Stock prediction based on LightGBM with feature selection and improved grid search. (13th July 2022)

• Record Type: Journal Article
• Authors: Zhou, Qihang; Zhou, Changjun; Liu, Zhiqiang
• Is Part Of: International journal of adaptive and innovative systems. Volume 3:Number 2(2022)
• Page Start: 100
• Page End: 118
• Publication Date: 2022-07-13
• Subjects: LightGBM -- grid search -- stock forecast -- optimisation -- feature selection; Adaptive computing systems -- Periodicals; Computer systems -- Technological innovations -- Periodicals; Systems software -- Periodicals; 004
• Languages: English
• ISSNs: 1740-2107

Abstract

In order to improve the accuracy of stock forecasting, a stock forecasting model based on LightGBM is proposed. Firstly, based on the grid search, an improved grid search is proposed, and the improved grid search is used to search out the best super parameters for LightGBM. At the same time, the normalised function and loss function suitable for the model are also searched out, and the IGS-LightGBM model is constructed. Then, the feature selection in LightGBM is used to further optimise the model and obtained FS-IGS-LightGBM. The model is then applied to the actual stock data. In this paper, the closing prices of Shanghai Composite Index, Shenzhen Composite Index, Shanghai and CSI 300, Growth Enterprise Index, Chuanneng Power and Baiyun Airport are taken as experimental data. The experimental results show that FS-IGS-LightGBM is superior to XGBoost and LightGBM in the evaluation index of MAPE.