Métodos de seleção de variáveis com aprendizado de máquina para estratégias de arbitragem estatística configuradas como predição de séries temporais

Abstract

Applied research in quantitative finance has high entry-barrier due the lack of high-quality data, which leads to false positive strategies. In present work the data is provided by a tournament, organized by a market-neutral hedge fund who configured a statistical arbitrage challenge as forecasting problem. In this context, there is a substantial amount of features and based on the feature selection literature the task is to develop strategies to consistently outperform a given benchmark. Initially, the predictions created by a tree-based estimator were decomposed into linear and non-linear portions to estimate the source of volatility excess. Next, a method which accounts for all trials in order to ensure positive returns were introduced. Several feature selection methods to attenuate the multicollinearity were implemented and a small extension to a feature clustering algorithm was proposed to decrease the risk of overestimating unimportant variables. To enhance better results we denoised the correlation matrix and proposed a codependency measure robust to nonlinear relationships. Then, was made an attempt to map regimes that describe market behavior over time, but it was not possible to accurately predict regime change. Despite reasonable results when the assumption is true. This research is limited to feature selection literature, but within this scope we were not able to state that the benchmark can be consistently overcomed, despite having done so most of the time. So we rely on the parity risk algorithms literature to present a workaround solution for the problem.