). With results comparable to those of the related literature, our model has the further advantage of being simpler and more tractable for its users. Also, along with the fact that it not only gives the next day closing direction, but also furnishes the set of indices that influence iBovespa the most, the model lends itself both to academic research purposes and as one of the building blocks in more robust decision support systems.

6. Conclusion

In this work, we set out to identify the applicability of a Bayesian Network structure in the investigation of the extent to which foreign markets influence the main index at the S˜ao Paulo Stock Exchange (i.e. iBOVESPA), by taking closing directions of stock market indices around the globe as input to the network, so as to try to forecast iBOVESPA’s next day direction. Being one of the few to move away from the Neural Network and Support Vector Machine mainstreams, our research is, to the best of our knowledge, the first machine learning based effort to take an “around the globe” approach, that is to base its underlying model on the assumption that markets are interrelated all around the world. We do so through a Bayesian Network that reflects the potential dependencies amongst indices from all continents while the planet rotates. Moreover, this is probably the machine learning effort trying to accommodate the highest number of different markets into a single model.

Since “the central idea to successful stock market prediction is achieving best results using minimum required input data and the least complex stock market model” (Atsalakis & Valavanis, 2009), and with results comparable to those of the related literature, our model moves towards this direction, by presenting the further advantage of being relatively simple and intuitive, when compared to other more “closed” models, such as those that deal with Neural Networks, for example. Also, the model makes no assumptions regarding error distribution, as do some time-series algorithms, thereby generalising over different datasets. Nevertheless, some of the more elaborate models, specially those that account for some traditional technical analysis indicators along with raw technical data, were found to be more accurate than ours, even though there are exceptions to this rule. This is a matter left for future investigation: whether by adding such information we can raise our system’s accuracy.