Seleção de carteiras por meio de machine learning e da influência da informação assimétrica

Abstract

This study had as main motivation the creation of an asset selection methodology, based on information from the Brazilian market microstructure acquired through the tool developed by Perlin and Ramos (2016) and applied to the informational asymmetry proxy model Probability Informed Trading developed by Easley, Hvidkjaer and O'Hara (2002; 2010) adapted by Lin and Ke (2011). Based on information from this model, specifically on its outputs, PIN and Delta, an asset selection model was developed, establishing these two variables as input of a machine learning algorithm developed by Chen and Guestrin (2016), called of eXtreme Gradient Boosting - XGBoost, with the objective of identifying the characteristics of the assets that have an abnormal return, that is, above the cost of capital captured by the Capital Asset Pricing Model (CAPM). The results showed that there is a possibility of obtaining an abnormal return using only the PIN and Delta as the asset selection parameter, found 32.92% and 16.46% for the 10 and 5 asset portfolios, respectively. In addition, at 57.62% of the occasions the investor would obtain a positive return, selecting the 10 assets with the highest predictions of abnormal return, and at least 52.38% a return was found above the CAPM price. These results were lower when only the first five assets were analyzed, which can be explained by the diversification of the portfolio proposed by Markowitz (1952).