A Genetic Algorithm Hyper-parameter Optimization of Ensemble Approach: A Strategy Sock Prediction

Abstract

Investors must predict stocks correctly in order to maximize profits while avoiding bankruptcy. But the stock market situation is difficult to detect. His changing behavior is influenced by various factors such as the political situation, the company's and global economy, that are available through the news. This study aims to develop a model that can predict stocks more accurately by combining stock technical indicators and news sentiment. Genetic algorithm (GA) optimizes several stacked decision tree-based ensembles using the stacked-generalization method with the meta-learner concept used in this study. There are five main stages of the methodology, starting with stock and news data collection, data preprocessing, feature extraction and data analysis, then model development. A series of GA crossover and mutation parameter trials gave optimum results for searching combinatoric hyper-parameter models with an accuracy of 81.63% and an f1-score of 82.21%. The evaluation of the model on the combination of dataset types was able to increase the prediction accuracy from 75.91% to 81.63%, and the f1-score from 77.56% to 82.21%. In terms of trading evaluation, give a fantastic return of 121.27% in a year. This evaluation surpassed the results of previous similar studies, even far above the performance of the buy & hold strategy.