Model design for stock statistical arbitrage using deep neural networks, random forests and gradient-boosted trees

Statistical arbitrage is a common investing strategy in inefficient markets which is market neutral and profits from both sides of the market without the need for initial capital. This research aims at designing suitable models for stock statistical arbitrage using deep neural network, random forest, gradient-boosted trees and equal-weighted ensemble of these methods whilst analyzes the returns and risks of the designed models. For this purpose, the information of all listed companies in Tehran Stock Exchange from 1385 until 1396 has been used to generate trading signals. The design of the research models and required coding also the testing of the research hypotheses which is analyzed by t-test were performed in R software. The research findings show that the highest daily return is 4.24% for k = 5 (prior transaction costs) which is for the simple equal-weighted ensemble (ENS). Also deep neural network (DNN) has the lowest value at risk (- 4.45%) and the lowest expected shortfall (- 5.57%) for k = 20. The highest value of the return to standard deviation ratio is 1.072 which belongs to the RAF model for k = 20. Moreover, research results show that recent returns have higher predictive power than previous returns.